LANGUAGE AND COGNITIVE PROCESSES 1997 12 (56) 657ndash693

Requests for reprints should be addressed to Ardi Roelofs Max Planck Institute forPsycholinguistics PO Box 310 6500 AH Nijmegen The Netherlands E-mail ardi6mpinl

I am indebted to Antje Meyer Pim Levelt Gary Dell and an anonymous reviewer fordiscussion and to Maarten van Casteren for his help in preparing and running the experiment

q 1997 Psychology Press Ltd

Syllabication in Speech Production Evaluation ofWEAVER

Ardi Roelofs

Max Planck Institute for Psycholinguistics Nijmegen The Netherlands

Models of speech production differ in their claims about syllabi cation In amemory-based approach such as that advanced by Dell (1986 1988) thesyllable structure of each word is stored in the mental lexicon In contrastaccording to a rule-based approach planning of speech involves theassignment of syllable positions to segments after they have been retrieved fora word from memory (eg Levelt 1992) Here a case is made for therule-based approach of the WEAVER model of speech production (Roelofs1994 in press a) First I argue that cross-morpheme and cross-wordsyllabi cation point to the need to deal with exibility of syllable membershipand therefore pose difculty to a memory-based approach but not toWEAVER Secondly I review empirical support for the specic form ofsyllabi cation realised in WEAVER Thirdly I report a new experiment onsyllabi cation which supports WEAVER rather than Dellrsquos model Finallythe issue of resyllabi cation is discussed

INTRODUCTION

In planning utterances speakers call on many facets of their storedknowledge about words including their meaning syntactic propertiesmorphological composition and phonological structure Lexical access is theprocess by which this information about words is retrieved from memory soas to construct articulatory programs for concepts to be verbally expressedThe access to a single word is achieved in two major steps lemma retrievaland word-form encoding (eg Butterworth 1989 Dell 1986 Garrett 19751988 Kempen amp Huijbers 1983 Kempen amp Hoenkamp 1987 Levelt 19891992 Roelofs 1992ab 1993) In lemma retrieval a lexical concept is used toretrieve the lemma of a corresponding word from memory A lemma as Iuse the term in this paper represents the syntactic properties of a word For

658 ROELOFS

example the lemma of the Dutch word juweel (gem) says that it is a noun andthat its grammatical gender is neuter In word-form encoding themorphophono logical properties of the word are retrieved from memory toconstruct an articulatory program It is generally assumed that this processinvolves the recovery or assignment of a syllable structure

Theories of the encoding of word forms differ in their claims aboutsyllabi cation On the one hand ldquomemory-basedrdquo theories assume that thesyllable structure of each word is stored in the mental lexicon For exampleon this view the memory representation of juweel would specify somethinglike onset j and nucleus y for the rst syllable and onset w nucleus eand coda l for the second syllable (eg Levelt 1989 Shattuck-Hufnagel1979) This view of syllabi cation is held by the classical computationalmodel of form encoding in speech production the model of Dell (19861988) On the other hand ldquorule-basedrdquo theories assume that memory doesnot contain such ready-made syllable structures for words Instead word-form encoding comprises a process that assigns syllable positions tosegments after they have been retrieved from memory (eg Beacuteland Caplanamp Nespoulous 1990 Levelt 1992 Levelt amp Wheeldon 1994) For examplethe memory representation of juweel says that its nal segment is l but doesnot specify the syllable position of this segment Whether the l occupies anonset or a coda position will only be determined when the word is actuallymade part of an utterance

Theories assign different roles to metrical structure in syllabi cationMetrical structures specify at least the number of syllables and the stresspatternmdashif they represent nothing else they are minimal Theories differ inwhether metrical structure is computed on-line (eg Beacuteland et al 1990) orstored (Levelt 1992) Stored metrical structures may be minimal (Roelofs1994 in press a) or may represent the precise CV structure of the syllables(Dell 1988) According to Levelt (1992) syllable structure is computedon-line by associating retrieved segments with retrieved metrical structuresThis view of syllabi cation is realised in a specic manner in the WEAVER(Wordform Encoding by Activation and VERication) model of formencoding in speech production (Roelofs 1994 in press a) According toWEAVER a principle of economy holds in that metrical structures arestored for some but not all words

In this paper I make a case for the view on syllabi cation held byWEAVER The paper is organised as follows First I discuss howsyllabi cation is treated in two computational models in the literature In thediscussion of a memory-based approach I focus on the model of Dell (19861988) because this model has been more explicit about syllabi cation thanother models of this kind Furthermore aspects of the model have beenexplored through simulation I argue that Dellrsquos model has difculty withcertain aspects of syllabi cation in particular syllabi cation across

SYLLABIFICATION 659

morpheme boundaries and across word boundaries Furthermore I explorehow these problems might be solved within the framework of this model Inthe discussion of a rule-based approach I focus on the WEAVER model ofword-form encoding I will show how WEAVER handles syllabi cationacross morpheme boundaries and word boundaries Next I review theoutcomes of two series of experiments that support WEAVERrsquos specicview on syllabi cation Then I report a new experiment that contrasts Dellrsquosmodel and WEAVER Finally I discuss the issue of resyllabi cation

THEORETICAL POSITIONS

To set the stage I discuss the role of word-form encoding in the process oflexical access in speech production Three major types of processes underliespeaking conceptualisation formulation and articulation (eg Caplan 1992Garrett 1975 Kempen amp Hoenkamp 1987 Levelt 1989) Conceptual-isation processes generate so-called messages that is conceptual structuresto be verbally expressed In the view held by WEAVER messages arespecied in terms of lexical concepts and their relationships Formulationprocesses take the message access appropriate words for the lexical con-cepts (ie lemma retrieval and word-form encoding) and build a syntacticstructuremdashin the case of sentence productionmdashand a morphophonologicalstructure The result is an articulatory program for the utterance Finallyarticulation processes execute the articulatory program which results inovert speech

As indicated lexical access consists of two major steps lemma retrievaland word-form encoding corresponding to the formulation stages ofsyntactic encoding and morphophonological encoding respectively Inlemma retrieval a target lexical concept is used to retrieve the lemma of acorresponding word from memory Lemmas represent the syntacticproperties of words Lemma retrieval makes these properties available forsyntactic encoding processes In word-form encoding the morpho-phonological properties of the word are retrieved from memory to constructan articulatory program Below I describe a widely accepted psycho-linguistic view on lexical access (cf Levelt 1989 1992) which has beenadopted by WEAVER The encoding stages are uncontroversial but theassumptions about the computations and their output may differ betweenmodels

Assume that a Dutch speaker wants to name a gem Lexical access consistsof mapping the lexical concept G EM(X) onto the articulatory program forjuweel The lemma retriever takes GEM(X) and outputs the lemma of juweelTo derive the singular form [jyrsquowel] instead of the plural form [jyrsquowel] thewordrsquos number has to be specied The lemma plus this abstractmorphosyntactic number parameter are then input to word-form encoding

660 ROELOFS

FIG 1 Lexical access in speech production Lemma retrieval and word-form encoding (cf theWEAVER model Roelofs 1994 in press a)

The articulatory program is derived in three major steps (cf Dell 1986Levelt 1989 1992) morphological encoding phonological encoding andphonetic encoding (see Fig 1) The morphological encoder takes the lemmaof juweel plus the parameter singular and produces the stem morpheme juweel This rst stage thus concerns what is traditionally called theldquosyntaxndashmorphology interfacerdquo (eg Spencer 1991) The phonologicalencoder takes the stem morpheme and produces a so-called phonologicalword representation This representation describes the singular form ofjuweel as a phonological word consisting of two feet the rst one metricallyweak (w) and the second one strong (s) The rst syllable has j as onset andy as nucleus the second syllable has w as onset e as nucleus and l ascoda This second stage thus comprises what is traditionally called the

SYLLABIFICATION 661

ldquomorphologyndashphonology interfacerdquo (eg Goldsmith 1990) Finally thephonetic encoder takes this phonological word representation and deliversthe articulatory program [jyrsquowel] In WEAVER phonetic encodinginvolves accessing a store of learned motor programs for syllables aso-called syllabary (Levelt amp Wheeldon 1994) and setting parameters forpitch loudness and duration This nal encoding stage includes what issometimes called the ldquopost-lexical phonologyrdquo (eg Goldsmith 1990)

Dellrsquos model assumes stages of access similar to those of WEAVERdescribed above but there are also some important differences Forexample Dellrsquos model does not make the claim that the output ofphonological encoding is a phonological word representation and it does notassume a syllabary How is memory organised according to the models andwhat is the lexical access mechanism In particular how is syllabi cationachieved

DELLrsquoS MEMORY-BASED APPROACH

Dell (1986) assumes that word forms are represented in a lexical networkwith nodes for morphemes syllables rimes segments and segment clustersand features Figure 2 illustrates this for the Dutch word juweel (rimes andfeatures are omitted to simplify the gure) The nodes for the segments andclusters are marked for syllable position The nodes are associated with eachother by weighted bidirectional connections In phonological encoding theactivation level of a morpheme node is enhanced by giving it a jolt ofactivation Activation then spreads through the network each node sendinga proportion r of its activation to its direct neighbours Mapping a morphemeonto its segments is accomplished by selecting the highest activated segmentor cluster nodes The speaking rate determines the amount of time that isdevoted to the encoding of a syllable (ie the time between providing the joltof activation and segment selection) Selected nodes are inserted into slots ofindependentl y created syllable frames The segment nodes for a syllable areselected in parallel

In the case of multisyllabic morphemes the links between the morphemenode and the syllable nodes are labelled for the serial position of thesyllables The syllables are serially encoded The successive encoding ofsyllables is accomplished by temporarily increasing the spreading rate forone syllable and decreasing it for the others In the encoding of a bisyllabicmorpheme activation rst spreads with a rate of 15 r to the rst syllablenode and with 05 r to the second syllable node After the rst syllable framehas been lled the weights are changed Activation now spreads at a rate of05 r to the rst syllable node and at a rate of 15 r to the second syllable nodeThe changing of weights prevents selection of segments or clusters of the

662 ROELOFS

FIG 2 Memory representation of the word form of juweel (gem) in the model of Dell (19861988)

second syllable for the slots of the rst syllable frame or vice versa (recallthat the highest activated nodes are selected)

In a subsequent paper Dell (1988) proposed an additional networkencoding for all possible ldquowordshapesrdquo Each word in the lexical network isconnected to a wordshape header node that represents its CV structure Awordshape node sequentially activates segment category nodes that isonset consonant (COn) vowel (V) and coda consonant (CCo) Each of thecategory nodes connects back to all possible segment nodes of its category inthe lexical network Instead of selecting in parallel nodes for the OnsetNucleus and Coda slot of a frame the successive activation of the segmentcategory nodes leads to serial selection This allows the model to account forthe seriality effects in phonological encoding obtained by Meyer (19901991) using the so-called implicit priming paradigm These ndings will bediscussed below in the section on implicit priming

What is the empirical support for the explicit representation of CVstructure Empirical evidence concerning CV structure is scarce (for areview see Roelofs amp Meyer in press) Concerning speech errors there issome evidence for a tendency towards increased similarity of the CVstructures of the words involved For example segment additions tend tocreate clusters rather than singletons when the source word also has a

SYLLABIFICATION 663

cluster as in ldquoprich playerrdquo instead of ldquorich playerrdquo (Stemberger 1990) Theevidence is not however very strong First the CV similarity has beenobserved for onsets but not for nuclei Secondly the CV similarities areobserved for some but not all error corpora For example Stemberger(1984) observed them for small corpora of German and Swedish errors butnot for an English corpus

In terms of experimental evidence in two priming experiments withDutch participants Meijer (1994) observed that word production wasfacilitated by spoken primes with the same CV structure as the targetsrelative to primes that differed in their number of consonants However inanother experiment this effect was not replicated In a fourth experiment inwhich primes and targets had the same or different structures of the nucleus(V or VV) no priming effect from shared CV structure was obtained

Sevald Dell and Cole (1995) used a repeated pronunciation task to studythe representation of CV structure in English Participants had to producepairs of one monosyllabic and one disyllabic pseudoword as often as theycould within 4 sec The sequences were produced faster when themonosyllable had the same CV structure as the rst syllable of the disyllablethan when this was not the case as in kul-parfen versus kult-parfen Basedon these results Sevald et al argued for the explicit representation of CVstructure However as Sevald et al measured how many targets theparticipants could produce within a given time period it is difcult toestablish the locus of the effect of CV structure It could arise during thecreation of the phonological representation as Sevald et al propose but itcould also originate during the retrieval or execution of motor programs (seeRoelofs amp Meyer in press)

PROBLEMS WITH A MEMORY-BASED APPROACH

I now argue that a memory-based approach to syllabi cation (eg Dell 19861988 Houghton 1990 Shattuck-Hufnagel 1979) has difculty withsyllabi cation across morpheme boundaries and word boundaries Theproblem is that a segment of one morpheme or word may be syllabi ed withanother morpheme or word This may occur in the production ofpolymorphemic words or connected speech (eg Chomsky amp Halle 1968Kaisse 1985 Levelt 1989 1992 Nespor amp Vogel 1986 Selkirk 1984Spencer 1991) Models that rigidly store words as sequences of syllablenodes or models that store each consonant as an onset or coda have adifcult time dealing with the need for exibility of syllable membership

Cross-morpheme Syllabication

Consider the production of the plural form juwelen of the Dutch wordjuweel The plural juwelen is created by adding en ([]) to the stem juweel The resulting form is syllabi ed as (jy) s (we) s (l) s Thus

664 ROELOFS

juxtaposing en changes the syllabi cation of l This segment occupies acoda position in juweel syllabi ed as (jy) s (wel) s but an onset position injuwelen The syllabi cation across morpheme boundaries poses difculty tomodels that assign segments to syllable slots on a morpheme-by-morphemebasis

According to the model of Dell at the morpheme level the plural of theword juweel would be represented by the stem node juweel and a nodefor the sufx en The question is how the model handles thesyllabi cation of the l that is triggered by juxtaposing the sufx The stem juweel is connected to the coda l node In Dell (1986) this node will beselected to ll the coda slot of the second syllable frame However for theplural form the l should be the onset of the third syllable Thus the onset lnode has to be selected for the third syllable frame However onset l willnot be activated let alone be selected

The wordshape headers and the sequence of category nodes proposed byDell (1988) provide for the specication of a syllable structure that is uniqueto a particular combination of morphemes A header and sequence mightspecify that the fth segment of the plural form juwelen occupies an onsetposition instead of the coda position it has in juweel Of course a categorynode alone does not uniquely identify a segment Recall that a category nodeconnects back to all segments in its category Thus to be selected onset lmust also be activated by the syllable node len But what causes this node tobecome active One might propose to connect the syllable node len to themorpheme node juweel as its third syllable However this codes thewrong information because len is not the third syllable of juweel but thethird syllable of the plural juwelen Alternatively len might be connected to en as one of its ldquoallomorphsrdquo and also to juweel without beingexplicitly marked as its third syllable By not marking the syllable it may goldquounnotedrdquo in the encoding of syllables of juweel in producing thesingular form juweel To prevent interference with the encoding of ju and wein the production of juwelen the spreading rate between juweel and lenshould be sufciently small Then activating en activates all itsallomorphs including len The onset l now receives activation from juweel en and COn so it will be the most highly activated onset nodeand will be selected For the correct production of juwelen however this isnot sufcient Note that we is the second syllable of juwelen thus the syllablenode we should have a labelled link to juweel Therefore we need a (yetto be specied) mechanism that selects the appropriate second syllablenode wel in producing singular juweel and we in producing plural juwelen

In sum cross-morpheme syllabi cation in Dellrsquos model might be achievedby connecting the critical third syllable (eg len) to both morphemesinvolved and connecting the syllablersquos segments to category nodes (eglinking len to the stem juweel and the sufx en and linking onset l

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

658 ROELOFS

example the lemma of the Dutch word juweel (gem) says that it is a noun andthat its grammatical gender is neuter In word-form encoding themorphophono logical properties of the word are retrieved from memory toconstruct an articulatory program It is generally assumed that this processinvolves the recovery or assignment of a syllable structure

Theories of the encoding of word forms differ in their claims aboutsyllabi cation On the one hand ldquomemory-basedrdquo theories assume that thesyllable structure of each word is stored in the mental lexicon For exampleon this view the memory representation of juweel would specify somethinglike onset j and nucleus y for the rst syllable and onset w nucleus eand coda l for the second syllable (eg Levelt 1989 Shattuck-Hufnagel1979) This view of syllabi cation is held by the classical computationalmodel of form encoding in speech production the model of Dell (19861988) On the other hand ldquorule-basedrdquo theories assume that memory doesnot contain such ready-made syllable structures for words Instead word-form encoding comprises a process that assigns syllable positions tosegments after they have been retrieved from memory (eg Beacuteland Caplanamp Nespoulous 1990 Levelt 1992 Levelt amp Wheeldon 1994) For examplethe memory representation of juweel says that its nal segment is l but doesnot specify the syllable position of this segment Whether the l occupies anonset or a coda position will only be determined when the word is actuallymade part of an utterance

