the representation of polysemous words - nyu … · the representation of polysemous words ......
TRANSCRIPT
Journal of Memory and Language 45,259–282 (2001)doi:10.1006/jmla.2001.2779, available online at http://www.academicpress.com on
The Representation of Polysemous Words
Devorah E. Klein and Gregory L. Murphy
Department of Psychology and Beckman Institute, University of Illinois at Urbana-Champaign
Words that have a number of related senses are polysemous. For example,paper refers to both a substanceand a publication printed on that substance. Five experiments investigated whether different senses are repre-
s wordse consis-dgmentse forthatinhibits
tations for
M one
Thand DellKaricom
AdInstiUrba
sented distinctly in the lexicon or if there is a common, core meaning. In all experiments, a polysemouwas used twice, in phrases that selected the same or different senses. Experiment 1 showed that sentency aided memory for the polysemous word. Experiment 2 extended this result to a timed sensicality jutask. Experiment 3 demonstrated that the effects for polysemous words were very similar to thohomonyms. Experiment 4 ruled out the possibility of modifier–modifier priming. Experiment 5 showedsense consistency facilitates comprehension relative to a neutral baseline, while sense inconsistencycomprehension. These experiments provide evidence that polysemous words have separate represeneach sense and that any core meaning is minimal.© 2001 Academic Press
Key Words:polysemy; lexical semantics; psycholinguistics; ambiguity.
uch research in lexical representation hasthese uses of the word paperare related to
dnt
idin
a
nde
s ofanceisednon-
o-eted.d’s(seeb-ob-at
ela- toes itond thee a as
compared homonyms with unambiguous worHomonyms usually arise through a historical acident in which two different word meaningconverge on the same phonological represetion, or in which a single word diverges invery different meanings. A typical example bank, with the unrelated meanings of a financinstitution and the side of a river. Most words not have such unrelated meanings, but linguand psycholinguists studying lexical meanihave nonetheless identified a wide rangemeanings within individual unambiguouwords. For example, the word papermay referto a substance made out of wood pulp, a blsheet of that substance, a daily publication, orarticle that is printed on that substance. Tmeaning of paperhas been extended so far thit is now possible for students to turn in a papby handing in a disk or sending a file electrocally, so that wood pulp is in no way involveUnlike the meanings of a homonym, howev
25
is research was supported by NIMH grant MH417an NSF graduate fellowship. We are grateful to G
and Kristine Onishi for insightful comments and tn Humphreys for linguistic assistance as well as helpments.dress correspondence to Gregory L. Murphy, Beckm
tute, University of Illinois, 405 No. Mathews Ave.na, IL 61801. E-mail: [email protected].
t
s.c-sta-
oisalostsgof
s
nk anheateri-.r,
another and clearly arose through a procesextension of similar meanings rather ththrough an arbitrary historical coinciden(Clark & Clark, 1979; Sweetser, 1990). Thphenomenon of words having multiple relatsenses is called polysemy.1 Rather than being aexception, polysemy can be found in most ctent words to at least some degree.
Whereas the alternative meanings of a homnym like bank have no obvious relation, thsenses of polysemous words are clearly relaCertain semantic relations between a worsenses appear over and over in polysemy Lehrer, 1990; Nunberg, 1979), for example, oject/substance, object/representation of that ject, type/token, and text/object containing thtext. To illustrate one, the object/substance rtion is found when the same word is usedrefer to an object and the substance that makup, often becoming a mass noun in the seccase. For example, an oak is a tree that issource of some oak (wood) used to maktable; a chicken is an animal that may end up
04ary 1 We will follow linguistic usage in calling the differen
0749-596X/01 $35.00Copyright © 2001 by Academic Press
All rights of reproduction in any form reserved.
9
oful
an,
meanings of polysemous words senses, generally referringto homonyms as having different meanings. However,meaningwill also be used as a term covering both of thesecases. As we will discuss, the distinction between a polyse-mous sense and a homonymous meaning can sometimes bedifficult to draw.
soas
me
eg
ee
t
ao
s-cim
or
Tnimto
ap-sta,
arery,arehehatnd
eni-oly-ted
sub-y isntly,ep-the
.g.,as a
le,f-eir.”e oftwoby
ase
-am--
hatirely are in-nstpre-that
260 KLEIN AND
some baked chicken (the food). If we werehear of a new species of plant called a delgar,we could say both “There is a delgar growingmy yard” (individual plant) and “This pen imade out of delgar” (the substance derived frthat plant). Thus, these forms of polysemy highly productive, and they are used quite eawhen new words enter the lexicon (Murph1997). For example, the word bookcan be usedto refer both to a physical object containingtext and to the content of that text. The saform of polysemy is present in recently inventwords for new information-storage devices suas videotape, CD, and DVD; for example,ThatCD is cracked(object) and That CD is brilliant(content of the CD).
The outlines of a theory of homonymic reprsentation are fairly clear. The different meaninof bank or calf are considered to be differenwords, so it is generally believed that they arepresented by differentlemmas(lexical units—see Levelt, 1989). In lexicology, there alsseems to be a belief that homonyms are differwords, as indicated by separate dictionarytries (Zgusta, 1971, p. 74; also shown belowThis is not to say that it is understood exachow listeners identify which meaning of a homonym is intended but that there is fairly gooagreement that these meanings are lexicseparated. There is no such agreement for psemy. Should the sense ofpapermeaning “sub-stance made from wood pulp” be in the samlexical entry as “sheet of writing material [madfrom that substance]”? Should the two sensebookmeaning “object with print” and “a particular text” be in the same lemma? Although susenses seem closely related, they are sometontologically different things. In the sentenc“Your book is green,” a physical object is beintalked about, but in “Your book was difficult tunderstand,” the textual content is beingferred to. If the senses are switched across stences, the sentences may no longer be true:textual content was not green; the object wasdifficult to understand. Linguists have variedtheir approach to this problem, ranging frosuggesting that there is a single represensense that accounts for all these uses of a w
(Ruhl, 1989) to arguing that each distinguishMURPHY
to
in
mreilyy,
ae
dch
-s
tre
ontn-).
ly-dllyly-
eeof
hes
eg
e-en-heot
n
edrd
able sense is separately represented (e.g.,proaches based on lexicology, such as Zgu1971).
Thus, the questions of how many senses represented, how they are linked in memoand how they are coordinated in processing the critical issues surrounding polysemy. Tpresent experiments aim to provide data twill constrain accounts of the representation aprocessing of polysemy.
Given the differences just described betwepolysemy and homonymy, it is obviously critcal to keep these two phenomena distinct. Psemy is the normal, expected presence of relasenses in a word, such as an object and the stance making up that object, and homonymthe unpredictable coincidence of two differewords having the same name. Unfortunatepsychologists have not been very good at keing these two terms separate. In particular,term polysemy, which is used in linguistics torefer to a word having related senses (eCruse, 1986; Geeraerts, 1993), is often used synonym for ambiguity(including homonymy)in the psychological literature. For exampHino and Lupker (1996) titled their article “Efects of Polysemy . . .” and then refer to thstimuli as “ambiguous” and “unambiguousThis is fairly typical of the terminology in thpsychological literature. Furthermore, studiesambiguity have sometimes combined these phenomena in their experimental designs treating homonyms and polysemous wordsboth “ambiguous.” In the present article, wwill use the term polysemyto refer to the phenomenon of related senses in otherwise unbiguous words (i.e., following linguistic practice), and we will use ambiguity to refer towords that are homonyms or homographs.
The Problem of Polysemous Representation
Whereas homonyms are different words thappen to share the same name, it is not entclear how polysemous words, whose sensesmore closely related, are represented. In anfluential paper, Nunberg (1979) argued agaithe idea that all distinct senses should be resented in the lexicon. Instead, he proposed
-pragmatic principles could be used to derive
O
d
thsegn
nourth
r2
lthic
eh sot
aehfrs oma6ihuocs
s
rei-
hegfreus
, as&
atgh a
wr-er-alc-e-nd
in-us,ichre6),et-r-nse
ly-eral.ex-edon. notple
an-b-
ista-
estheialt
arti-ly,mery
REPRESENTATI
word senses from others. For example, if dogrefers to a class of animals, one could easilyrive the use of this word to refer to an individuanimal in this class. We should emphasize Nunberg’s argument concerned how a linguitheory of the lexicon should represent differsenses. It was not presented as a psycholotheory of representation and processiNonetheless, one can readily construct a pchological theory from this proposal. Accordito this theory, all that is represented is a cmeaning of a word. The different polysemoextensions are generated on the fly, using pmatics and plausible reasoning. Thus, on view, different senses are not prestored butrather computed from contextual features.