Theories assign different roles to metrical structure in syllabi cationMetrical structures specify at least the number of syllables and the stresspatternmdashif they represent nothing else they are minimal Theories differ inwhether metrical structure is computed on-line (eg Beacuteland et al 1990) orstored (Levelt 1992) Stored metrical structures may be minimal (Roelofs1994 in press a) or may represent the precise CV structure of the syllables(Dell 1988) According to Levelt (1992) syllable structure is computedon-line by associating retrieved segments with retrieved metrical structuresThis view of syllabi cation is realised in a specic manner in the WEAVER(Wordform Encoding by Activation and VERication) model of formencoding in speech production (Roelofs 1994 in press a) According toWEAVER a principle of economy holds in that metrical structures arestored for some but not all words

In this paper I make a case for the view on syllabi cation held byWEAVER The paper is organised as follows First I discuss howsyllabi cation is treated in two computational models in the literature In thediscussion of a memory-based approach I focus on the model of Dell (19861988) because this model has been more explicit about syllabi cation thanother models of this kind Furthermore aspects of the model have beenexplored through simulation I argue that Dellrsquos model has difculty withcertain aspects of syllabi cation in particular syllabi cation across

SYLLABIFICATION 659

morpheme boundaries and across word boundaries Furthermore I explorehow these problems might be solved within the framework of this model Inthe discussion of a rule-based approach I focus on the WEAVER model ofword-form encoding I will show how WEAVER handles syllabi cationacross morpheme boundaries and word boundaries Next I review theoutcomes of two series of experiments that support WEAVERrsquos specicview on syllabi cation Then I report a new experiment that contrasts Dellrsquosmodel and WEAVER Finally I discuss the issue of resyllabi cation

THEORETICAL POSITIONS

To set the stage I discuss the role of word-form encoding in the process oflexical access in speech production Three major types of processes underliespeaking conceptualisation formulation and articulation (eg Caplan 1992Garrett 1975 Kempen amp Hoenkamp 1987 Levelt 1989) Conceptual-isation processes generate so-called messages that is conceptual structuresto be verbally expressed In the view held by WEAVER messages arespecied in terms of lexical concepts and their relationships Formulationprocesses take the message access appropriate words for the lexical con-cepts (ie lemma retrieval and word-form encoding) and build a syntacticstructuremdashin the case of sentence productionmdashand a morphophonologicalstructure The result is an articulatory program for the utterance Finallyarticulation processes execute the articulatory program which results inovert speech

As indicated lexical access consists of two major steps lemma retrievaland word-form encoding corresponding to the formulation stages ofsyntactic encoding and morphophonological encoding respectively Inlemma retrieval a target lexical concept is used to retrieve the lemma of acorresponding word from memory Lemmas represent the syntacticproperties of words Lemma retrieval makes these properties available forsyntactic encoding processes In word-form encoding the morpho-phonological properties of the word are retrieved from memory to constructan articulatory program Below I describe a widely accepted psycho-linguistic view on lexical access (cf Levelt 1989 1992) which has beenadopted by WEAVER The encoding stages are uncontroversial but theassumptions about the computations and their output may differ betweenmodels

Assume that a Dutch speaker wants to name a gem Lexical access consistsof mapping the lexical concept G EM(X) onto the articulatory program forjuweel The lemma retriever takes GEM(X) and outputs the lemma of juweelTo derive the singular form [jyrsquowel] instead of the plural form [jyrsquowel] thewordrsquos number has to be specied The lemma plus this abstractmorphosyntactic number parameter are then input to word-form encoding

660 ROELOFS

FIG 1 Lexical access in speech production Lemma retrieval and word-form encoding (cf theWEAVER model Roelofs 1994 in press a)

The articulatory program is derived in three major steps (cf Dell 1986Levelt 1989 1992) morphological encoding phonological encoding andphonetic encoding (see Fig 1) The morphological encoder takes the lemmaof juweel plus the parameter singular and produces the stem morpheme juweel This rst stage thus concerns what is traditionally called theldquosyntaxndashmorphology interfacerdquo (eg Spencer 1991) The phonologicalencoder takes the stem morpheme and produces a so-called phonologicalword representation This representation describes the singular form ofjuweel as a phonological word consisting of two feet the rst one metricallyweak (w) and the second one strong (s) The rst syllable has j as onset andy as nucleus the second syllable has w as onset e as nucleus and l ascoda This second stage thus comprises what is traditionally called the

SYLLABIFICATION 661

ldquomorphologyndashphonology interfacerdquo (eg Goldsmith 1990) Finally thephonetic encoder takes this phonological word representation and deliversthe articulatory program [jyrsquowel] In WEAVER phonetic encodinginvolves accessing a store of learned motor programs for syllables aso-called syllabary (Levelt amp Wheeldon 1994) and setting parameters forpitch loudness and duration This nal encoding stage includes what issometimes called the ldquopost-lexical phonologyrdquo (eg Goldsmith 1990)

Dellrsquos model assumes stages of access similar to those of WEAVERdescribed above but there are also some important differences Forexample Dellrsquos model does not make the claim that the output ofphonological encoding is a phonological word representation and it does notassume a syllabary How is memory organised according to the models andwhat is the lexical access mechanism In particular how is syllabi cationachieved

DELLrsquoS MEMORY-BASED APPROACH

Dell (1986) assumes that word forms are represented in a lexical networkwith nodes for morphemes syllables rimes segments and segment clustersand features Figure 2 illustrates this for the Dutch word juweel (rimes andfeatures are omitted to simplify the gure) The nodes for the segments andclusters are marked for syllable position The nodes are associated with eachother by weighted bidirectional connections In phonological encoding theactivation level of a morpheme node is enhanced by giving it a jolt ofactivation Activation then spreads through the network each node sendinga proportion r of its activation to its direct neighbours Mapping a morphemeonto its segments is accomplished by selecting the highest activated segmentor cluster nodes The speaking rate determines the amount of time that isdevoted to the encoding of a syllable (ie the time between providing the joltof activation and segment selection) Selected nodes are inserted into slots ofindependentl y created syllable frames The segment nodes for a syllable areselected in parallel

In the case of multisyllabic morphemes the links between the morphemenode and the syllable nodes are labelled for the serial position of thesyllables The syllables are serially encoded The successive encoding ofsyllables is accomplished by temporarily increasing the spreading rate forone syllable and decreasing it for the others In the encoding of a bisyllabicmorpheme activation rst spreads with a rate of 15 r to the rst syllablenode and with 05 r to the second syllable node After the rst syllable framehas been lled the weights are changed Activation now spreads at a rate of05 r to the rst syllable node and at a rate of 15 r to the second syllable nodeThe changing of weights prevents selection of segments or clusters of the

662 ROELOFS

FIG 2 Memory representation of the word form of juweel (gem) in the model of Dell (19861988)

second syllable for the slots of the rst syllable frame or vice versa (recallthat the highest activated nodes are selected)

In a subsequent paper Dell (1988) proposed an additional networkencoding for all possible ldquowordshapesrdquo Each word in the lexical network isconnected to a wordshape header node that represents its CV structure Awordshape node sequentially activates segment category nodes that isonset consonant (COn) vowel (V) and coda consonant (CCo) Each of thecategory nodes connects back to all possible segment nodes of its category inthe lexical network Instead of selecting in parallel nodes for the OnsetNucleus and Coda slot of a frame the successive activation of the segmentcategory nodes leads to serial selection This allows the model to account forthe seriality effects in phonological encoding obtained by Meyer (19901991) using the so-called implicit priming paradigm These ndings will bediscussed below in the section on implicit priming

What is the empirical support for the explicit representation of CVstructure Empirical evidence concerning CV structure is scarce (for areview see Roelofs amp Meyer in press) Concerning speech errors there issome evidence for a tendency towards increased similarity of the CVstructures of the words involved For example segment additions tend tocreate clusters rather than singletons when the source word also has a

SYLLABIFICATION 663

cluster as in ldquoprich playerrdquo instead of ldquorich playerrdquo (Stemberger 1990) Theevidence is not however very strong First the CV similarity has beenobserved for onsets but not for nuclei Secondly the CV similarities areobserved for some but not all error corpora For example Stemberger(1984) observed them for small corpora of German and Swedish errors butnot for an English corpus

In terms of experimental evidence in two priming experiments withDutch participants Meijer (1994) observed that word production wasfacilitated by spoken primes with the same CV structure as the targetsrelative to primes that differed in their number of consonants However inanother experiment this effect was not replicated In a fourth experiment inwhich primes and targets had the same or different structures of the nucleus(V or VV) no priming effect from shared CV structure was obtained

Sevald Dell and Cole (1995) used a repeated pronunciation task to studythe representation of CV structure in English Participants had to producepairs of one monosyllabic and one disyllabic pseudoword as often as theycould within 4 sec The sequences were produced faster when themonosyllable had the same CV structure as the rst syllable of the disyllablethan when this was not the case as in kul-parfen versus kult-parfen Basedon these results Sevald et al argued for the explicit representation of CVstructure However as Sevald et al measured how many targets theparticipants could produce within a given time period it is difcult toestablish the locus of the effect of CV structure It could arise during thecreation of the phonological representation as Sevald et al propose but itcould also originate during the retrieval or execution of motor programs (seeRoelofs amp Meyer in press)

PROBLEMS WITH A MEMORY-BASED APPROACH

I now argue that a memory-based approach to syllabi cation (eg Dell 19861988 Houghton 1990 Shattuck-Hufnagel 1979) has difculty withsyllabi cation across morpheme boundaries and word boundaries Theproblem is that a segment of one morpheme or word may be syllabi ed withanother morpheme or word This may occur in the production ofpolymorphemic words or connected speech (eg Chomsky amp Halle 1968Kaisse 1985 Levelt 1989 1992 Nespor amp Vogel 1986 Selkirk 1984Spencer 1991) Models that rigidly store words as sequences of syllablenodes or models that store each consonant as an onset or coda have adifcult time dealing with the need for exibility of syllable membership

Cross-morpheme Syllabication

Consider the production of the plural form juwelen of the Dutch wordjuweel The plural juwelen is created by adding en ([]) to the stem juweel The resulting form is syllabi ed as (jy) s (we) s (l) s Thus

664 ROELOFS

juxtaposing en changes the syllabi cation of l This segment occupies acoda position in juweel syllabi ed as (jy) s (wel) s but an onset position injuwelen The syllabi cation across morpheme boundaries poses difculty tomodels that assign segments to syllable slots on a morpheme-by-morphemebasis

According to the model of Dell at the morpheme level the plural of theword juweel would be represented by the stem node juweel and a nodefor the sufx en The question is how the model handles thesyllabi cation of the l that is triggered by juxtaposing the sufx The stem juweel is connected to the coda l node In Dell (1986) this node will beselected to ll the coda slot of the second syllable frame However for theplural form the l should be the onset of the third syllable Thus the onset lnode has to be selected for the third syllable frame However onset l willnot be activated let alone be selected

The wordshape headers and the sequence of category nodes proposed byDell (1988) provide for the specication of a syllable structure that is uniqueto a particular combination of morphemes A header and sequence mightspecify that the fth segment of the plural form juwelen occupies an onsetposition instead of the coda position it has in juweel Of course a categorynode alone does not uniquely identify a segment Recall that a category nodeconnects back to all segments in its category Thus to be selected onset lmust also be activated by the syllable node len But what causes this node tobecome active One might propose to connect the syllable node len to themorpheme node juweel as its third syllable However this codes thewrong information because len is not the third syllable of juweel but thethird syllable of the plural juwelen Alternatively len might be connected to en as one of its ldquoallomorphsrdquo and also to juweel without beingexplicitly marked as its third syllable By not marking the syllable it may goldquounnotedrdquo in the encoding of syllables of juweel in producing thesingular form juweel To prevent interference with the encoding of ju and wein the production of juwelen the spreading rate between juweel and lenshould be sufciently small Then activating en activates all itsallomorphs including len The onset l now receives activation from juweel en and COn so it will be the most highly activated onset nodeand will be selected For the correct production of juwelen however this isnot sufcient Note that we is the second syllable of juwelen thus the syllablenode we should have a labelled link to juweel Therefore we need a (yetto be specied) mechanism that selects the appropriate second syllablenode wel in producing singular juweel and we in producing plural juwelen

In sum cross-morpheme syllabi cation in Dellrsquos model might be achievedby connecting the critical third syllable (eg len) to both morphemesinvolved and connecting the syllablersquos segments to category nodes (eglinking len to the stem juweel and the sufx en and linking onset l

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 659

morpheme boundaries and across word boundaries Furthermore I explorehow these problems might be solved within the framework of this model Inthe discussion of a rule-based approach I focus on the WEAVER model ofword-form encoding I will show how WEAVER handles syllabi cationacross morpheme boundaries and word boundaries Next I review theoutcomes of two series of experiments that support WEAVERrsquos specicview on syllabi cation Then I report a new experiment that contrasts Dellrsquosmodel and WEAVER Finally I discuss the issue of resyllabi cation

THEORETICAL POSITIONS

To set the stage I discuss the role of word-form encoding in the process oflexical access in speech production Three major types of processes underliespeaking conceptualisation formulation and articulation (eg Caplan 1992Garrett 1975 Kempen amp Hoenkamp 1987 Levelt 1989) Conceptual-isation processes generate so-called messages that is conceptual structuresto be verbally expressed In the view held by WEAVER messages arespecied in terms of lexical concepts and their relationships Formulationprocesses take the message access appropriate words for the lexical con-cepts (ie lemma retrieval and word-form encoding) and build a syntacticstructuremdashin the case of sentence productionmdashand a morphophonologicalstructure The result is an articulatory program for the utterance Finallyarticulation processes execute the articulatory program which results inovert speech

As indicated lexical access consists of two major steps lemma retrievaland word-form encoding corresponding to the formulation stages ofsyntactic encoding and morphophonological encoding respectively Inlemma retrieval a target lexical concept is used to retrieve the lemma of acorresponding word from memory Lemmas represent the syntacticproperties of words Lemma retrieval makes these properties available forsyntactic encoding processes In word-form encoding the morpho-phonological properties of the word are retrieved from memory to constructan articulatory program Below I describe a widely accepted psycho-linguistic view on lexical access (cf Levelt 1989 1992) which has beenadopted by WEAVER The encoding stages are uncontroversial but theassumptions about the computations and their output may differ betweenmodels

Assume that a Dutch speaker wants to name a gem Lexical access consistsof mapping the lexical concept G EM(X) onto the articulatory program forjuweel The lemma retriever takes GEM(X) and outputs the lemma of juweelTo derive the singular form [jyrsquowel] instead of the plural form [jyrsquowel] thewordrsquos number has to be specied The lemma plus this abstractmorphosyntactic number parameter are then input to word-form encoding

660 ROELOFS

FIG 1 Lexical access in speech production Lemma retrieval and word-form encoding (cf theWEAVER model Roelofs 1994 in press a)

The articulatory program is derived in three major steps (cf Dell 1986Levelt 1989 1992) morphological encoding phonological encoding andphonetic encoding (see Fig 1) The morphological encoder takes the lemmaof juweel plus the parameter singular and produces the stem morpheme juweel This rst stage thus concerns what is traditionally called theldquosyntaxndashmorphology interfacerdquo (eg Spencer 1991) The phonologicalencoder takes the stem morpheme and produces a so-called phonologicalword representation This representation describes the singular form ofjuweel as a phonological word consisting of two feet the rst one metricallyweak (w) and the second one strong (s) The rst syllable has j as onset andy as nucleus the second syllable has w as onset e as nucleus and l ascoda This second stage thus comprises what is traditionally called the

SYLLABIFICATION 661

ldquomorphologyndashphonology interfacerdquo (eg Goldsmith 1990) Finally thephonetic encoder takes this phonological word representation and deliversthe articulatory program [jyrsquowel] In WEAVER phonetic encodinginvolves accessing a store of learned motor programs for syllables aso-called syllabary (Levelt amp Wheeldon 1994) and setting parameters forpitch loudness and duration This nal encoding stage includes what issometimes called the ldquopost-lexical phonologyrdquo (eg Goldsmith 1990)

Dellrsquos model assumes stages of access similar to those of WEAVERdescribed above but there are also some important differences Forexample Dellrsquos model does not make the claim that the output ofphonological encoding is a phonological word representation and it does notassume a syllabary How is memory organised according to the models andwhat is the lexical access mechanism In particular how is syllabi cationachieved