A similar account was proposed by Camazza and Grober (1976), who identified separate but related senses for the word line.They suggested that these senses are all reto a core meaning and that “it is precisely core meaning that is stored in the psychologrepresentation for the meaning of line” (p. 188).They argued against the notion that each sis explicitly stored in the mental lexicon. Ru(1989) went so far as to argue that there is agle, defining sense for words (even mhomonyms), with distinct senses neither creaor stored.
Lehrer (1990) agreed with Nunberg thmuch polysemy can be predicted through geral principles of meaning extension, but salso noted that these principles sometimes She argued that the lexicon is simply unpdictable to some degree and that language umust learn which words can be extendedwhich ways, rather than relying entirely pragmatic principles. That is, at least sosenses must be explicitly represented (see Rice, 1992). In lexicology, Zgusta (1971, p. 6argued that it is usually impossible to find a sgle basic sense of a word from which the otsenses can be derived. Thus, within the lingtic literature there is a variety of views on hexplicit the lexicon must be—whether eaword sense must be represented or instead irived from a more basic or core meaning.
Few psychological studies have addres
this issue. Williams (1992) found that contextuN OF POLYSEMY 261
e-alat
ticnticalg.sy-gres
ag-is
are
a-6
atedeal
nselin-sted
tn-e
ail.e-ersinnelso)
n-eris-wh de-
ed
ally irrelevant senses of polysemous words aactive even over long delays in a lexical decsion task. He compared this to results from thomonym literature, which show that priminfor the contextually irrelevant meaning ohomonymous words is short-lived. TherefoWilliams argued that the senses of polysemowords cannot be represented independentlyhomonym meanings are (see also DurkinManning, 1989). One possibility, then, is ththe polysemous senses are connected throucommon core.
Additional support for the core concept vieof polysemy can be found in the work of Andeson and Ortony (1975), who argued that undstanding is more than finding the correct lexicentry in a semantic associative network. Acording to their work, a polysemous word drives rich representations from its sentence, aboth context and world knowledge must be volved in deriving those representations. Ththey argued that semantic memory is not renough to explain how polysemous words ainterpreted. Like Caramazza and Grober (197they seem to be suggesting that the lexical nwork contains core information, and other infomation necessary to understand the exact seof the word is supplied by context.
Core-meaning theories suggest a view of posemous senses as being somewhat ephemLexical meanings can be augmented or tended in a given context, but those extendsenses are not permanently stored in the lexicEven having created an extension once doesmake a subsequent creation easier on a simcore-meaning account, since it is the core meing that is retained. (We will discuss more elaorate views in the General Discussion.)
Another view of polysemy representation one that is much closer to homonym represention. According to this view, common senswould have separate entries connected to same lemma. For example, the writing matersense of paper might be its core meaning, buother senses such as a daily periodical or an cle would be represented distinctly. Presumabthese senses would all be linked to the salemma, unlike homonyms. Such views can va
-considerably, depending on how many senses
ecn oavtiapi8
9
xonshe)r
ho
ogpe
inin,
wbm
brdin
thdaus
ns.ord the
notrd.ssi- is in-hethat all
con-an-hecal the.
are asore all andightor-ely
ofd-ion
ap-ingsedred-inol-ofin-ot
at-ininadif-ndi-etion
on-
262 KLEIN AND
they claim are represented and whether onthese senses is picked out as being the meaning. An important question surroundithis approach is how to decide which sensesdistinct (e.g.,dog used to refer to the class dogs vs an individual dog? dogas an animal vs kind of meat?) and when a sense rises abomere nonce usage to deserve full representaMany linguists appear to take the view thsome reasonable number of senses are resented, rather than only a single (core) meanor every possible sense (e.g., Cruse, 19Deane, 1988; Langacker, 1987; Rice, 19Tuggy, 1993).