DELLrsquoS MEMORY-BASED APPROACH

Dell (1986) assumes that word forms are represented in a lexical networkwith nodes for morphemes syllables rimes segments and segment clustersand features Figure 2 illustrates this for the Dutch word juweel (rimes andfeatures are omitted to simplify the gure) The nodes for the segments andclusters are marked for syllable position The nodes are associated with eachother by weighted bidirectional connections In phonological encoding theactivation level of a morpheme node is enhanced by giving it a jolt ofactivation Activation then spreads through the network each node sendinga proportion r of its activation to its direct neighbours Mapping a morphemeonto its segments is accomplished by selecting the highest activated segmentor cluster nodes The speaking rate determines the amount of time that isdevoted to the encoding of a syllable (ie the time between providing the joltof activation and segment selection) Selected nodes are inserted into slots ofindependentl y created syllable frames The segment nodes for a syllable areselected in parallel

In the case of multisyllabic morphemes the links between the morphemenode and the syllable nodes are labelled for the serial position of thesyllables The syllables are serially encoded The successive encoding ofsyllables is accomplished by temporarily increasing the spreading rate forone syllable and decreasing it for the others In the encoding of a bisyllabicmorpheme activation rst spreads with a rate of 15 r to the rst syllablenode and with 05 r to the second syllable node After the rst syllable framehas been lled the weights are changed Activation now spreads at a rate of05 r to the rst syllable node and at a rate of 15 r to the second syllable nodeThe changing of weights prevents selection of segments or clusters of the

662 ROELOFS

FIG 2 Memory representation of the word form of juweel (gem) in the model of Dell (19861988)

second syllable for the slots of the rst syllable frame or vice versa (recallthat the highest activated nodes are selected)

In a subsequent paper Dell (1988) proposed an additional networkencoding for all possible ldquowordshapesrdquo Each word in the lexical network isconnected to a wordshape header node that represents its CV structure Awordshape node sequentially activates segment category nodes that isonset consonant (COn) vowel (V) and coda consonant (CCo) Each of thecategory nodes connects back to all possible segment nodes of its category inthe lexical network Instead of selecting in parallel nodes for the OnsetNucleus and Coda slot of a frame the successive activation of the segmentcategory nodes leads to serial selection This allows the model to account forthe seriality effects in phonological encoding obtained by Meyer (19901991) using the so-called implicit priming paradigm These ndings will bediscussed below in the section on implicit priming

What is the empirical support for the explicit representation of CVstructure Empirical evidence concerning CV structure is scarce (for areview see Roelofs amp Meyer in press) Concerning speech errors there issome evidence for a tendency towards increased similarity of the CVstructures of the words involved For example segment additions tend tocreate clusters rather than singletons when the source word also has a

SYLLABIFICATION 663

cluster as in ldquoprich playerrdquo instead of ldquorich playerrdquo (Stemberger 1990) Theevidence is not however very strong First the CV similarity has beenobserved for onsets but not for nuclei Secondly the CV similarities areobserved for some but not all error corpora For example Stemberger(1984) observed them for small corpora of German and Swedish errors butnot for an English corpus

In terms of experimental evidence in two priming experiments withDutch participants Meijer (1994) observed that word production wasfacilitated by spoken primes with the same CV structure as the targetsrelative to primes that differed in their number of consonants However inanother experiment this effect was not replicated In a fourth experiment inwhich primes and targets had the same or different structures of the nucleus(V or VV) no priming effect from shared CV structure was obtained

Sevald Dell and Cole (1995) used a repeated pronunciation task to studythe representation of CV structure in English Participants had to producepairs of one monosyllabic and one disyllabic pseudoword as often as theycould within 4 sec The sequences were produced faster when themonosyllable had the same CV structure as the rst syllable of the disyllablethan when this was not the case as in kul-parfen versus kult-parfen Basedon these results Sevald et al argued for the explicit representation of CVstructure However as Sevald et al measured how many targets theparticipants could produce within a given time period it is difcult toestablish the locus of the effect of CV structure It could arise during thecreation of the phonological representation as Sevald et al propose but itcould also originate during the retrieval or execution of motor programs (seeRoelofs amp Meyer in press)

PROBLEMS WITH A MEMORY-BASED APPROACH

I now argue that a memory-based approach to syllabi cation (eg Dell 19861988 Houghton 1990 Shattuck-Hufnagel 1979) has difculty withsyllabi cation across morpheme boundaries and word boundaries Theproblem is that a segment of one morpheme or word may be syllabi ed withanother morpheme or word This may occur in the production ofpolymorphemic words or connected speech (eg Chomsky amp Halle 1968Kaisse 1985 Levelt 1989 1992 Nespor amp Vogel 1986 Selkirk 1984Spencer 1991) Models that rigidly store words as sequences of syllablenodes or models that store each consonant as an onset or coda have adifcult time dealing with the need for exibility of syllable membership

Cross-morpheme Syllabication

Consider the production of the plural form juwelen of the Dutch wordjuweel The plural juwelen is created by adding en ([]) to the stem juweel The resulting form is syllabi ed as (jy) s (we) s (l) s Thus

664 ROELOFS

juxtaposing en changes the syllabi cation of l This segment occupies acoda position in juweel syllabi ed as (jy) s (wel) s but an onset position injuwelen The syllabi cation across morpheme boundaries poses difculty tomodels that assign segments to syllable slots on a morpheme-by-morphemebasis

According to the model of Dell at the morpheme level the plural of theword juweel would be represented by the stem node juweel and a nodefor the sufx en The question is how the model handles thesyllabi cation of the l that is triggered by juxtaposing the sufx The stem juweel is connected to the coda l node In Dell (1986) this node will beselected to ll the coda slot of the second syllable frame However for theplural form the l should be the onset of the third syllable Thus the onset lnode has to be selected for the third syllable frame However onset l willnot be activated let alone be selected

The wordshape headers and the sequence of category nodes proposed byDell (1988) provide for the specication of a syllable structure that is uniqueto a particular combination of morphemes A header and sequence mightspecify that the fth segment of the plural form juwelen occupies an onsetposition instead of the coda position it has in juweel Of course a categorynode alone does not uniquely identify a segment Recall that a category nodeconnects back to all segments in its category Thus to be selected onset lmust also be activated by the syllable node len But what causes this node tobecome active One might propose to connect the syllable node len to themorpheme node juweel as its third syllable However this codes thewrong information because len is not the third syllable of juweel but thethird syllable of the plural juwelen Alternatively len might be connected to en as one of its ldquoallomorphsrdquo and also to juweel without beingexplicitly marked as its third syllable By not marking the syllable it may goldquounnotedrdquo in the encoding of syllables of juweel in producing thesingular form juweel To prevent interference with the encoding of ju and wein the production of juwelen the spreading rate between juweel and lenshould be sufciently small Then activating en activates all itsallomorphs including len The onset l now receives activation from juweel en and COn so it will be the most highly activated onset nodeand will be selected For the correct production of juwelen however this isnot sufcient Note that we is the second syllable of juwelen thus the syllablenode we should have a labelled link to juweel Therefore we need a (yetto be specied) mechanism that selects the appropriate second syllablenode wel in producing singular juweel and we in producing plural juwelen

In sum cross-morpheme syllabi cation in Dellrsquos model might be achievedby connecting the critical third syllable (eg len) to both morphemesinvolved and connecting the syllablersquos segments to category nodes (eglinking len to the stem juweel and the sufx en and linking onset l

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

660 ROELOFS

FIG 1 Lexical access in speech production Lemma retrieval and word-form encoding (cf theWEAVER model Roelofs 1994 in press a)

The articulatory program is derived in three major steps (cf Dell 1986Levelt 1989 1992) morphological encoding phonological encoding andphonetic encoding (see Fig 1) The morphological encoder takes the lemmaof juweel plus the parameter singular and produces the stem morpheme juweel This rst stage thus concerns what is traditionally called theldquosyntaxndashmorphology interfacerdquo (eg Spencer 1991) The phonologicalencoder takes the stem morpheme and produces a so-called phonologicalword representation This representation describes the singular form ofjuweel as a phonological word consisting of two feet the rst one metricallyweak (w) and the second one strong (s) The rst syllable has j as onset andy as nucleus the second syllable has w as onset e as nucleus and l ascoda This second stage thus comprises what is traditionally called the

SYLLABIFICATION 661

ldquomorphologyndashphonology interfacerdquo (eg Goldsmith 1990) Finally thephonetic encoder takes this phonological word representation and deliversthe articulatory program [jyrsquowel] In WEAVER phonetic encodinginvolves accessing a store of learned motor programs for syllables aso-called syllabary (Levelt amp Wheeldon 1994) and setting parameters forpitch loudness and duration This nal encoding stage includes what issometimes called the ldquopost-lexical phonologyrdquo (eg Goldsmith 1990)

Dellrsquos model assumes stages of access similar to those of WEAVERdescribed above but there are also some important differences Forexample Dellrsquos model does not make the claim that the output ofphonological encoding is a phonological word representation and it does notassume a syllabary How is memory organised according to the models andwhat is the lexical access mechanism In particular how is syllabi cationachieved

DELLrsquoS MEMORY-BASED APPROACH

Dell (1986) assumes that word forms are represented in a lexical networkwith nodes for morphemes syllables rimes segments and segment clustersand features Figure 2 illustrates this for the Dutch word juweel (rimes andfeatures are omitted to simplify the gure) The nodes for the segments andclusters are marked for syllable position The nodes are associated with eachother by weighted bidirectional connections In phonological encoding theactivation level of a morpheme node is enhanced by giving it a jolt ofactivation Activation then spreads through the network each node sendinga proportion r of its activation to its direct neighbours Mapping a morphemeonto its segments is accomplished by selecting the highest activated segmentor cluster nodes The speaking rate determines the amount of time that isdevoted to the encoding of a syllable (ie the time between providing the joltof activation and segment selection) Selected nodes are inserted into slots ofindependentl y created syllable frames The segment nodes for a syllable areselected in parallel

In the case of multisyllabic morphemes the links between the morphemenode and the syllable nodes are labelled for the serial position of thesyllables The syllables are serially encoded The successive encoding ofsyllables is accomplished by temporarily increasing the spreading rate forone syllable and decreasing it for the others In the encoding of a bisyllabicmorpheme activation rst spreads with a rate of 15 r to the rst syllablenode and with 05 r to the second syllable node After the rst syllable framehas been lled the weights are changed Activation now spreads at a rate of05 r to the rst syllable node and at a rate of 15 r to the second syllable nodeThe changing of weights prevents selection of segments or clusters of the

662 ROELOFS

FIG 2 Memory representation of the word form of juweel (gem) in the model of Dell (19861988)

second syllable for the slots of the rst syllable frame or vice versa (recallthat the highest activated nodes are selected)

In a subsequent paper Dell (1988) proposed an additional networkencoding for all possible ldquowordshapesrdquo Each word in the lexical network isconnected to a wordshape header node that represents its CV structure Awordshape node sequentially activates segment category nodes that isonset consonant (COn) vowel (V) and coda consonant (CCo) Each of thecategory nodes connects back to all possible segment nodes of its category inthe lexical network Instead of selecting in parallel nodes for the OnsetNucleus and Coda slot of a frame the successive activation of the segmentcategory nodes leads to serial selection This allows the model to account forthe seriality effects in phonological encoding obtained by Meyer (19901991) using the so-called implicit priming paradigm These ndings will bediscussed below in the section on implicit priming

What is the empirical support for the explicit representation of CVstructure Empirical evidence concerning CV structure is scarce (for areview see Roelofs amp Meyer in press) Concerning speech errors there issome evidence for a tendency towards increased similarity of the CVstructures of the words involved For example segment additions tend tocreate clusters rather than singletons when the source word also has a

SYLLABIFICATION 663

cluster as in ldquoprich playerrdquo instead of ldquorich playerrdquo (Stemberger 1990) Theevidence is not however very strong First the CV similarity has beenobserved for onsets but not for nuclei Secondly the CV similarities areobserved for some but not all error corpora For example Stemberger(1984) observed them for small corpora of German and Swedish errors butnot for an English corpus

In terms of experimental evidence in two priming experiments withDutch participants Meijer (1994) observed that word production wasfacilitated by spoken primes with the same CV structure as the targetsrelative to primes that differed in their number of consonants However inanother experiment this effect was not replicated In a fourth experiment inwhich primes and targets had the same or different structures of the nucleus(V or VV) no priming effect from shared CV structure was obtained

Sevald Dell and Cole (1995) used a repeated pronunciation task to studythe representation of CV structure in English Participants had to producepairs of one monosyllabic and one disyllabic pseudoword as often as theycould within 4 sec The sequences were produced faster when themonosyllable had the same CV structure as the rst syllable of the disyllablethan when this was not the case as in kul-parfen versus kult-parfen Basedon these results Sevald et al argued for the explicit representation of CVstructure However as Sevald et al measured how many targets theparticipants could produce within a given time period it is difcult toestablish the locus of the effect of CV structure It could arise during thecreation of the phonological representation as Sevald et al propose but itcould also originate during the retrieval or execution of motor programs (seeRoelofs amp Meyer in press)

PROBLEMS WITH A MEMORY-BASED APPROACH

I now argue that a memory-based approach to syllabi cation (eg Dell 19861988 Houghton 1990 Shattuck-Hufnagel 1979) has difculty withsyllabi cation across morpheme boundaries and word boundaries Theproblem is that a segment of one morpheme or word may be syllabi ed withanother morpheme or word This may occur in the production ofpolymorphemic words or connected speech (eg Chomsky amp Halle 1968Kaisse 1985 Levelt 1989 1992 Nespor amp Vogel 1986 Selkirk 1984Spencer 1991) Models that rigidly store words as sequences of syllablenodes or models that store each consonant as an onset or coda have adifcult time dealing with the need for exibility of syllable membership

Cross-morpheme Syllabication

Consider the production of the plural form juwelen of the Dutch wordjuweel The plural juwelen is created by adding en ([]) to the stem juweel The resulting form is syllabi ed as (jy) s (we) s (l) s Thus

664 ROELOFS

juxtaposing en changes the syllabi cation of l This segment occupies acoda position in juweel syllabi ed as (jy) s (wel) s but an onset position injuwelen The syllabi cation across morpheme boundaries poses difculty tomodels that assign segments to syllable slots on a morpheme-by-morphemebasis

According to the model of Dell at the morpheme level the plural of theword juweel would be represented by the stem node juweel and a nodefor the sufx en The question is how the model handles thesyllabi cation of the l that is triggered by juxtaposing the sufx The stem juweel is connected to the coda l node In Dell (1986) this node will beselected to ll the coda slot of the second syllable frame However for theplural form the l should be the onset of the third syllable Thus the onset lnode has to be selected for the third syllable frame However onset l willnot be activated let alone be selected

The wordshape headers and the sequence of category nodes proposed byDell (1988) provide for the specication of a syllable structure that is uniqueto a particular combination of morphemes A header and sequence mightspecify that the fth segment of the plural form juwelen occupies an onsetposition instead of the coda position it has in juweel Of course a categorynode alone does not uniquely identify a segment Recall that a category nodeconnects back to all segments in its category Thus to be selected onset lmust also be activated by the syllable node len But what causes this node tobecome active One might propose to connect the syllable node len to themorpheme node juweel as its third syllable However this codes thewrong information because len is not the third syllable of juweel but thethird syllable of the plural juwelen Alternatively len might be connected to en as one of its ldquoallomorphsrdquo and also to juweel without beingexplicitly marked as its third syllable By not marking the syllable it may goldquounnotedrdquo in the encoding of syllables of juweel in producing thesingular form juweel To prevent interference with the encoding of ju and wein the production of juwelen the spreading rate between juweel and lenshould be sufciently small Then activating en activates all itsallomorphs including len The onset l now receives activation from juweel en and COn so it will be the most highly activated onset nodeand will be selected For the correct production of juwelen however this isnot sufcient Note that we is the second syllable of juwelen thus the syllablenode we should have a labelled link to juweel Therefore we need a (yetto be specied) mechanism that selects the appropriate second syllablenode wel in producing singular juweel and we in producing plural juwelen

In sum cross-morpheme syllabi cation in Dellrsquos model might be achievedby connecting the critical third syllable (eg len) to both morphemesinvolved and connecting the syllablersquos segments to category nodes (eglinking len to the stem juweel and the sufx en and linking onset l

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 661

ldquomorphologyndashphonology interfacerdquo (eg Goldsmith 1990) Finally thephonetic encoder takes this phonological word representation and deliversthe articulatory program [jyrsquowel] In WEAVER phonetic encodinginvolves accessing a store of learned motor programs for syllables aso-called syllabary (Levelt amp Wheeldon 1994) and setting parameters forpitch loudness and duration This nal encoding stage includes what issometimes called the ldquopost-lexical phonologyrdquo (eg Goldsmith 1990)

Dellrsquos model assumes stages of access similar to those of WEAVERdescribed above but there are also some important differences Forexample Dellrsquos model does not make the claim that the output ofphonological encoding is a phonological word representation and it does notassume a syllabary How is memory organised according to the models andwhat is the lexical access mechanism In particular how is syllabi cationachieved