As already mentioned, there is very little eperimental evidence to support either the cor a multiple-sense theory or to provide costraints on either view, and what evidence exiis often muddied by the use of homonyms in tpolysemous stimuli. One study that focussolely on polysemy was that of Murphy (1997which showed that novel extensions of a wothat were closely related to previously knowsenses were more acceptable than were mdistantly related extensions. This suggests tpolysemy could develop by the constructiona chain of extensions, each building on its preecessors (as proposed by Heine, 1992; Lak1987; Malt, Sloman, Gennari, Shi, & Wan1999). Furthermore, this is evidence that peocan create novel extensions of words in contand that not all senses of a word need to be pstored to be understood—a conclusion reforced by research on nonce uses of existwords (e.g., Clark & Gerrig, 1983; Gerrig1989)—which is consistent with a core vieHowever, although novel uses of a word maycomprehensible, this does not mean that comonly encountered senses are not stored. Mphy used novel words and novel extensions,it is possible that many senses of actual woare represented in memory. It is our goal tovestigate this question.
The main question being investigated in present experiments is the degree to which ferent senses of polysemous words use the sor different representations. If polysemowords have only a core meaning, specific sen
being derived online, then different uses ofMURPHY
oforegaref
e aon.tre-
ng6;
2;
-re-tsed,dnoreatf
d-ff,,lextre--g
.e-
ur-uts-
eif-meses
word would have highly shared representatioConversely, if each sense of a polysemous wis encoded and represented separately, thenrepresentation of one use of the word might overlap that of a different use of the same woThere are also a number of intermediate pobilities, in which a core part of the meaningshared by most senses, varying in how muchformation is in the core and how much is in tsenses. For example, one might believe there is a very abstract core that is present inthe uses of a word, but each sense provides siderable detail to flesh out its particular meing. Alternatively, one might propose that tcore is a rich representation of the prototypiuse of the word, and the senses provide onlyminimal information that distinguishes themAlong with each representational view there a number of possible processing accountswell. For example, one might argue that the csense is activated first, since it is common toor many uses; one could argue that the coreother senses are activated in parallel; one mclaim that all senses are activated until the crect one is selected, or that only the most liksense is activated.
The problem in beginning an investigationpolysemy is that there are few explicit moels of the representation and comprehensprocesses of polysemous words. Linguisticproaches virtually never discuss processissues. The psychological literature has focuon the single-sense notion but has not explomost of the other possibilities. It is clearly impossible to sort through all these possibilitiesa single study. Our approach has been to clect data that will act as constraints on allthese theories. We will argue that the resultsdeed rule out some of them, even if they do nyet determine which one is correct. Our stregy was to investigate the amount of overlapdifferent senses, using a priming techniquewhich one use of a word was followed bysubsequent use that involved the same or aferent sense. Differences between these cotions indicate the amount of overlap of thsenses’ semantic representations. The quesis whether there is enough semantic comm
aality across different uses of a word to indicate
O
arp
nrl
s
rtscr
u
rismannoae
o
oeins
pths
,Itn
oe
r
its
in
-oly-tede aenteri-by
y,
han
r-edrds.in
ryer,bee-hesep-le
enon
ex-aremsht
his
d-onu-ext,onle-ad-seee
REPRESENTATI
a possible shared core meaning. As we hjust pointed out, even if one believes in a coit might range from being the entire stored reresentation of a word to a minimal, abstracomponent. If our results show that differesenses have considerable semantic ovethen single-sense views or any view withsubstantial core will be supported. In contraif the results show minimal overlap, then seprate-sense views or a view with a minimal cowill be supported. The initial experimenbegin by looking at whether there are distintions between different senses of a single woLater experiments attempt to gauge the sizesuch distinctions by comparing polysemowords to homonyms.
Although our results will not be able to narow the field down to a single model, itnonetheless important to begin to perform epirical work that will elucidate the representtion of polysemy, because of its implicatiofor our understanding of lexical processiand representation. Gerrig (1986) pointedthat considerable psycholinguistic researchdresses how meaning is used in lexical accand discourse comprehension, yet there istle agreement on exactly what semantic infmation is included in lexical representationFor example, psychologists argue abwhether meaning is selectively or exhaustivaccessed (especially for homonyms) durcomprehension, yet neither they nor linguiagree on just what that meaning is—whatand is not included in a word’s semantic reresentation. Without an understanding ofcontent of lexical representations, it is imposible to decide whether access is selectivein what ways it is and is not selective.Caramazza and Grober (1976) are correcsaying that there are (at least) 26 distisenses of the wordline and 40 senses forrun,for example, one might be reluctant to prpose that all these senses are accessedtime the word is encountered. In contrast,the representation of such polysemous wois a core sense or only a few senses, thenhaustive access seems a likely possibilThus, specifying the representation of poly
mous words is a necessary part of explaininN OF POLYSEMY 263
vehow meaning is represented and involved
e,-cttap,at,
a-e
-d.ofs
-
--sgutd-ss
lit-r-s.utlyg
tsis-e-or
fin
ct
-achifdsex-y.e-
production and comprehension.
EXPERIMENT 1
The first experiment used memory performance as a measure of the representation of psemous senses. In particular, it investigawhether people are better able to recognizword used in the same sense or a differsense than its original presentation. The expment was based on a paradigm developedLight and Carter-Sobell (1970). In their studsubjects saw phrases liketraffic jam and thensaw a phrase likestrawberry jam and wereasked if they had previously seen the wordjam.Performance was worse in such cases twhen they sawraspberry jam followed bystrawberry jam, which usesjam in the sameway. In Light and Carter-Sobell’s (1970) expeiment, most of the test words would be countas homonyms rather than as polysemous woIt is perhaps not surprising that using a wordone way makes it difficult to retrieve a memoof the word used in an unrelated way. Howevit is not so clear that the same effect wouldfound with the highly related senses of polysmous words. For example, would one find tsame decline in recognition if one were to uthe wordpaper to mean newspaper and wraping paper? If the two uses draw on a singcore meaning, recall might be quite high evwhen the sense is changed. If the senses relyseparate representations, recall would bepected to be lower when different sensesused in learning and test. Because homonydo not share a common core meaning, Ligand Carter-Sobell’s results do not speak to tpossibility.
A related experiment was Perfetti and Gooman’s (1970) study of semantic constraintsdisambiguation. They found that when ambigous words were presented in a sentence contsubjects later false-alarmed in a recognitimemory test to words associated with the revant sense of the word. For example, after reing a sentence like “Many families rent a houin the country for the summer months” for thcritical word country, subjects later were mor
glikely to say that the wordcity (related to the
a
fu
od
ss
aruoicoa
nelels.ilds
te
azhop e
e
gr-e ofed.
ith-nt
y re-by
ns,ous
n-
hesenoordererd.todsesen-sedheforiven
uldof
gh
r-an
althease
264 KLEIN AND
rural sense ofcountry) but not the wordnation(related to the political sense of country) happeared in the list. In another experiment, Pfetti and Goodman presented a list of wordssubjects to learn. They did not find analogocontext effects when the context was the won the list just prior to the ambiguous worThus, their results suggest that the context osingle word may not be sufficient to distinguithe senses of a polysemous word in the preexperiment.
However, the Perfetti and Goodman (197study is not entirely appropriate for investiging the issues raised in the present enterpThey did not look at memory for the ambiguoword itself but instead at false alarms to assates. In Experiment 1, we looked at recognitmemory for the polysemous word itself, whimay be a more sensitive measure. Furthermthe Perfetti and Goodman stimuli seem to hincluded both polysemous items (like country)and homonyms. They argued (p. 427) that o4 of the 30 items used in most of their expments were polysemous. So, again, it is uncwhether evidence of distinct memory represtation will be found for words that have highrelated semantic representations. The preexperiment attempts to answer that question
The experiment had two parts. In the learnphase, subjects read phrases and were tostudy them for a later memory test. The phracontained polysemous words, biased in interptation toward one of two senses. In the phase, subjects viewed similar phrases in whone of the two words was capitalized. Their twas to decide if they had seen the capitaliword before. In experimental conditions, tword that was capitalized was the polysemword, which could be in the same phrase as viously seen, in a different phrase that usedsame sense, or in a phrase that used the alttive sense. However, the capitalized word walways the repeated word. Therefore, in theperimental conditions, the response shouldways be YES. In foils, the noncapitalized wowas repeated.
If polysemous words are represented wseparate sense representations, memory sh
be better if a word is used in the same sense tMURPHY
der-ors
rd.f ahent
0)t-ise.sci-onhre,ve
lyri-earn-yent
ng toes
re-st
ichskedeusre-therna-asx-
al-rd
ithould
if it is in a different sense. If the core meaninview is correct, there should be no such diffeence, as subjects will access the single sensa polysemous word every time it is encounter
Method
Materials. Twenty-four polysemous wordswere used (see the Appendix). We started wDurkin and Manning’s (1989) list of 175 polysemous words and their most common differesenses and selected senses that were clearllated. Polysemy of senses was also ensuredconsulting the words’ listings in theOxford Eng-lish Dictionary. The meanings of homonyms ithis dictionary are given separate entriewhereas the related senses of our polysemitems were all listed under a single entry.