DELLrsquoS MEMORY-BASED APPROACH

Dell (1986) assumes that word forms are represented in a lexical networkwith nodes for morphemes syllables rimes segments and segment clustersand features Figure 2 illustrates this for the Dutch word juweel (rimes andfeatures are omitted to simplify the gure) The nodes for the segments andclusters are marked for syllable position The nodes are associated with eachother by weighted bidirectional connections In phonological encoding theactivation level of a morpheme node is enhanced by giving it a jolt ofactivation Activation then spreads through the network each node sendinga proportion r of its activation to its direct neighbours Mapping a morphemeonto its segments is accomplished by selecting the highest activated segmentor cluster nodes The speaking rate determines the amount of time that isdevoted to the encoding of a syllable (ie the time between providing the joltof activation and segment selection) Selected nodes are inserted into slots ofindependentl y created syllable frames The segment nodes for a syllable areselected in parallel

In the case of multisyllabic morphemes the links between the morphemenode and the syllable nodes are labelled for the serial position of thesyllables The syllables are serially encoded The successive encoding ofsyllables is accomplished by temporarily increasing the spreading rate forone syllable and decreasing it for the others In the encoding of a bisyllabicmorpheme activation rst spreads with a rate of 15 r to the rst syllablenode and with 05 r to the second syllable node After the rst syllable framehas been lled the weights are changed Activation now spreads at a rate of05 r to the rst syllable node and at a rate of 15 r to the second syllable nodeThe changing of weights prevents selection of segments or clusters of the

662 ROELOFS

FIG 2 Memory representation of the word form of juweel (gem) in the model of Dell (19861988)

second syllable for the slots of the rst syllable frame or vice versa (recallthat the highest activated nodes are selected)

In a subsequent paper Dell (1988) proposed an additional networkencoding for all possible ldquowordshapesrdquo Each word in the lexical network isconnected to a wordshape header node that represents its CV structure Awordshape node sequentially activates segment category nodes that isonset consonant (COn) vowel (V) and coda consonant (CCo) Each of thecategory nodes connects back to all possible segment nodes of its category inthe lexical network Instead of selecting in parallel nodes for the OnsetNucleus and Coda slot of a frame the successive activation of the segmentcategory nodes leads to serial selection This allows the model to account forthe seriality effects in phonological encoding obtained by Meyer (19901991) using the so-called implicit priming paradigm These ndings will bediscussed below in the section on implicit priming

What is the empirical support for the explicit representation of CVstructure Empirical evidence concerning CV structure is scarce (for areview see Roelofs amp Meyer in press) Concerning speech errors there issome evidence for a tendency towards increased similarity of the CVstructures of the words involved For example segment additions tend tocreate clusters rather than singletons when the source word also has a

SYLLABIFICATION 663

cluster as in ldquoprich playerrdquo instead of ldquorich playerrdquo (Stemberger 1990) Theevidence is not however very strong First the CV similarity has beenobserved for onsets but not for nuclei Secondly the CV similarities areobserved for some but not all error corpora For example Stemberger(1984) observed them for small corpora of German and Swedish errors butnot for an English corpus

In terms of experimental evidence in two priming experiments withDutch participants Meijer (1994) observed that word production wasfacilitated by spoken primes with the same CV structure as the targetsrelative to primes that differed in their number of consonants However inanother experiment this effect was not replicated In a fourth experiment inwhich primes and targets had the same or different structures of the nucleus(V or VV) no priming effect from shared CV structure was obtained

Sevald Dell and Cole (1995) used a repeated pronunciation task to studythe representation of CV structure in English Participants had to producepairs of one monosyllabic and one disyllabic pseudoword as often as theycould within 4 sec The sequences were produced faster when themonosyllable had the same CV structure as the rst syllable of the disyllablethan when this was not the case as in kul-parfen versus kult-parfen Basedon these results Sevald et al argued for the explicit representation of CVstructure However as Sevald et al measured how many targets theparticipants could produce within a given time period it is difcult toestablish the locus of the effect of CV structure It could arise during thecreation of the phonological representation as Sevald et al propose but itcould also originate during the retrieval or execution of motor programs (seeRoelofs amp Meyer in press)

PROBLEMS WITH A MEMORY-BASED APPROACH

I now argue that a memory-based approach to syllabi cation (eg Dell 19861988 Houghton 1990 Shattuck-Hufnagel 1979) has difculty withsyllabi cation across morpheme boundaries and word boundaries Theproblem is that a segment of one morpheme or word may be syllabi ed withanother morpheme or word This may occur in the production ofpolymorphemic words or connected speech (eg Chomsky amp Halle 1968Kaisse 1985 Levelt 1989 1992 Nespor amp Vogel 1986 Selkirk 1984Spencer 1991) Models that rigidly store words as sequences of syllablenodes or models that store each consonant as an onset or coda have adifcult time dealing with the need for exibility of syllable membership

Cross-morpheme Syllabication

Consider the production of the plural form juwelen of the Dutch wordjuweel The plural juwelen is created by adding en ([]) to the stem juweel The resulting form is syllabi ed as (jy) s (we) s (l) s Thus

664 ROELOFS

juxtaposing en changes the syllabi cation of l This segment occupies acoda position in juweel syllabi ed as (jy) s (wel) s but an onset position injuwelen The syllabi cation across morpheme boundaries poses difculty tomodels that assign segments to syllable slots on a morpheme-by-morphemebasis

According to the model of Dell at the morpheme level the plural of theword juweel would be represented by the stem node juweel and a nodefor the sufx en The question is how the model handles thesyllabi cation of the l that is triggered by juxtaposing the sufx The stem juweel is connected to the coda l node In Dell (1986) this node will beselected to ll the coda slot of the second syllable frame However for theplural form the l should be the onset of the third syllable Thus the onset lnode has to be selected for the third syllable frame However onset l willnot be activated let alone be selected

The wordshape headers and the sequence of category nodes proposed byDell (1988) provide for the specication of a syllable structure that is uniqueto a particular combination of morphemes A header and sequence mightspecify that the fth segment of the plural form juwelen occupies an onsetposition instead of the coda position it has in juweel Of course a categorynode alone does not uniquely identify a segment Recall that a category nodeconnects back to all segments in its category Thus to be selected onset lmust also be activated by the syllable node len But what causes this node tobecome active One might propose to connect the syllable node len to themorpheme node juweel as its third syllable However this codes thewrong information because len is not the third syllable of juweel but thethird syllable of the plural juwelen Alternatively len might be connected to en as one of its ldquoallomorphsrdquo and also to juweel without beingexplicitly marked as its third syllable By not marking the syllable it may goldquounnotedrdquo in the encoding of syllables of juweel in producing thesingular form juweel To prevent interference with the encoding of ju and wein the production of juwelen the spreading rate between juweel and lenshould be sufciently small Then activating en activates all itsallomorphs including len The onset l now receives activation from juweel en and COn so it will be the most highly activated onset nodeand will be selected For the correct production of juwelen however this isnot sufcient Note that we is the second syllable of juwelen thus the syllablenode we should have a labelled link to juweel Therefore we need a (yetto be specied) mechanism that selects the appropriate second syllablenode wel in producing singular juweel and we in producing plural juwelen

In sum cross-morpheme syllabi cation in Dellrsquos model might be achievedby connecting the critical third syllable (eg len) to both morphemesinvolved and connecting the syllablersquos segments to category nodes (eglinking len to the stem juweel and the sufx en and linking onset l

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

662 ROELOFS

FIG 2 Memory representation of the word form of juweel (gem) in the model of Dell (19861988)

second syllable for the slots of the rst syllable frame or vice versa (recallthat the highest activated nodes are selected)

In a subsequent paper Dell (1988) proposed an additional networkencoding for all possible ldquowordshapesrdquo Each word in the lexical network isconnected to a wordshape header node that represents its CV structure Awordshape node sequentially activates segment category nodes that isonset consonant (COn) vowel (V) and coda consonant (CCo) Each of thecategory nodes connects back to all possible segment nodes of its category inthe lexical network Instead of selecting in parallel nodes for the OnsetNucleus and Coda slot of a frame the successive activation of the segmentcategory nodes leads to serial selection This allows the model to account forthe seriality effects in phonological encoding obtained by Meyer (19901991) using the so-called implicit priming paradigm These ndings will bediscussed below in the section on implicit priming

What is the empirical support for the explicit representation of CVstructure Empirical evidence concerning CV structure is scarce (for areview see Roelofs amp Meyer in press) Concerning speech errors there issome evidence for a tendency towards increased similarity of the CVstructures of the words involved For example segment additions tend tocreate clusters rather than singletons when the source word also has a

SYLLABIFICATION 663

cluster as in ldquoprich playerrdquo instead of ldquorich playerrdquo (Stemberger 1990) Theevidence is not however very strong First the CV similarity has beenobserved for onsets but not for nuclei Secondly the CV similarities areobserved for some but not all error corpora For example Stemberger(1984) observed them for small corpora of German and Swedish errors butnot for an English corpus

In terms of experimental evidence in two priming experiments withDutch participants Meijer (1994) observed that word production wasfacilitated by spoken primes with the same CV structure as the targetsrelative to primes that differed in their number of consonants However inanother experiment this effect was not replicated In a fourth experiment inwhich primes and targets had the same or different structures of the nucleus(V or VV) no priming effect from shared CV structure was obtained

Sevald Dell and Cole (1995) used a repeated pronunciation task to studythe representation of CV structure in English Participants had to producepairs of one monosyllabic and one disyllabic pseudoword as often as theycould within 4 sec The sequences were produced faster when themonosyllable had the same CV structure as the rst syllable of the disyllablethan when this was not the case as in kul-parfen versus kult-parfen Basedon these results Sevald et al argued for the explicit representation of CVstructure However as Sevald et al measured how many targets theparticipants could produce within a given time period it is difcult toestablish the locus of the effect of CV structure It could arise during thecreation of the phonological representation as Sevald et al propose but itcould also originate during the retrieval or execution of motor programs (seeRoelofs amp Meyer in press)

PROBLEMS WITH A MEMORY-BASED APPROACH

I now argue that a memory-based approach to syllabi cation (eg Dell 19861988 Houghton 1990 Shattuck-Hufnagel 1979) has difculty withsyllabi cation across morpheme boundaries and word boundaries Theproblem is that a segment of one morpheme or word may be syllabi ed withanother morpheme or word This may occur in the production ofpolymorphemic words or connected speech (eg Chomsky amp Halle 1968Kaisse 1985 Levelt 1989 1992 Nespor amp Vogel 1986 Selkirk 1984Spencer 1991) Models that rigidly store words as sequences of syllablenodes or models that store each consonant as an onset or coda have adifcult time dealing with the need for exibility of syllable membership

Cross-morpheme Syllabication

Consider the production of the plural form juwelen of the Dutch wordjuweel The plural juwelen is created by adding en ([]) to the stem juweel The resulting form is syllabi ed as (jy) s (we) s (l) s Thus

664 ROELOFS

juxtaposing en changes the syllabi cation of l This segment occupies acoda position in juweel syllabi ed as (jy) s (wel) s but an onset position injuwelen The syllabi cation across morpheme boundaries poses difculty tomodels that assign segments to syllable slots on a morpheme-by-morphemebasis

According to the model of Dell at the morpheme level the plural of theword juweel would be represented by the stem node juweel and a nodefor the sufx en The question is how the model handles thesyllabi cation of the l that is triggered by juxtaposing the sufx The stem juweel is connected to the coda l node In Dell (1986) this node will beselected to ll the coda slot of the second syllable frame However for theplural form the l should be the onset of the third syllable Thus the onset lnode has to be selected for the third syllable frame However onset l willnot be activated let alone be selected

The wordshape headers and the sequence of category nodes proposed byDell (1988) provide for the specication of a syllable structure that is uniqueto a particular combination of morphemes A header and sequence mightspecify that the fth segment of the plural form juwelen occupies an onsetposition instead of the coda position it has in juweel Of course a categorynode alone does not uniquely identify a segment Recall that a category nodeconnects back to all segments in its category Thus to be selected onset lmust also be activated by the syllable node len But what causes this node tobecome active One might propose to connect the syllable node len to themorpheme node juweel as its third syllable However this codes thewrong information because len is not the third syllable of juweel but thethird syllable of the plural juwelen Alternatively len might be connected to en as one of its ldquoallomorphsrdquo and also to juweel without beingexplicitly marked as its third syllable By not marking the syllable it may goldquounnotedrdquo in the encoding of syllables of juweel in producing thesingular form juweel To prevent interference with the encoding of ju and wein the production of juwelen the spreading rate between juweel and lenshould be sufciently small Then activating en activates all itsallomorphs including len The onset l now receives activation from juweel en and COn so it will be the most highly activated onset nodeand will be selected For the correct production of juwelen however this isnot sufcient Note that we is the second syllable of juwelen thus the syllablenode we should have a labelled link to juweel Therefore we need a (yetto be specied) mechanism that selects the appropriate second syllablenode wel in producing singular juweel and we in producing plural juwelen

In sum cross-morpheme syllabi cation in Dellrsquos model might be achievedby connecting the critical third syllable (eg len) to both morphemesinvolved and connecting the syllablersquos segments to category nodes (eglinking len to the stem juweel and the sufx en and linking onset l

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 663

cluster as in ldquoprich playerrdquo instead of ldquorich playerrdquo (Stemberger 1990) Theevidence is not however very strong First the CV similarity has beenobserved for onsets but not for nuclei Secondly the CV similarities areobserved for some but not all error corpora For example Stemberger(1984) observed them for small corpora of German and Swedish errors butnot for an English corpus

In terms of experimental evidence in two priming experiments withDutch participants Meijer (1994) observed that word production wasfacilitated by spoken primes with the same CV structure as the targetsrelative to primes that differed in their number of consonants However inanother experiment this effect was not replicated In a fourth experiment inwhich primes and targets had the same or different structures of the nucleus(V or VV) no priming effect from shared CV structure was obtained

Sevald Dell and Cole (1995) used a repeated pronunciation task to studythe representation of CV structure in English Participants had to producepairs of one monosyllabic and one disyllabic pseudoword as often as theycould within 4 sec The sequences were produced faster when themonosyllable had the same CV structure as the rst syllable of the disyllablethan when this was not the case as in kul-parfen versus kult-parfen Basedon these results Sevald et al argued for the explicit representation of CVstructure However as Sevald et al measured how many targets theparticipants could produce within a given time period it is difcult toestablish the locus of the effect of CV structure It could arise during thecreation of the phonological representation as Sevald et al propose but itcould also originate during the retrieval or execution of motor programs (seeRoelofs amp Meyer in press)

PROBLEMS WITH A MEMORY-BASED APPROACH

I now argue that a memory-based approach to syllabi cation (eg Dell 19861988 Houghton 1990 Shattuck-Hufnagel 1979) has difculty withsyllabi cation across morpheme boundaries and word boundaries Theproblem is that a segment of one morpheme or word may be syllabi ed withanother morpheme or word This may occur in the production ofpolymorphemic words or connected speech (eg Chomsky amp Halle 1968Kaisse 1985 Levelt 1989 1992 Nespor amp Vogel 1986 Selkirk 1984Spencer 1991) Models that rigidly store words as sequences of syllablenodes or models that store each consonant as an onset or coda have adifcult time dealing with the need for exibility of syllable membership

Cross-morpheme Syllabication

Consider the production of the plural form juwelen of the Dutch wordjuweel The plural juwelen is created by adding en ([]) to the stem juweel The resulting form is syllabi ed as (jy) s (we) s (l) s Thus

664 ROELOFS

juxtaposing en changes the syllabi cation of l This segment occupies acoda position in juweel syllabi ed as (jy) s (wel) s but an onset position injuwelen The syllabi cation across morpheme boundaries poses difculty tomodels that assign segments to syllable slots on a morpheme-by-morphemebasis

According to the model of Dell at the morpheme level the plural of theword juweel would be represented by the stem node juweel and a nodefor the sufx en The question is how the model handles thesyllabi cation of the l that is triggered by juxtaposing the sufx The stem juweel is connected to the coda l node In Dell (1986) this node will beselected to ll the coda slot of the second syllable frame However for theplural form the l should be the onset of the third syllable Thus the onset lnode has to be selected for the third syllable frame However onset l willnot be activated let alone be selected

The wordshape headers and the sequence of category nodes proposed byDell (1988) provide for the specication of a syllable structure that is uniqueto a particular combination of morphemes A header and sequence mightspecify that the fth segment of the plural form juwelen occupies an onsetposition instead of the coda position it has in juweel Of course a categorynode alone does not uniquely identify a segment Recall that a category nodeconnects back to all segments in its category Thus to be selected onset lmust also be activated by the syllable node len But what causes this node tobecome active One might propose to connect the syllable node len to themorpheme node juweel as its third syllable However this codes thewrong information because len is not the third syllable of juweel but thethird syllable of the plural juwelen Alternatively len might be connected to en as one of its ldquoallomorphsrdquo and also to juweel without beingexplicitly marked as its third syllable By not marking the syllable it may goldquounnotedrdquo in the encoding of syllables of juweel in producing thesingular form juweel To prevent interference with the encoding of ju and wein the production of juwelen the spreading rate between juweel and lenshould be sufciently small Then activating en activates all itsallomorphs including len The onset l now receives activation from juweel en and COn so it will be the most highly activated onset nodeand will be selected For the correct production of juwelen however this isnot sufcient Note that we is the second syllable of juwelen thus the syllablenode we should have a labelled link to juweel Therefore we need a (yetto be specied) mechanism that selects the appropriate second syllablenode wel in producing singular juweel and we in producing plural juwelen

In sum cross-morpheme syllabi cation in Dellrsquos model might be achievedby connecting the critical third syllable (eg len) to both morphemesinvolved and connecting the syllablersquos segments to category nodes (eglinking len to the stem juweel and the sufx en and linking onset l

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

664 ROELOFS

juxtaposing en changes the syllabi cation of l This segment occupies acoda position in juweel syllabi ed as (jy) s (wel) s but an onset position injuwelen The syllabi cation across morpheme boundaries poses difculty tomodels that assign segments to syllable slots on a morpheme-by-morphemebasis

According to the model of Dell at the morpheme level the plural of theword juweel would be represented by the stem node juweel and a nodefor the sufx en The question is how the model handles thesyllabi cation of the l that is triggered by juxtaposing the sufx The stem juweel is connected to the coda l node In Dell (1986) this node will beselected to ll the coda slot of the second syllable frame However for theplural form the l should be the onset of the third syllable Thus the onset lnode has to be selected for the third syllable frame However onset l willnot be activated let alone be selected

The wordshape headers and the sequence of category nodes proposed byDell (1988) provide for the specication of a syllable structure that is uniqueto a particular combination of morphemes A header and sequence mightspecify that the fth segment of the plural form juwelen occupies an onsetposition instead of the coda position it has in juweel Of course a categorynode alone does not uniquely identify a segment Recall that a category nodeconnects back to all segments in its category Thus to be selected onset lmust also be activated by the syllable node len But what causes this node tobecome active One might propose to connect the syllable node len to themorpheme node juweel as its third syllable However this codes thewrong information because len is not the third syllable of juweel but thethird syllable of the plural juwelen Alternatively len might be connected to en as one of its ldquoallomorphsrdquo and also to juweel without beingexplicitly marked as its third syllable By not marking the syllable it may goldquounnotedrdquo in the encoding of syllables of juweel in producing thesingular form juweel To prevent interference with the encoding of ju and wein the production of juwelen the spreading rate between juweel and lenshould be sufciently small Then activating en activates all itsallomorphs including len The onset l now receives activation from juweel en and COn so it will be the most highly activated onset nodeand will be selected For the correct production of juwelen however this isnot sufcient Note that we is the second syllable of juwelen thus the syllablenode we should have a labelled link to juweel Therefore we need a (yetto be specied) mechanism that selects the appropriate second syllablenode wel in producing singular juweel and we in producing plural juwelen

In sum cross-morpheme syllabi cation in Dellrsquos model might be achievedby connecting the critical third syllable (eg len) to both morphemesinvolved and connecting the syllablersquos segments to category nodes (eglinking len to the stem juweel and the sufx en and linking onset l

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 665

to COn) Also we have to postulate a mechanism that selects the correctsecond syllable node This wordshape network solution will be empiricallytested in the new experiment reported below

The new suggestion for cross-morpheme syllabi cation in Dellrsquos modelmay seem rather complicated and ad hoc Instead one might propose tostore complex forms not in a morphological ly decomposed way but as aunitary representation for example a single ldquomorphemerdquo node juwelen for the plural juwelen There is some empirical evidence from speech errorsthat suggests that inected forms of high frequency are stored in such amanner (Stemberger amp MacWhinney 1986 but see Roelofs amp Meyer inprep) Still the problem for items of lower frequency would remainFurthermore the problems in syllabi cation do not only arise with inectedforms but also with derivations For example juwelier ( juweller) issyllabi ed as (jy) s (w) s (lir) s Again after juxtaposing a sufx the l ofjuweel no longer occupies a coda position but becomes a syllable onset Theoption of unitary representations does not seem to be available herebecause empirical evidence from speech errors (eg Dell 1986 Garrett1975 1980 1988 Levelt 1989 Stemberger 1985) and production latencies(Roelofs 1996 ab Roelofs Baayen amp Van den Brink submitted) suggeststhat derivational forms are stored in a morphologically decomposedmanner I will discuss the speech-error evidence now and the latencyevidence later

The evidence from speech errors for morphological decompositionconcerns failures in the serial ordering of morphemes in an utterance Somemorphemic errors seem to concern the lemma level whereas others involvethe word-form level (eg Dell 1986 Garrett 1980 1988) Consider forexample the morphemic error in ldquohow many PIEs does it take to make anAPPLErdquo (from Garrett 1988) The interacting stems belong to the samesyntactic category (ie noun) and come from distinct phrases which is alsocharacteristic of whole-word exchanges such as in ldquowe completely forgot toadd the LIST to the ROOFrdquo (from Garrett 1980) Whole-word exchangestypically involve items of the same syntactic category and ignore phraseboundaries (Garrett 1975) This suggests that these morpheme errors andwhole-word errors occur at the same level of processing They occur duringsyntactic encoding when lemmas in a developing syntactic structure tradeplaces Note that the plurality of apple is stranded that is it is realised on pieThis suggests that the abstract morphosyntactic number parameter is setafter the exchange By contrast the exchanging morphemes in an error suchas ldquoSLICEly THINNedrdquo (from Stemberger 1982) belong to differentsyntactic categories (adjective and verb) and come from the same phrasewhich is also characteristic of segment exchanges such as in ldquoRack Patrdquo forldquopack ratrdquo (from Garrett 1988) Segment errors are typically not affected bylemma information such as syntactic category and occur on words within a

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

666 ROELOFS

single phrase This suggests that this second type of morpheme error andsegment errors occur at the same level of processing namely the level atwhich word forms are retrieved and the morphophonological form of theutterance is constructed The errors occur when morphemes or segments in adeveloping morphophonological structure trade places

In the classication of speech errors a distinction is made betweencontextual and non-contextual errors Contextual errors involve amisordering within the intended utterance whereas for non-contextualerrors there does not exist a clear source within the utterance As concernsderivational forms morphemes of both prexed words and compounds areinvolved in speech errors (all examples of errors below are from Stemberger1985) Examples of contextual errors involving prexes are the anticipationerror ldquowe have twenty-ve DEdollars deductible rdquo (intended as ldquowe havetwenty-ve dollars deductible rdquo) the perseveratio n error ldquoit does notexplain how an apparent case of rule EXsertion may ariserdquo (intended as ldquoitdoes not explain how an apparent case of rule insertion may ariserdquo) and theexchange error ldquoa self-INstruct DE rdquo (intended as ldquoa self-destructinstructionrdquo) These errors involve words of different syntactic classeswhich suggests that the errors are due to encoding failures at the word-formlevel Examples of non-contextual errors involving prexes are thesubstitution error ldquoshersquos so EXquisitiverdquo (intended as ldquoshersquos so inquisitiverdquo)the addition error ldquopositively or negatively REmarked as rdquo (intended asldquopositively or negatively marked as rdquo) and the deletion error ldquotheywerenrsquot jealrdquo (intended as ldquothey werenrsquot conjealingrdquo) Similar errorsinvolving prexes have also been observed for Dutch These errors aredifcult to explain purely in phonological terms because phonologicalerrors rarely involve more than a single segment or syllable constituent (egDell 1986 Stemberger 1985) Speech error evidence also suggests thatcompounds have internal morphological structure in the mental lexiconExamples of misorderings are ldquooh you were just closing the LIDBOXesrdquo(intended as ldquooh you were just closing the boxlidsrdquo) and ldquodid we miss theTURN TRAIL-offrdquo (intended as ldquodid we miss the trail turn-offrdquo) Againdue to the large number of segments involved these errors cannot beexplained phonological ly

One objection to the speech error evidence may be that speech errors arerare events By denition speech errors reect unusual circumstances thatcannot straightforwardly be taken to represent the norm It may be possiblethat speakers normally do not assemble a wordrsquos form out of its constituentmorphemes but start to make errors when they occasionally try to do soBelow in a section on morphological structure I will review ndings from achronometric paradigm that independentl y support the conclusion that thelexicon contains morphological ly decomposed word forms

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 667

In summary the distributional characteristics of certain morpheme errorsare similar to those of segment errors and differ from those of whole-worderrors This suggests that the lexicon contains morphologically decomposedword forms Thus the empirical evidence refutes the idea of solving thecross-morpheme syllabi cation problem by a single ldquomorphemerdquo node for amorphological ly complex form Something like the wordshape-networksolution is needed for Dellrsquos model

Cross-word Syllabication

The wordshape-network suggestion for syllabi cation across morphemeboundaries in Dellrsquos model however cannot be a general solution It breaksdown on syllabi cation across word boundaries For cross-wordsyllabi cation header nodes for combinations of words rather than forcombinations of morphemes would be needed Cross-word syllabi cationhowever often concerns combinations of open-class items and closed-classitems Since the combinations involve open-class words a speaker cannotstore header nodes and category sequences for all alliances Thus thesuggestion above for syllabi cation across morpheme boundaries cannot beused as a general solution to the problem of syllabi cation across wordboundaries

Consider for example the production of the utterance ldquoleg mijn juweel inde doosrdquo (ldquoput my gem in the boxrdquo) To increase the speed and uency ofarticulation a speaker might syllabify the utterance as ldquo (jy) s (we) s (lIn) s

rdquo instead of ldquo (jy) s (wel) s (In) s rdquo The coda l of the second syllable ofjuweel is made the onset of the syllable lin The lexical words juweel and inare combined into the new phonological word juwelin The creation of a newphonological word is optional in this example but it is obligatory for cliticsClitics are function words such as pronouns determiners particlesauxiliary verbs prepositions and conjunctions which unlike words of majorlexical categories are phonologically dependent on a host (eg Booij 1995Levelt 1989) For example the reduced form rsquos [s] of the Dutch adverb eens(now) cannot stand alone In producing ldquogeef me het juweel rsquosrdquo (ldquoplease giveme the gem nowrdquo) rsquos is adjoined to juweel This yields the new phonologicalword juweel rsquos which is syllabi ed as (jy) s (we) s (ls) s Syllabi cation acrossword boundaries poses a difculty for models that assign segments tosyllable slots on a word-by-word basis (for a discussion see Levelt 1992)

To conclude cross-morpheme and cross-word syllabi cation point to theneed to deal with exibility of syllable membership The WEAVER modeladdresses this issue in a particular way

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

668 ROELOFS

WEAVERrsquoS RULE-BASED APPROACH

Like Dellrsquos model WEAVER assumes that the mental lexicon is a networkof nodes and links that is accessed by spreading activation (Meyer Roelofsamp Schiller in prep Roelofs 1992ab 1993 1994 in press ab submitted abRoelofs amp Meyer in press Roelofs Meyer amp Levelt 1996) The networkconsists of three strata a conceptual stratum with lexical-concept nodes andlinks (eg GEM(X) STO NE(X) IS-A) a syntactic stratum with lemma nodes(eg juweel) nodes and links for syntactic properties (eg LEXICA L

CATEGORY noun GENDER neuter) and slots and llers for diacritics (egHAS-NUM BER sg pl) and relevant for the present paper a word-formstratum with metrical structure morpheme segment and syllable programnodes and links The word-form stratum is connected to a syllabary which isa store of ready-made motor programs for syllables (Levelt 1989 1992Levelt amp Wheeldon 1994)

Figure 3 illustrates the memory representation of the form of juweel inWEAVER The form network consists of three layers of nodes morphemenodes segment nodes and syllable program nodes Morpheme nodes standfor roots and afxes Morpheme nodes are connected to the lemma and itsparameters (diacritic feature nodes) The stem juweel is connected tothe lemma of juweel and singular A morpheme node points to the segmentsthat make up its underlying form and for some words to its metricalstructure For the storage of metrical structures a principle of economyholds WEAVER assumes that the main accent of Dutch words is on the rstsyllable containing a full vowel (which holds for about 90 of the wordtokens) unless the lexical form representation indicates otherwise (Meyeret al in prep) Thus for polysyllabic words that do not have main stress onthe rst stressable syllable the metrical structure is stored as part of thelexical entry but for monosyllabic words and for all other polysyllabicwords it is not stored but computed For example the metrical structure forjuweel [jyrsquowel] is stored but for tafel [rsquotafl] (table) it is not Metricalstructures describe abstract groupings of syllables ( s ) into feet ( S ) and feetinto phonological words ( w ) Importantly it is not speci ed which segmentsmake up the syllables nor is the CV pattern specied The links betweenmorpheme and segment nodes indicate the serial position of the segmentswithin the morpheme Possiblemdashas opposed to actualmdashsyllable positions(onset nucleus coda) of the segments are specied by the links betweensegment nodes and syllable program nodes For example the links in thenetwork specify that l is the coda of [wel] and the onset of [lIn] These linksare used in retrieving a motor program for a syllable after the actual syllablepositions of the segments have been determined by the syllabi cationprocess

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 669

FIG 3 Memory representation of the word form of juweel (gem) in WEAVER (Roelofs1994 in press a)

Information is retrieved from the network through spreading ofactivation Encoding starts when a morpheme node receives activation froma lemma Activation then spreads through the network in a forward fashionEach node sends a proportion of its activation to its direct neighbours Thereis also spontaneous decay of activation No further constraints are imposedon the activation process

The form encoders follow simple selection rules The rules areimplemented in a parallel distributed manner Attached to each node in thenetwork there is a procedure that veries the label on the link between thenode and a target node one level up Procedures are stored with the relevantdata structures Thus a representation scheme is proposed that is sometimesreferred to as ldquoobject-orientedrdquo (cf Bobrow amp Winograd 1977) Averication procedure is triggered when the nodersquos activation level exceeds athreshold The procedures may run in parallel Thus word forms are notplanned by a central agent that overlooks the whole process but by anumber of procedures that work in parallel on small parts of the word form(like several spiders making a single web) Similar production systems havebeen proposed by Anderson (1983) and Kempen and Hoenkamp (1987)

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

670 ROELOFS

The morphological encoder selects the morpheme nodes that are linked toa selected lemma and its parameters Thus juweel is selected for juweeland singular

The phonological encoder selects the segments and if available themetrical structures that are linked to the selected morpheme nodes Nextthe segments are input to a prosodication process that associates thesegments to the syllable nodes within the metrical structure (for ldquoexceptionrdquowords) or constructs metrical structures based on segmental information(see Fig 4) Thus when stored metrical structures are retrieved and woveninto the phonetic plan otherwise they are constructed on the spot Likeweaving a fabric the process has a certain direction The prosodicationproceeds from the segment whose link is labelled rst to the one labelledsecond and so forth In the prosodication process syllable positions (onsetnucleus coda) are assigned to the segments following the syllabi cationrules of the language Essentially each vowel and diphthong is assigned to adifferent syllable node and consonants are treated as onsets unlessphonotactically illegal onset clusters arise Syllabi cation rules refer tofeatures so the syllabi cation process may have to consult a segmentrsquosfeatural composition (part of the syllabary)mdashthis is one of the functions ofthe direct connection in Fig 3 from segment node w to the phonetic levelthe other function is that it provides for the creation of novel syllableprograms directly from segments In the encoding of juweel the j ismade syllable onset and the y nucleus of the rst syllable the w onset thee nucleus and the l coda of the second syllable When metrical structure iscomputed word accent is assigned to the rst stressable syllable

The prosodication process provides for cross-morpheme and cross-wordsyllabi cation In planning polymorphemic words or connected speechadjacent morphemes or words may be prosodied together This leads tonew phonological words (Booij 1995 Levelt 1989 1992) For exampleWEAVER may syllabify juweel with en (for the plural juwelen) or juweel with in (for the cliticisation juwelin) Following the maximalonset principle in syllabi cation (eg Goldsmith 1990) l will be madeonset of the third syllable instead of coda of the second In this wayWEAVER achieves syllabi cation across morpheme boundaries and acrossword boundaries

WEAVERrsquos planning of polymorphemic words and cliticisations differsfrom that proposed by Levelt (1992) in an important respect According toLevelt (1992) cliticisations like juwelin are produced by rst combining theretrieved metrical frames for juweel and in followed by segment-to-frameassociation By contrast WEAVER does not rst combine metrical framesbut starts syllabifying juweel and includes in only later in the prosodicationprocess The same holds for the stem juweel and the plural sufx -en in theproduction of the plural form juwelen

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 671

FIG 4 Phonological encoding in WEAVER Segmental spellout and for some wordsmetrical spellout followed by a prosodication process that assigns syllable positions to thesegments