These potential stimuli were then normed oour subject population. Thirty-six subjects produced definitions of polysemous words, and ttwo most common senses produced were chofor use in the first experiment. After the twsenses were chosen, two phrases using the wwere constructed for each sense so that thwere four phrases for each polysemous woFor example,paper is a polysemous word thacan mean sheets of material made from wopulp or a newspaper. (Note that our paraphraof these senses are simply expository conviences. Since the descriptions were never uin the experiments, they have no bearing on tresults.) There were two phrases createdeach of these senses, as in the examples gbelow.Sheets of a material Newspaper
wrapping paper daily papershredded paper liberal paper
The phrases were selected so that there wobe minimal semantic overlap in the modifiers each sense. For example,daily and liberal arenot from the same semantic field, even thouthey both serve to mark paper as indicating anewspaper. In addition, there was minimal mophological overlap between the modifiers of item (e.g., two modifiers ending in -ing), andconsistent pairs did not share morphologicstructure more than did inconsistent pairs. In learning phase, subjects might see a phr
hanusing one of the senses of paper, either wrap-
n
e
b
r
tut
ce
2o
o
i
ae
a
uc-edtestec-idehadd,ey toordslso
ingut
-er-.
ass)
of
tely
ing
, it
notn-atesed
reses,
sesn-
t
REPRESENTATIO
ping paperor daily paper. In the test phase thewould see either the same phrase repeatedsame phrase condition), a new phrase using thsame sense (the consistent sense condition), or anew phrase using a different sense (the inconsis-tent sense condition). So, the test item shreddedPAPERwas in the consistent condition whewrapping paperwas studied and was in the iconsistent condition when daily paperwas stud-ied. Note that the critical comparison (of constent vs inconsistent conditions) always involva new modifier.
Counterbalancing items across conditionsquired six lists (each of two test phrases inquadruple was preceded by a phrase that identical, consistent, or inconsistent). Thunone of the effects could be attributed to osense being more familiar or memorable ththe other, since both senses appeared in consistent and inconsistent conditions.
Foils also consisted of pairs of phrases, oappearing in the study portion of the expement and the second appearing in test. Asthe experimental items, one word of the phrawas repeated, but unlike in the experimenitems, the repeated word was not the word sjects were asked to judge. For example,study phrase might betennis ball, which wouldbe followed in the test phase bytennis SHIRT,where subjects were asked if they had seshirt before. These foils were used to forsubjects to focus on the critical word, raththan allowing them to respond positively if onof the words seemed familiar. There werefoils and 24 experimental items, for a total48 test trials.
Procedure. Subjects viewed the materials a Macintosh Quadra 630 computer, which wconnected to a PsyScope button box (CohMacWhinney, Flatt, & Provost, 1993). Thedominant hands were assigned to the YESsponse and their nondominant hands to the response. For the first phase, subjects werestructed to study the phrases that would viewed for a later memory test. Phrases peared on the computer screen for 2 s apiwith 2 s between trials. Each phrase was viewonce. At the end of this phase, instructions
peared telling subjects about the recognitiN OF POLYSEMY 265
y(thee
n-
is-d
re- awass,neanoth
nei-insealb-
he
ener
e4f
nasen,rre-NO in-bep-ce,edp-
memory test. They also received verbal instrtions from the experimenter and were allowto ask any questions about the test. The phase began immediately afterward. In this sond part, subjects were instructed to decwhether the capitalized word in each phrase appeared in the first part of the study. If it hathey pressed the YES button; if it had not, thpressed the NO button. Subjects were toldread the whole phrase, because the other wmight serve as a memory aid. They were atold to go as fast as possible, without makmistakes, although no explicit feedback aboerrors was given.
Subjects. Subjects were 61 introductory psychology students who participated in the expiment for partial fulfillment of course creditThey were all native speakers of English.
Results and Discussion
One subject who had over 50% errors wdropped from the study. Reaction times (RTwere not analyzed, given the large numbermissing RTs (due to memory errors).