The phonetic encoder selects the syllable program nodes whose labelledlinks to the segments correspond with the syllable positions assigned to thesegments For example [wel] is selected for the second phonological syllableof juweel because the link between [wel] and w is labelled onset between[wel] and e nucleus and between [wel] and l coda Note that vericationprevents a potential selection problem between [wel] and [we] (as there is inDellrsquos model as we saw earlier) In planning singular juweel the node [we]will not be selected because there is no coda link between [we] and lVerication also guarantees that the phonetic encoder selects [we] and [l]for the plural form juwelen and [we] and [lIn] for the cliticised form juwelinProvided that the selection conditions of a syllable program node are metthe actual selection of the node at any moment in time is a random eventThe probability of selecting a node at a particular moment in time is equal tothe ratio of its level of activation and the sum of the activation levels of allsyllable program nodes in the network Roelofs (1992ab 1993 1996a inpress a) gives the equations that formalise WEAVER There areexpressions for the activation dynamics the hazard rate (Luce 1986) of theencoding of a syllable and the mathematical expectation of the encodingtime as a function of these hazard rates

As a nal step the phonetic encoder addresses the syllable programs inthe syllabary uses the metrical representation to set the parameters forloudness pitch and duration and makes the programs available to thearticulators for the control of the articulatory movements (Levelt 1992

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

672 ROELOFS

Levelt amp Wheeldon 1994) The syllabary does not have to contain manyprograms Statistical analyses by Schiller Meyer Baayen and Levelt (1996)revealed that 500 different syllables cover almost 85 of the Dutch syllabletokens Finally the hierarchical phonetic plan will govern articulation

In sum word-form encoding is achieved by a spreading-ac tivation-basednetwork with labelled links combined with a parallel object-orientedproduction system WEAVER also provides for a suspensionresumptionmechanism that supports incremental generation of phonetic plansIncremental production means that encoding processes can be triggered by afragment of their characteristic input (Levelt 1989) For examplesyllabi cation can start on the initial segments of a word without having allits segments Only initial segments and for some words the metricalstructure are needed When given partial information computations arecompleted as far as possible after which they are put on hold When givenfurther information the encoding processes continue from where theystopped

THE IMPLICIT-PRIMING PARADIGM

Although WEAVER has been designed to handle among other issuesproblems such as the exibility of syllable membership the scope of themodel is much wider For example I have shown elsewhere (Roelofs 1994in press a submitted a) by computer simulation that WEAVER accounts forkey empirical ndings about the time course of phonological facilitation andinhibition from spoken distractor words in picture naming (Meyer ampSchriefers 1991) for effects from the order of encoding inside and betweenthe syllables of a word (Meyer 1990 1991) for frequency effects(Jescheniak amp Levelt 1994 Levelt amp Wheeldon 1994 Roelofs 1996b) andfor speech error effects (Nooteboom 1969) Furthermore new predictionsof the model have been validated in recent extensive series of experiments(eg Ferrand Segui amp Grainger 1996 Hendriks amp McQueen 1996 Meyeret al in prep Pederson amp Roelofs 1995 Roelofs 1996ab in press absubmitted ab Roelofs Baayen amp Van den Brink submitted Roelofs ampMeyer submitted)

A number of key ndings about phonological encoding have beenobtained with the so-called implicit-priming paradigm developed by Meyer(1990 1991) This paradigm involves producing words from learnedpaired-associates In Meyerrsquos experiments participants rst learned smallsets of promptndashresponse word pairs such as steenndashjuweel wetndashjurist etc(stonendashgem lawndashjurist etc) After learning a set on each trial one of theprompts (the rst word of a pair) was at random visually presented on acomputer screen The task for a participant was to produce the second wordof a pair (eg juweel) upon the visual presentation of the rst word (steen)

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 673

The instruction was to respond as quickly as possible without makingmistakes The production latency (ie the interval between prompt onsetand speech onset) was the main dependent variable An experimentcomprised homogeneous and heterogeneou s response sets In ahomogeneous set the response words shared part of their form and in aheterogeneou s set they did not For example the responses shared the rstsyllable (JUweel JUrist etc) or the second syllable (juWEEL houWEEL(pick-axe) etc) or they were unrelated ( jurist houweel etc)Heterogeneous sets in the experiments were created by regrouping the pairsfrom the homogeneous sets Therefore each word pair was tested bothunder the homogeneous and the heterogeneou s condition and alluncontrolled item effects were kept constant across these conditions Meyerfound a facilitatory effect from homogeneity only when the overlap wasfrom the beginning of the response words onwards Thus a facilitatory effectwas obtained for the set that included JUweel and JUrist but not for the setthat included juWEEL and houWEEL Furthermore facilitation increasedwith the number of shared segments

According to WEAVER this seriality phenomenon reects thesuspensionndashresumption mechanism that underlies the incrementalprosodication of an utterance Assume the response set consists of juweeljurist and so on (ie the rst syllable is shared) Before the beginning of atrial the morphological encoder can do nothing the phonological encodercan construct the rst phonological syllable (jy) s and the phonetic encodercan recover the rst phonetic syllable [jy] When the prompt steen is giventhe morphological encoder will retrieve juweel Segmental spelloutmakes available the segments of this morpheme which includes thesegments of the second syllable The phonological and phonetic encoderscan start working on the second syllable In the heterogeneou s condition( jurist houweel etc) nothing can be prepared There will be nomorphological encoding no phonological encoding and no phoneticencoding In the end-homogeneous condition ( juweel houweel etc)nothing can be done either Although the segments of the second syllable areknown the phonological word cannot be computed because the remainingsegments are to the left of the suspension point In WEAVER this meansthat the syllabi cation process has to go to the initial segments of the wordwhich amounts to restarting the whole process (like unravelling a wovenfabric) Thus a facilitatory effect will be obtained for the homogeneouscondition relative to the heterogeneou s condition for the begin conditiononly

Computer simulations of the experiments of Meyer (1990) can be found inRoelofs (1994 in press a) Advance knowledge about a syllable wassimulated by completing the phonological and phonetic encoding of thesyllable before the beginning of a trial For the begin condition the model

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

674 ROELOFS

yielded a facilitatory effect of 2 43 msec (real 2 49 msec collapsed acrosstrochaic feet and iambs) whereas for the end condition it predicted an effectof 0 msec (real 1 5 msec) Thus WEAVER captures the empiricalphenomenon

A similar suspensionndashresumption mechanism might work for Dellrsquos(1988) model The speech production system might work its way through thesequence of segment categories corresponding to the n shared segments (seeFig 2) After the nth category node has been reached and the correspondingsegment has been selected the encoding process is suspended The process isresumed when segment n 1 1 has been made available by the morphemethat is derived from the prompt

EVIDENCE FOR WEAVERrsquoS SYLLABIFICATION

I now review the outcomes of two series of implicit-priming experiments thatspecically support WEAVERrsquos view on syllabi cation For details of theexperiments I refer to Meyer et al (in prep) Roelofs (1996ab) and Roelofsand Meyer (in press in prep)

Morphological Structure

WEAVERrsquos syllabi cation algorithm requires morphologicallydecomposed form entries The reason for this is that in languages such asDutch morphemes such as prexes particles the base verbs of particleverbs and prexed verbs and some sufxes constitute independent domainsof syllabi cation (Booij 1983 1995) In Dutch this holds for prexes such asver- and ont- particles such as op af aan uit and so forth and sufxes suchas -achtig but not for sufxes such as -in -er and -ing For example thesegment r of the prex ver- of the Dutch verb vereren (honour) is notsyllabi ed with the base verb eren as the maximal onset principle insyllabi cation would predict but is syllabi ed as the coda of ver- This doesnot hold for a pseudo-pre xed verb such as verieumlren (verify) where the r isthe onset of the second syllable ri instead of the coda of ver- Similar to theprexed verb vereren the t of the particle verb uitademen (breathe out) issyllabi ed with the particle uit (out) and not with the base verb ademen(breathe) Also the nal segment r of the base kinder (child) inkinderachtig (childish) is syllabi ed with the base and not with the sufx-achtig (contrary to the English translation equivalent) By contrast the nalsegment w of leeuw (lion) in leeuwin (lioness) is syllabi ed with the sufx-in (-ess)

The component morphemes of Dutch compounds also constituteindependent domains of syllabi cation For example the Dutch nominalcompound handappel (eating-apple ) consists of the morphemes hand

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 675

and appel The segment d of hand in handappel is not syllabi ed withappel as the principle of maximization of onset would predict but issyllabi ed as the coda of hand Thus handappel is syllabi ed as (hAnd) s

(Ap) s (pl) s This does not hold for synchronically pseudo-complex wordssuch as aardappel (ldquoearth applerdquo potato) which is syllabi ed as (ar) s (dAp) s

(pl) s In WEAVER aardappel is represented by one node aardappel atthe morphological level whereas handappel is represented by two nodesone for hand and one for appel Thus in morphological encodingthe lemma node of the word aardappel is mapped onto one morpheme nodejust like the lemma node of a simple word like juweel (gem) is whereas thelemma node of the compound handappel is mapped onto two morphemenodes In the next processing step phonological procedures select thesegments linked to these morpheme nodes and syllabify the segments takingeach morpheme as a domain of syllabi cation

Recent implicit-priming experiments in Dutch have tested for effects ofmorpheme preparation predicted by WEAVER (Roelofs 1996ab RoelofsBaayen amp Van den Brink submitted) For monomorphemic words such asbijbel (bible) consisting of the morpheme bijbel sharing the rst syllablebij allows phonological preparation only In contrast for polymorphemicwords such as bijrol (supporting role) consisting of the morphemes bij and rol additional morphological preparation is possible In ahomogeneous condition where the responses share the syllable bij (b e i) s

and [b e i] will have been planned for the monomorphemic word bijbel beforethe beginning of a trial The morpheme bijbel and the second syllable belwill be planned during the trial itself In a heterogeneou s condition wherethe responses do not share part of their form the whole monomorphemicword has to be planned during the trial If the rst morpheme bij and(b e i) s and [b e i] have been planned for the polymorphemic word bijrol beforethe beginning of a trial in the homogeneous condition the secondmorpheme rol can be selected during the trial itself and the secondsyllable rol can be computed In the heterogeneou s condition however theinitial morpheme bij has to be selected rst before segments can beselected for the second morpheme rol so that the second syllable rol canbe computed Thus in the case of a polymorphemic word such as bijroladditional morphological preparation is possible before the beginning of atrial Consequently extra facilitation should be obtained Thus thefacilitatory effect for bij in bijrol (consisting of the morphemes bij and rol ) should be larger than the effect for bij in bijbel ( bijbel )

The outcomes conrmed the predictions of WEAVER In producingdisyllabic simple and compound nouns a larger facilitatory effect wasobtained when a shared initial syllable constituted a morpheme than when itdid not For example the effect was larger for bij in bijrol ( bij and rol ) than for bij in bijbel ( bijbel )

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

676 ROELOFS

In the simulation of the experiment the homogeneous response setsconsisted of bijrol ( bij rol supporting role) bijnier ( bij nier kidney) and so forth (the real condition) versus bijbel ( bijbel bible)bijster ( bijster loss) and so forth (the pseudo condition) Theheterogeneou s sets were created by recombining the responses of differenthomogeneous sets The critical items were embedded in a network of 50randomly selected words (no embedding produced the same simulationoutcomes) Advance knowledge about the form of the response word wassimulated by completing the morphological phonological and phoneticencoding of the word form as far as possible before the beginning of a trialThe parameters for WEAVER that t these data were identical with ts ofWEAVER to other data

Figure 5 presents the results of the simulations Sharing bij in bijrol andbijbel yields a facilitatory effect The facilitatory effect for bij in bijrol(consisting of the morphemes bij and rol the real condition) is largerthan the facilitatory effect for bij in bijbel ( bijbel the pseudo condition)This corresponds to the empirical ndings To conclude WEAVERaccounts for the empirical phenomenon

In addition WEAVER predicts an effect of morpheme frequency for theconstituents of polymorphemic lexical items Frequency effects in the modeloriginate from differences in the speed of running procedures Speeddepends on frequency of usage (more experienced spiders work faster inmaking a web) This has been tested by Roelofs (1996b in press b) using theimplicit-priming paradigm High-frequency morphemes are retrieved fasterfrom memory than morphemes of low frequency so the bene t frompreparation should be larger for low-frequency morphemes than forhigh-frequency ones This prediction was empirically conrmed Forexample in producing compounds (Roelofs 1996b) the facilitatory effectwas larger for response sets sharing a low-frequency morpheme like bloem (ower) as in bloemkool (cauliower) than for response setssharing a high-frequency morpheme like bloed (blood) as in bloedspoor(trace of blood) Also in producing particle verbs (Roelofs in press b) thefacilitatory effect was larger for a morpheme like veeg (low frequency) inldquoveeg oprdquo (ldquoclean uprdquo) than for a morpheme like geef (high frequency) inldquogeef oprdquo (ldquogive uprdquo)

The results concerning the pseudo and real morphemes and the effect ofmorpheme frequency show that the implicit-priming paradigm is sensitive tomorphological structure In a compound such as handappel morphemestructure is needed to arrive at the correct syllabi cation For compoundswhose rst morpheme ends in a vowel like bijrol however morphemestructure is not needed for reasons of syllabi cation (the correctsyllabi cation would be produced even without access to the wordrsquosmorpheme structure) but for other formal behaviour such as gapping in a

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 677

FIG 5 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for pseudo and real morphemes Empiricaldata (Roelofs 1996a) and data simulated by WEAVER n Empirical N WEAVER

conjunction (eg Booij 1995) The non-initial morpheme of a compound canbe omitted under identity with the non-initial morpheme of an adjacentcompound as in ldquobij- en hoofdrolrdquo (ldquosupporting and leading rolerdquo) Thisonly holds when the morphemes involved are independent phonologicalwords Prexes dene domains of syllabi cation but they are notphonological words (Booij 1995) To test the idea that morpheme structureis stored when it is required for syllabi cation Roelofs Baayen and Van denBrink (submitted) compared the preparation effect for prexed verbs thatneed morpheme structure for correct syllabi cation (eg vereren (honour))with the preparation effect for prexed verbs that do not need morphemestructure for correct syllabi cation (eg verkopen (sell) where rk is anillegal onset cluster in Dutch) In homogeneous sets the responses shared theprex syllable (eg ver-) whereas in heterogeneou s sets there was nooverlap As predicted the effect of morpheme preparation was onlyobtained when morpheme structure was needed for syllabi cation Thus themorphemic effect was only observed for words like vereren but not for wordslike verkopen

To conclude WEAVERrsquos syllabi cation algorithm requiresmorphological ly decomposed form entries because morphemes denedomains of syllabi cation Larger preparation effects for real than forpseudo morphemes and effects of morpheme frequency show that implicitpriming is sensitive to morphological structure In producing prexed verbs

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

678 ROELOFS

effects of morpheme preparation were only obtained for words that requiremorpheme structure for a correct syllabi cation

Metrical Structure

The experiments on morphological structure concern the input tophonological encoding Other experiments tested WEAVERrsquos claim thatphonological encoding involves a parallel spellout of metrical structure andsegments followed by segment-to-frame association (see Fig 4)WEAVER assigns a specic role to metrical structures in syllabi cation Forwords like the trochee tafel metrical structures are computed on-line but forwords like the iamb juweel metrical structures are stored The storedmetrical structures specify the number of syllables and the stress pattern butnot the precise CV structure of the syllables as the CV headers of Dell (1988)do The prosodication process in WEAVER associates segments with thesyllable nodes within the metrical structure for ldquoexceptionrdquo words (egjuweel) or constructs syllable and metrical structures based on segmentalinformation (eg for tafel) Roelofs and Meyer (in press in prep) and Meyeret al (in prep) have conducted a series of implicit-priming experimentsdesigned to test WEAVERrsquos view on metrical structure in syllabi cation

On each trial participants had to produce one Dutch word out of a set ofthree or four as quickly as possible In homogeneous sets the responsesshared a number of word-initial segments whereas in heterogeneou s setsthey did not As we have seen earlier research has shown that sharing initialsegments reduces production latencies (Meyer 1990 1991 Roelofs1996ab) The responses shared their metrical structure (the constant sets) orthey did not (the variable sets)

A rst series of experiments (Roelofs amp Meyer in press) testedpredictions of WEAVER about the role of metrical structure in theproduction of polysyllabic words that do not have main stress on the rststressable syllable such as juweel According to the model the metricalstructures of these words are stored in memory WEAVERrsquos view ofsyllabi cation implies that preparation for word-initial segments should onlybe possible if such response words have an identical metrical structure If theresponses in a set have different metrical structures segment-to-frameassociation cannot take place before the beginning of a trial This predictionwas tested by comparing the effect of segmental overlap for response setswith a constant number of syllables such as manier (manner) matras(mattress) makreel (mackerel) (all 2 syllables) with that for sets having avariable number of syllables such as majoor (major) materie (matter)malaria (malaria) (respectively 2 3 and 4 syllables) In the example theresponses share the rst syllable ma Word stress was always on the secondsyllable As predicted facilitation was obtained for the metrically constant