Same phrase items were the most accuraevaluated (79% correct) (SD 5 19%), followedby the consistent sense items (64%) (SD 525%), with the inconsistent sense items bethe most error prone (56%) (SD 5 24%),F1(2,118) 5 23.14,p , .001; F2(2,46) 5 25.75,p , .001. For example, when paper was seeninitially in a phrase like wrapping paper, whichsupported the “sheets of a material” sensewas easier to verify having seen paper whenshredded PAPERwas presented at test thawhen liberal PAPERwas presented at test. Nsurprisingly, seeing a word in an identical cotext was the most helpful. These results indicthat the way a polysemous word is procesinitially affects later memory access.
The repeated items were significantly moaccurate than the consistent sense phrat(59) 5 4.25,p , .001;t(23) 5 5.80,p , .001.Most importantly, the consistent sense phrawere reliably more accurate than the incosistent sense phrases,t(59) 5 2.75, p , .01;t(23) 5 2.23,p , .05. In fact, the inconsistencondition was not reliably different from
onchance (p’s . .05).
D
e em
oote st
rtb
is nrl r
eeiduu
raa
inatraedre
rgleoti
ote ofandrentsur-n-singsk
t onoryg-
gaintingentf aon-
devers-g
asendg,ceon-ntean-s,aedthetlityof
g.,
rusde-on
266 KLEIN AN
These results lend support to the hypoththat senses are stored separately and aredence against a single core meaning hypothIf the words were interpreted in terms of a comon core meaning (as suggested by a msemy view such as Ruhl’s, 1989), memwould have been equivalent for the consisand inconsistent senses. Thus, these resultsgest that polysemous senses may be stored rately. At the very least, they suggest that senses are functionally distinct in that samsense uses are more related than are diffesenses uses of the same word. The fact thasense-inconsistent condition was not reliadifferent from chance indicates that if there core sense, it is not at all strong, since it didprovide a basis for memory. Semantic oveacross senses is less, then, than might bepected, a conclusion that is examined furtheExperiment 3.
This conclusion contrasts with that of Pefetti and Goodman’s (1970) study describabove. In their experiment, a single contword was not sufficient to force subjects to dtinguish the senses of the learned worwhereas a sentence context was. Althothere are many differences between their stand ours, perhaps the most important onethat our subjects read the words as a phand therefore presumably interpreted themcoherent concept (Gerrig & Murphy, 1992For example, in order to understandliberalpaper, a subject would have had to determthat paper referred to a newspaper, rather thto a sheet of blank paper, say, or elsephrase would not have made sense. In contPerfetti and Goodman’s context was a precing item on a list of to-be-remembered worso subjects were not encouraged to integthe stimuli. Also, as mentioned earlier, our dpendent measure was memory for the taword itself, whereas their measure was thedirect measure of false alarms to another wrelated to one sense of the target word. Aseffect size was only 8% in our experiment, itperhaps not surprising that a less direct me
ure would not obtain a significant differencbetween consistent and inconsistent senses.MURPHY
sisevi-sis.-
no-rynt
sug-epa-hee-ent- thely aot
apex- in
r-dxts-s,ghdyisse
s a).
en
hest,d-s,ate-etssrdhesas-
EXPERIMENT 2Our main interest in Experiment 1 was n
memory for senses per se, but rather the issuhow polysemous words are represented processed. The results did suggest that diffesenses are stored separately. This result is prising enough to follow up in a task that ivolves comprehension and semantic procesrather than memory. In this experiment, the tawas to make a sense/nonsense judgmenphrases similar to those used in the first memexperiment. The RT and accuracy of this judment were the dependent measures. We amanipulated sense consistency by presenphrases that involved the same or a differsense of a word and by looking for priming oconsistent use of a word compared to the incsistent use.
For example, subjects might seeliberalpaper and have to say whether or not it masense. On the previous trial, they would haseen eitherdaily paper (consistent sense) owrapping paper(inconsistent sense). The quetion, then, is whether the difficulty in decidinthe sensicality ofliberal paperdepends on theconsistency in sense of the prior use ofpaper.(Foils were used so that a sensical first phrdid not predict the sensicality of the secophrase.) Again, if there is only a core meaninconsistency should not make a difference, sinall phrases would require access to the core ccept of paper. We used a sensicality judgmebecause it requires subjects to access the ming of the word, unlike lexical decision taskwhich only require subjects to verify thatstring is a word. Although meaning may be usin such a task, it may not be accessed atlevel of detail that would distinguish differensenses. Furthermore, judging phrase sensicahas been shown to be a sensitive measureconceptual processing in previous work (e.Murphy, 1991).