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 679

sets but not for the variable sets The same predictions were also testedby comparing the effect of segmental overlap for response sets with aconstant stress pattern such as marine (navy) materie (matter) malaise(depression) madonna (madonna) (all responses having stress on thesecond syllable) with that for sets having a variable stress pattern such asmarine (navy) materie (matter) manuscript (manuscript) madelief(daisy) (rst two responses having stress on the second syllable and the lasttwo responses having stress on the third syllable) All response words weretrisyllabic Again as predicted facilitation was observed for the constantsets but not for the variable sets In sum the experiments have yieldedcounterintuitive results that support WEAVERrsquos approach to metricalstructure

The results of these experiments suggest that constancy in CV structure isnot necessary to observe a preparation effect because in none of thehomogeneous sets of the experiments was this structure identical acrossresponse words (though one could perhaps argue that the response words inthe metrically constant sets were more similar in CV structure than those inmetrically variable sets) Results obtained by Meyer (1990 1991) andRoelofs (1996ab in press b) also suggest that implicit priming effects can beobtained for homogeneous sets with variable CV structures Howeveralthough constancy in CV structure does not appear to be necessary forobtaining a facilitatory effect it is still possible that stronger effects arise forsets with constant than with variable CV structure This is not predicted byWEAVER but as explained earlier others (eg Dell 1988) have argued foran explicit representation of CV structure Therefore we explicitly testedwhether the size of the preparation effect was affected by the constancyversus variability of the CV structure of the response words We comparedthe effect of segmental overlap for response sets having a constant CVstructure such as bres (breach) bril (glasses) brok (piece) brug (bridge)(all CCVC responses) with that for sets having a variable CV structure suchas brij (porridge) brief (letter) bron (source) brand (re) (CCVVCCVVC CCVC CCVCC responses respectively) In the example theresponses share the onset cluster br Facilitation from segmental overlap wasobtained for both the constant and the variable sets The size of the effectwas the same for both types of set

These results suggest that the exact CV structure is not stored therebyrefuting the CV headers of Dell With constant CV structure the systemmight select a header (ie CCVC) and work its way through the sequence ofsegment categories corresponding to the shared segments After the lastshared category node has been reached and the corresponding segment hasbeen selected the encoding process is suspended The process is resumedwhen the rst non-shared segment has been made available by themorpheme that is derived from the prompt With variable CV structure

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

680 ROELOFS

FIG 6 Mean difference in milliseconds between the production latencies in thehomogeneous and the heterogeneous conditions for constant and variable metrics Empiricaldata (Roelofs amp Meyer in press) and data simulated by WEAVER n Empirical N WEAVER

however the system cannot select a header and work its way through thesequence of segment categories corresponding to the shared segments Forexample suspension in a CCVV sequence leads to wrong results if it turnsout that a CCVVC word has actually to be produced Neverthelessempirically the same amount of facilitation from segmental overlap wasobtained for the constant and the variable sets

WEAVER explains why preparation for word-initial segments is onlypossible for response words with an identical number of syllables and stresspattern and why identical CV is not needed Figure 6 shows the results ofsimulations comparing the effect of segmental overlap for response sets witha constant number of syllables such as manier matras makreel with thatfor sets having a variable number of syllables such as majoor materiemalaria Varying the place of stress while keeping the number of syllablesxed gives the same results The parameters for WEAVER that t thesedata were identical with ts of WEAVER to other data As can be seenWEAVER accounts for the key empirical nding concerning metricalstructure In contrast if metrical structures are not involved in advanceplanning or if metrical structures are computed on-line on the basis ofsegments for these words sharing metrical structure should be irrelevant forpreparation Then preparation (ie on-line computing the syllable mabefore the beginning of a trial and computing the remainder of the word

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 681

form during the trial itself) should be possible for both the metricallyconstant and variable sets

Meyer et al (in prep) tested predictions of WEAVER about the role ofmetrical structure in producing monosyllabic words and polysyllabic wordswhose main stress is on the rst stressable syllable like tafel According to themodel the metrical structures of these words are computed on-line by theprosodication process That is syllabi cation and stress assignment is doneon the basis of retrieved segments Consequently implicit priming of initialsegments should now be possible for both metrically constant and variablesets This prediction was tested by comparing the effect of segmental overlapfor response sets with a constant number of syllables such as borstel (brush)botsing (crash) bochel (hump) bonje (ght) (all disyllables stressed on therst syllable) with that for sets having a variable number of syllables such asborstel botsing bok (goat) bom (bomb) (two disyllables stressed on therst syllable and two monosyllables respectively) In the example theresponses share the onset and nucleus (bo) As predicted an equal amountof facilitation was obtained for the constant and the variable sets The sameresult is predicted for varying the number of syllables of polysyllabic wordswith an unstressable rst syllable (ie words with a schwa as the rst vowel)and stress on the second syllable This prediction was tested by comparingthe effect of segmental overlap for response sets with a constant number ofsyllables such as gebit (teeth) gezin (family) getal (number) gewei(antlers) (all disyllables having stress on the second syllable) with that forsets having a variable number of syllables such as geraamte (skeleton)getuige (witness) gebit gezin (two trisyllables stressed on the secondsyllable and two disyllables stressed on the second syllable respectively) Aspredicted an equal amount of facilitation was obtained for the constant andthe variable sets

Finally Roelofs and Meyer (in prep) tested predictions of WEAVERabout the production of cliticisations and sufxed forms According toLevelt (1992) cliticisations like juweel rsquos (eg in ldquogeef het juweel rsquosrdquo ldquogivethe gem nowrdquo) are produced by rst combining the retrieved metrical framesfor juweel and rsquos followed by segment-to-frame association By contrastWEAVER does not rst combine metrical frames but starts syllabifyingjuweel and includes rsquos only later in the prosodication process The sameholds for the stem juweel and the plural sufx -en in the production of theplural form juwelen These different views were tested by comparing theeffect of segmental overlap for response sets that combine disyllabic nouns(eg dozijn [dorsquoz e in] (dozen)) with disyllabic verb stems (eg doneer[dorsquoner] (donate)) with that of sets combining the disyllabic nouns withtrisyllabic cliticised forms of the verbs (doneer rsquos) or with trisyllabicinnitival forms of the verbs (doneren) If metrical frames are combinedbefore segment-to-frame association then the sets with the cliticisations

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

682 ROELOFS

(eg the set including disyllabic dozijn and trisyllabic doneer rsquos) and the setswith the innitives (eg the set including disyllabic dozijn and trisyllabicdoneren) would be metrically variable and preparation should not bepossible According to WEAVER however all three types of set aremetrically constant since the clitic and the plural sufx are metricallyindependent elements that are adjoined after the prosodication process hasstarted syllabifying the verb stem The outcomes of the experimentsupported the predictions of WEAVER The sets with the cliticisations (egthe set including dozijn and doneer rsquos) and the sets with the innitives (egthe set including dozijn and doneren) yielded an equal amount of segmentalfacilitation and the size of the facilitation was the same as that for thecontrol sets including the disyllabic verb stems (eg the set including dozijnand doneer)

EXPERIMENT

I now report a new experiment testing WEAVER The experiment had twomain objectives First it examined whether the facilitatory effect of responsehomogeneity that Meyer (1990 1991) Roelofs (1996a) and Roelofs andMeyer (in press) obtained for single words can be generalised to theproduction of sentential forms Bock (1987) found that phonologicalsimilarity had overall inhibitory inuences in a sentence production taskWEAVER predicts facilitation instead of inhibition for implicit priming ofsentence production Secondly the experiment contrasted Dellrsquos model andWEAVER In particular the experiment tested the new suggestion forcross-morpheme syllabi cation discussed in the Theoretical Positionssection There it is suggested that cross-morpheme syllabi cation in Dellrsquosmodel might be achieved by connecting a critical syllable to bothmorphemes involved and the syllablersquos segments to category nodes I arguedthat this suggestion breaks down on syllabi cation across word boundariesStill it might be a solution for cross-morpheme syllabi cation Thesentences in the experiment were chosen such that the computation of theirsound structure involved encoding across morpheme boundaries Forexample a promptresponse pair in the experiment was ets ldquolenen we uitrdquo(literally bicycle ldquolend we outrdquo) The morphological and phonologicalstructures of the utterance relate to each other as

Morphology [ [leen]STEM [en]AFF ]V [we]PRO [uit]P

Phonology ( (le) s (n) s (w) s ) w (( L yt) s ) w

The responses in the experiment are particle verbs such as uitlenen (lendout) which are combined with a personal pronoun into an elliptical sentence

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 683

which leaves out the grammatical object The object corresponds to theprompt in the example ets (bicycle) The pronoun is the reduced form we[w] of the rst-person plural nominative form wij [w e i] so the particle verbhas to be produced in the plural form The verb is in the indicative moodMost importantly the nal segment of the stem morpheme is syllabi edwith the plural sufx The nal segment n of the morpheme leen issyllabi ed with the plural sufx -en yielding the form (le) s (nn) s The pluralsufx in Dutch is not an independent domain of syllabi cation The stemplus sufx together make up a single phonological word Furthermore thereduced form of wij [w] cannot stand alone so it is adjoined to lenen Theresulting form lenen we is syllabi ed as (le) s (nn) s (w) s Dutch has apost-lexical rule that deletes the nal n of a syllable after a schwa at the endof a morpheme that is not a verbal stem (Booij 1995) Since this rule appliesto the nal segment of lenen it is deleted yielding the form (le) s (n) s (w) s Finally lenen we plus the particle uit make up the response lenen weuit

WEAVER predicts a facilitatory effect from response-set homogeneityfor the production of these sentential forms similar to that obtained byMeyer (1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) forsingle words In the homogeneous condition the encoding process can besuspended after having encoded the rst two segments before the beginningof a trial whereas in the heterogeneou s condition such preparation is notpossible In Dellrsquos model however advance planning of these sentences inthe homogeneous condition is problematic if not impossible

The problem posed by these sentences to Dellrsquos model is as followsAssume that in the homogeneous condition the rst surface syllable le isprepared before the beginning of a trial and that the remainder of theutterance (nen we uit) is encoded during the trial itself After havingprepared the rst syllable le supplying a jolt of activation to leen (duringthe trial itself) to retrieve the onset n so as to make it the onset of the nextsyllable nen the onset l will also be activated Supplying a jolt of activationto en alone is not sufcient because en is connected to all itsldquoallomorphsrdquo nen len and so forthmdashat least that is the solution we aretesting here Furthermore the category node COn connects to onset nonset l and so forth Note further that leen is a monosyllabicmorpheme so there will be no changing of spreading rates as would be thecase in the encoding of polysyllabic morphemes

In the heterogeneou s condition the model may account for performanceby assuming that the strength of the connection between leen and le (andthus onset l) is stronger than that between leen and nen (and thus onsetn) Thus after having supplied a jolt of activation to leen rst the onsetl node will be selected After selection its activation will be set to zeroConsequently onset n may be the most highly activated onset node in the

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

684 ROELOFS

encoding of the next syllable nen However when preparing the rst twosegments onset l and nucleus e before the beginning of a trial in thehomogeneous condition supplying (during the trial itself) a jolt of activationto leen to retrieve onset n will also activate onset l In fact given thatthe strength of the connection between leen and le (ie onset l) wasassumed to be stronger than that between leen and nen (ie onset n)the onset l node will be more active than the onset n and onset l willerroneously be selected Adding a jolt of activation to leen and en simultaneously instead of to leen only will not help Recall that en isconnected to nen len and so forth

In short preparation of the sound structure of the sentences in theexperiment causes an encoding problem in Dellrsquos model The simplestsolution to this problem is to refrain from preparing the response in thehomogeneous condition This would predict that for the sentences in theexperiment no facilitatory effect from homogeneity will be obtainedcontrary to the prediction of WEAVER Consequently if a facilitatoryeffect is obtained this would pose a challenge to the model of Dell (19861988)

Method

Participants The experiment was conducted with a group of 12 paidparticipants from the pool of the Max Planck Institute All participants werenative speakers of Dutch

Materials and Design The materials were obtained from the Dutch partof the CELEX lexical database (Baayen Piepenbrock amp van Rijn 1993)The stimulus materials consisted of two practice sets and six experimentalsets of three promptndashresponse pairs each listed in Table 1 There were ninedifferent responses which consisted of elliptical sentences having the samestructure as lenen we uit Each set of pairs was tested in a separate block oftrials In three experimental sets (the homogeneous sets) the response wordsshared the onset and nucleus of the rst syllable and in the remaining threesets (the heterogeneou s sets) they were unrelated in form Thus in thehomogeneous condition each response was tested together with otherresponses with the same onset and nucleus of the rst syllable whereas in theheterogeneou s condition the responses tested together in a block did notshare part of the rst syllable Following Meyer (1990) the rst independentvariablemdash homogeneous vs heterogeneou s setsmdashwill be called ldquocontextrdquoThe same promptndashresponse pairs were tested in the homogeneous andheterogeneou s conditions only their combinations into sets differed Theshared onset and nucleus of the syllables in the homogeneous sets were [z e ][le] and [vu] Each of these syllables was used in one homogeneous set The

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 685

TABLE 1Stimulus Materials

Pairsgewoontendashleren we af arrestantndashvoeren we af reisndashzeggen we afetsndashlenen we uit snelwegndashvoegen we uit lmndashzenden we uittekstndashlezen we op kindndashvoeden we op kraagndashzetten we op

ContextHomogeneous Heterogeneous

Set 1 Set 1VOEren we af zeggen we afVOEgen we uit lenen we uitVOEden we op voeden we op

Set 2 Set 2LEren we af voeren we afLEnen we uit zenden we uitLEzen we op lezen we op

Set 3 Set 3ZEggen we af leren we afZEnden we uit voegen we uitZEtten we op zetten we op

Approximate English translation of the verbsgewoontendashleren we af [rsquoler w rsquoAf] habitndashunlearnetsndashlenen we uit [rsquolenw rsquo L yt] bicyclendash lend outtekstndashlezen we op [rsquolezw rsquoOp] textndashread aloudarrestantndashvoeren we af [rsquovur w rsquoAf] prisonerndashcarry awaysnelwegndashvoegen we uit [rsquovux w rsquo L yt] highwayndashtake a turnkindndashvoeden we op [rsquovudw rsquoOp] childndashbring upreisndashzeggen we af [rsquoz e x w rsquoAf] tripndashcancel lmndashzenden we uit [rsquoz e ndw rsquo L yt] lmndashbroadcastkraagndashzetten we op [rsquoz e tw rsquoOp] collarndashturn up

second independent variable which had three levels ([z e ] [le] and [vu]) willbe called ldquofragmentrdquo

Each response was coupled with a prompt that I considered a strong andunambiguous retrieval cue for the corresponding target All prompts werenouns and all responses were elliptical sentences The prompt named atypical themepatient for the verb (eg ets ldquolenen we uitrdquo bicycle ldquolend weoutrdquo) so that the association between prompt and response would be anatural one and be easy to remember

Each participant was tested on all sets The order of the sets was rotatedacross participants in the following way A group of six participants was rsttested on the homogeneous sets and then on the heterogeneou s sets For theremaining six participants the order of testing was reversed The sets were

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

686 ROELOFS

tested in three successive sections of the experimental session A differentorder of the six sets within a section was used for each participant of a groupThe three sections will be the third independent variable which I will referto as ldquorepetitionrdquo

Each of the three promptndashresponse pairs of a set was tested six timeswithin each block of trials The order of testing the pairs was random exceptthat immediate repetitions of pairs were excluded A different order wasused for each block of trials and each participant

Procedure and Apparatus The participants were tested individuallyThey were seated in a quiet room in front of a computer screen (NECMultisync 30) and a microphone (Sennheisser ME40) After the participanthad read the instructions two practice blocks (a homogeneous and aheterogeneou s one with the same structure as an experimental block butwith different items) were administered followed by the 18 experimentalblocks In the learning phase before each block the three pairs of a set werepresented on the screen As soon as the participant indicated having studiedthe pairs sufciently the experimenter started the test phase The structureof a trial was as follows First the participant saw a warning signal (anasterisk) for 500 msec Next the screen was cleared for 500 msec followedby the display of the prompt for 1500 msec The asterisk and prompt werepresented in white on a black background Finally before the start of thenext trial there was a blank interval of 500 msec Thus the total duration ofa trial was 3 sec The experiment was controlled by a Hermac 386 SXcomputer

Analyses After each trial the experimenter coded the response forerrors Experimental sessions were recorded on audiotape by a Sony DTC55DAT recorder The recordings contained the participantrsquos speech and tonesindicating the onset of the prompt (1 kHz) and the moment of the triggeringof the voice key (25 kHz) These tones were also heard by the experimenter(via closed headphones) for each trial The recordings were consulted afterthe experiment when the experimenter was in doubt about whether aresponse was completely correct Four types of incorrect responses weredistinguished First a participant might have produced a wrong responseSecondly the response might have exhibited a disuency that is theparticipant stuttered paused within the utterance or repaired the utteranceThirdly the voice key might have been triggered by a non-speech sound(noise in the environment or a smacking sound produced by the lips ortongue) Fourthly the participant might have failed to respond within atime-out period of 1500 msec Incorrect responses were excluded from thestatistical analysis of the production latencies

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 687

The production latencies and error rates were submitted to by-participantand by-item analyses of variance with context repetition and fragment asrepeated-m easures factors

Results and Discussion

A signicant facilitation effect of 28 msec was obtained for context [meansfor the homogeneous and heterogeneou s conditions 680 and 708 msecrespectively with by-participant standard errors of 56 and 56 msecrespectively F1(110) 5 3729 MSe 5 3372 P 0001 F2(16) 5 1153 MSe5 908 P 0015] Also main effects of repetition [F1(220) 5 403 MSe 59814 P 0035 F2(212) 5 879 MSe 5 398 P 0005] and fragment[F1(220) 5 4613 MSe 5 5254 P 0001 F2(26) 5 527 MSe 5 3830 P 005] were obtained Context did not interact with repetition [F1(220) 5108 MSe 5 5531 P 035 F2(212) 5 270 MSe 5 185 P 011] orfragment [F1(220) 5 664 MSe 5 2631 P 0006 F2(26) 5 160 MSe 5908 P 028]

In the experiment the overall error rate (wrong responses anddysuencies) for the homogeneous and heterogeneou s conditions was 11and 12 respectively The percentage of time-outs was 02 for thehomogeneous condition and 01 for the heterogeneou s condition and thepercentage of false triggering of the voice-key was 03 and 03respectively The statistical analyses of the errors did not yield signicantresults

The results support the following two conclusions First the experimentshows that the facilitatory effect from response homogeneity that Meyer(1990 1991) Roelofs (1996a) and Roelofs and Meyer (in press) obtained forsingle words can also be found for the production of sentential forms Thussuch a facilitatory effect is also obtained outside naming Secondly thefacilitatory effect from response homogeneity supports WEAVER ratherthan the model of Dell WEAVER predicted the facilitatory effect for theproduction of these sentential forms In the homogeneous condition theencoding process can be suspended after having encoded the rst twosegments before the beginning of a trial whereas in the heterogeneou scondition such preparation is not possible By contrast it is difcult forDellrsquos model to explain the advance planning of these sentences in thehomogeneous condition

One objection to the experiment may be that there is no specic evidencethat it speaks to syllabi cation The small priming effect observed couldsimply be due to the shared initial consonant only Although theexperiments on metrical structure discussed earlier suggest that the implicit-priming paradigm taps into syllabi cation the present experiment may bean exception Thus additional controls would be required in subsequent

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

688 ROELOFS

research Note however that even if the advantage for the homogeneouscondition would be due to the shared onset segment only the experimentwould still test the proposed new mechanism for cross-morphemesyllabi cation in Dellrsquos model Crucial for the test is that responsepreparation (eg le) includes selection of the shared initial consonant (ieonset l) because this would interfere with the selection of the resyllabi edonset segment for the second morpheme (ie the selection of onset n inproducing ldquolenen we uitrdquo)

Another objection to the experiment may be that implicit priming doesnot tap into morphophonological encoding processes only and thereforedoes not need to reect the relevant level of encoding in Dellrsquos modelPerhaps Dell may explain the facilitation effect at a phonetic level ratherthan at a phonological level However the problem with this view is againthat it leaves unexplained the effects from metrical structure discussedearlier If implicit priming can reect preparation of segments at a phoneticlevel only at least some facilitation should have been obtained in themetrically variable conditions (with the metrically exception words)However these conditions yielded no facilitation at all Thus the facilitationin the current experiment cannot reveal preparation of segments at aphonetic level only but should involve phonological encoding and thereforeposes a challenge to the model of Dell (1986 1988)

GENERAL DISCUSSION

Models of speech production differ in their claims about syllabi cation In amemory-based approach (eg Dell 1986 1988) the syllable structure ofeach word is stored in the mental lexicon In contrast according to therule-based approach the planning of speech involves the assignment ofsyllable positions to segments after they have been retrieved for a word frommemory (eg Beacuteland et al 1990 Levelt 1992) This paper has made a casefor the rule-based approach to syllabi cation realised in the WEAVERmodel of speech production (Roelofs 1994 in press a) I have argued thatcross-morpheme and cross-word syllabi cation pose difculties for amemory-based approach but not for WEAVER I have also reviewedempirical support for WEAVERrsquos specic form of syllabi cation Finally Ireported a new experiment on syllabi cation which supports WEAVERrather than Dellrsquos model

In this last section I will discuss an important issue in syllabi cation thatmay be investigated (further) in future research namely the issue ofresyllabi cation in speech production Is rule-based syllabi cation achievedin one or two steps

In Leveltrsquos (1992) view syllabi cation occurs only once in the speechproduction process In a single syllabi cation pass segments receive their

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 689

ultimate syllable positions in the utterance For example the newphonological word (le) s (nt) s (leen rsquot cf lend it syllabi ed as (le n) s (dIt) s )made up of the lexical words leen (lend) and het (it) is created by rstcombining the metrical structures of leen and het and next associating theirsegments with this new composed structure (Levelt amp Wheeldon 1994) Incontrast according to the theory of Lexical Phonology in phonology(Kiparsky 1982 for reviews see Goldsmith 1990 Kenstowicz 1994Mohanan 1986) a clitic would be adjoined to an already syllabi ed basefollowed by resyllabi cation For example (le) s (nt) s is created by rst fullysyllabifying leen adjoining het and resyllabifying the n WEAVER canhandle both possibilities so for the model it is more of an empirical issuethan a theoretical one In the case of syllabi cation in a single pass theprosodication of leen would already have to take the presence of het intoaccount In contrast when leen is rst fully syllabi ed a second step ofresyllabi cation would be needed Note that on both accountssyllabi cation is an active process Syllable structures of words are not storedin memory as in models such as that of Dell (1988) I now discuss someevidence for the view of a two-step construction of new phonological wordsthat combine lexical words

Syllable- nal devoicing in Dutch bears on the issue of resyllabi cation(eg Berendsen 1986 Booij 1995) For example the pronunciation of theDutch word voed (feed) is [vut] However underlyingly the nal stem-consonant of the word is voiced (ie d ) which explains the appearance ofthe d in the plural form [vudn] What happens with this segment when forexample the word voed is combined with the reduced form of the pronounhem rsquom [m] in ldquovoed rsquomrdquo (ldquofeed himrdquo) With combination of metricalstructures prior to association (Levelt 1992) the underlying d shouldsurface as d (ie voiced) but when a clitic is adjoined to an alreadysyllabi ed base and the d is resyllabi ed the underlying d should surfaceas t (ie voiceless) In the case of resyllabi cation the d would haveoccupied a syllable coda position in the derivation where it wouldobligatorily undergo the Dutch devoicing rule Linguistic intuition(Berendsen 1986 Booij 1995) acoustic measurements and perceptualjudgements by participants (Baumann 1996) suggest that Dutch speakerssay [vutm] that is the obstruents are voiceless in syllable-onset positionHowever native speakers differ in their pronunciations For somecliticisations the variants with voiced obstruents are found This mayindicate that syllabi cation occurs in a single pass for these forms (asproposed by Levelt) or may suggest that certain cliticisations are stored inmemory (eg high-frequency combinations of hosts and clitics as proposedby Booij 1995)

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

690 ROELOFS

CONCLUSION

In this paper I have summarised the support for the claim that syllabi cationis computed by rule as opposed to being directly stored in the mental lexicon(direct storage is claimed by Dell 1988) First I argued that models thatrigidly store words as sequences of syllable nodes or models that store eachconsonant as an onset or coda have a difcult time dealing with the need forexibility of syllable membership (ie in cross-morpheme and cross-wordsyllabi cation) Secondly I reviewed empirical evidence that supportsWEAVERrsquos claim that syllable structure is computed on-line and in aleft-to-right fashion The evidence suggests that syllable structure iscomputed by associating retrieved segments with the syllable nodes withinretrieved metrical structures for polysyllabic words that do not have mainstress on the rst stressable syllable (which holds for roughly 10 of theDutch word tokens) and by constructing syllable and metrical structuresbased on segmental information for monosyllabic words and for all otherpolysyllabic words (the remaining 90 of the words) Furthermore Ireviewed evidence that supports WEAVERrsquos claim that cross-morphemeand cross-word syllabi cation is achieved by adjoining a sufx or a clitic to analready partly syllabi ed base rather than by rst combining the metricalframes for the base and for the sufx or clitic followed by segment-to-frameassociation (contrary to Levelt 1989 1992) Whether speech productioninvolves resyllabi cation is an open issue

REFERENCESAnderson JR (1983) The architecture of cognition Cambridge MA Harvard University

PressBaayen RH Piepenbrock R amp van Rijn H (1993) The CELEX Lexical Database

(CD-ROM) Linguistic Data Consortium University of Pennsylvania Philadelphia PABaumann M (1996) The production of syllables in connected speech PhD dissertation

University of NijmegenBeacuteland R Caplan D amp Nespoulous J-L (1990) The role of abstract phonological

representations in word production Evidence from phonemic paraphasias Journal ofNeurolinguistics 5 125ndash164

Berendsen E (1986) The phonology of cliticizations Dordrecht ForisBobrow DG amp Winograd T (1977) An overview of KRL a knowledge representation

language Cognitive Science 1 3ndash46Bock JK (1987) An effect of the accessibility of word forms on sentence structures Journal

of Memory and Language 26 119ndash137Booij GE (1983) Principles and parameters in prosodic phonology Linguistics 21

249ndash280Booij GE (1995) The phonology of Dutch Oxford Oxford University PressButterworth B (1989) Lexical access in speech production In W Marslen-Wilson (Ed)

Lexical representation and process Cambridge MA MIT PressCaplan D (1992) Language Structure processing and disorders Cambridge MA MIT

Press

SYLLABIFICATION 691

Chomsky N amp Halle M (1968) The sound pattern of English New York Harper and RowDell GS (1986) A spreading-activation theory of retrieval in sentence production

Psychological Review 93 283ndash321Dell GS (1988) The retrieval of phonological forms in production Tests of predictions

from a connectionist model Journal of Memory and Language 27 124ndash142Ferrand L Segui J amp Grainger J (1996) Masked priming of word and picture naming

The role of syllabic units Journal of Memory and Language 35 708ndash723Garrett MF (1975) The analysis of sentence production In GH Bower (Ed) The

psychology of learning and motivation New York Academic PressGarrett MF (1980) Levels of processing in sentence production In B Butterworth (Ed)

Language production Vol 1 Speech and talk London Academic PressGarrett MF (1988) Processes in language production In FJ Newmeyer (Ed) Linguistics

The Cambridge survey Vol 3 Cambridge MA Harvard University PressGoldsmith J (1990) Autosegmental and metrical phonology Cambridge MA BlackwellHendriks H amp McQueen J (Eds) (1996) Annual Report 1995 Nijmegen Max Planck

Institute for PsycholinguisticsHoughton G (1990) The problem of serial order A neural network model of sequence

learning and recall In R Dale C Mellish amp M Zock (Ed) Current research in naturallanguage generation London Academic Press

Jescheniak JD amp Levelt WJM (1994) Word frequency effects in speech productionRetrieval of syntactic information and phonological form Journal of ExperimentalPsychology Learning Memory and Cognition 20 824ndash843

Kaisse EM (1985) Connected speech The interaction of syntax and phonology San DiegoCA Academic Press

Kempen G amp Huijbers P (1983) The lexicalization process in sentence production andnaming Indirect election of words Cognition 14 185ndash209

Kempen G amp Hoenkamp E (1987) An incremental procedural grammar for sentenceformulation Cognitive Science 11 201ndash258

Kenstowicz M (1994) Phonology in generative grammar Oxford BlackwellKiparsky P (1982) Lexical morphology and phonology In IS Yang (Ed) Linguistics in the

morning calm Seoul HanshinLevelt WJM (1989) Speaking From intention to articulation Cambridge MA MIT PressLevelt WJM (1992) Accessing words in speech production Stages processes and

representations Cognition 42 1ndash22Levelt WJM amp Wheeldon L (1994) Do speakers have access to a mental syllabary

Cognition 50 239ndash269Luce RD (1986) Response times Their role in inferring elementary mental organization

New York Oxford University PressMeijer PJA (1994) Phonological encoding The role of suprasegmental structures PhD

dissertation University of NijmegenMeyer AS (1990) The time course of phonological encoding in language production The

encoding of successive syllables of a word Journal of Memory and Language 29 524ndash545Meyer AS (1991) The time course of phonological encoding in language production The

phonological encoding inside a syllable Journal of Memory and Language 30 69ndash89Meyer AS amp Schriefers H (1991) Phonological facilitation in picturendashword interference

experiments Effects of stimulus onset asynchrony and types of interfering stimuli Journalof Experimental Psychology Learning Memory and Cognition 17 1146ndash1160

Meyer AS Roelofs A amp Schiller NO (in prep) Prosodication of metrically regularand exception words in production

Mohanan KP (1986) The theory of lexical phonology Dordrecht ReidelNespor M amp Vogel I (1986) Prosodic phonology Dordrecht Foris

692 ROELOFS

Nooteboom SG (1969) The tongue slips into patterns In AG Sciarone AJ van Essen ampAA van Raad (Eds) Nomen Leyden studies in linguistics and phonetics The HagueMouton

Pederson E amp Roelofs A (Eds) (1995) Annual Report 1994 Nijmegen Max PlanckInstitute for Psycholinguistics

Roelofs A (1992a) A spreading-activation theory of lemma retrieval in speakingCognition 42 107ndash142

Roelofs A (1992b) Lemma retrieval in speaking A theory computer simulations andempirical data Doctoral dissertation NICI Technical Report 92ndash08 University ofNijmegen

Roelofs A (1993) Testing a non-decompositional theory of lemma retrieval in speakingRetrieval of verbs Cognition 47 59ndash87

Roelofs A (1994) On-line versus off-line priming of word-form encoding in spoken wordproduction In A Ram amp K Eiselt (Eds) Proceedings of the Sixteenth Annual Conference ofthe Cognitive Science Society pp 772ndash777 Hillsdale NJ Lawrence Erlbaum Associates Inc

Roelofs A (1996a) Serial order in planning the production of successive morphemes of aword Journal of Memory and Language 35 854ndash876

Roelofs A (1996b) Morpheme frequency in speech production Testing WEAVER InGE Booij amp J van Marle (Eds) Yearbook of Morphology 1996 pp 135ndash154 DordrechtKluwer Academic Publishers

Roelofs A (in press a) The WEAVER model of word-form encoding in speech productionCognition

Roelofs A (in press b) Rightward incrementality in encoding simple phrasal forms inspeech production Verbndashparticle combinations Journal of Experimental PsychologyLearning Memory and Cognition

Roelofs A (submitted a) Implicit and explicit priming of speech production TestingWEAVER

Roelofs A (submitted b) WEAVER 1 1 and other computational models of lemmaretrieval and word-form encoding

Roelofs A Baayen H amp Van den Brink D (submitted) Semantic transparency inproducing polymorphemic words

Roelofs A amp Meyer AS (in press) Metrical structure in planning the production ofspoken words Journal of Experimental Psychology Learning Memory and Cognition

Roelofs A amp Meyer AS (in prep) Time course of metrical planning in speechproduction Sufxes and clitics

Roelofs A Meyer AS amp Levelt WJM (1996) Interaction between semantic andorthographic factors in conceptually driven naming Comment on Starreveld and La Heij(1995) Journal of Experimental Psychology Learning Memory and Cognition 22 246ndash251

Schiller NO Meyer AS Baayen RH amp Levelt WJM (1996) A comparison of lexemeand speech syllables in Dutch Journal of Quantitative Linguistics 3 8ndash28

Selkirk E (1984) Phonology and syntax The relation between sound and structureCambridge MA MIT Press

Sevald CA Dell G amp Cole JS (1995) Syllable structure in speech production Aresyllables chunks or schemas Journal of Memory and Language 34 807ndash820

Shattuck-Hufnagel S (1979) Speech errors as evidence for a serial-order mechanism insentence production In WE Cooper amp ECT Walker (Eds) Sentence processingPsycholinguistic studies presented to Merrill Garrett Hillsdale NJ Lawrence ErlbaumAssociates Inc

Spencer A (1991) Morphological theory Cambridge MA BlackwellStemberger JP (1982) The lexicon in a model of language production Unpublished

doctoral dissertation University of California San Diego CA

SYLLABIFICATION 693

Stemberger JP (1984) Length as a suprasegmental Evidence from speech errorsLanguage 60 895ndash913

Stemberger JP (1985) The lexicon in a model of language production New York GarlandStemberger JP (1990) Wordshape errors in language production Cognition 35 123ndash157Stemberger JP amp MacWhinney B (1986) Frequency and the lexical storage of regularly

inected forms Memory and Cognition 14 17ndash26