Bainbridge, Lewandowsky, and Kirsne(1993) performed a similar study on polysemowords using sentence contexts and a lexicalcision task. Subjects made lexical decisions
ewords twice—preceded by contexts that evokedeither the same or different senses across trials.
ereeesao
dimn1
id
a
dfee
theg
n
tar
hit.
enthnh
1
theitlyowldutselfftered-theas
250 be-asat
ess
y-ri-
t.
ndectsad
gorsen-of ereorTs
ananthub-se,
la-he
esthe the
they
REPRESENTATIO
They found that polysemous words were judgfaster in the second trial when they were pceded by the same-sense context. Bainbridgal. concluded (p. 624) that the different sensare represented separately, and that priming ia large degree dependent on activating the ssense of a word. However, their study did nseparate repetition of a sense from repetitionthe exact sentence context. That is, their contion with different senses (necessarily) had dferent sentence contexts, whereas their sasense condition had the exact same sentecontext on both trials. As our Experimentshowed, seeing the exact same context provmore priming than simply activating the samsense. (In fact, the repetition effect there wlarger than the sense consistency effect.)order to fairly compare the same-sense andferent-sense conditions, one needs to use difent contexts in both cases, as in the presentperiment.
Method
Materials. The same phrases were used incritical trials as in Experiment 1. These phraswere counterbalanced, so that a subject misee the prime phrasewrapping paperfollowedby eithershredded paper(consistent) orliberalpaper (inconsistent) as targets, and a differesubject might seedaily paperas a prime, againfollowed by eitherliberal paper (now consis-tent) orshredded paper(inconsistent), requiringfour counterbalancing lists. Within each lishalf the critical items were consistent and hinconsistent. Each subject saw 24 total expemental pairs of phrases and 72 foil pairs. Tfoils were also two phrases sharing a word, wat least one phrase that did not make senseone-third of the foils, the first phrase (historylecture) made sense while the second (yellowlecture) did not. One-third of the pairs had threverse pattern, and the final third had two nosensical phrases. As a result of these foils,sensicality of the first phrase did not predict sesicality of the second across the experiment. Tsame equipment was used as in Experiment
Procedure. To ensure that subjects understood the instructions, there was a set of eig
N OF POLYSEMY 267
d-etsto
metofi-
f-e-ce
esesInif-r-x-
esht
t
,lfi-ehIn
-e-e
.
practice trials. Subjects were asked to judgesensicality of each phrase. They were explicinstructed that sets of two phrases in a rwould have a word in common. They were toto respond as quickly as they could withomaking errors. Each phrase appeared by iton the screen until subjects responded. Aevery trial, subjects received feedback: Feback telling them they were correct was on screen for 1 s, while feedback after an error won the screen for 2 s. There was a pause ofms between the end of the feedback and theginning of the subsequent item, and there wno particular marking of the phrase pairs thshared a word. The entire experiment took lthan 20 min.
Subjects. Subjects were 27 introductory pschology students who participated in the expement for partial fulfillment of course crediThey were all native speakers of English.
Results and Discussion
In order to be able to analyze RTs, in this asubsequent experiments we eliminated subjwho made too many errors and therefore hfew trials per cell. Our criterion for droppinsubjects was making more than 20% erroverall (which suggested a general lack of atttion) or making errors on more than 20% the experimental trials. Seven subjects wdropped from Experiment 2 on this basis. Fthe remaining 20 subjects, experimental Rlarger than 3 SDs above each subject’s mewere omitted. Trials in which subjects made error on the prime were omitted from boanalyses, since we could not be sure that sjects had correctly processed the prime phrawhich constituted the experimental manipution. This resulted in the exclusion of 1% of tdata.
The error analysis of the critical phrasshowed that it was easier to understand phrase when the repeated word was used insame sense than in a different sense,F1(1,19) 514.77,p , .005; F2(1,23) 5 10.12,p , .005.When subjects received consistent phrases,
-htwere correct 96% of the time (SD 5 5%),whereas when they received a prime and target
nmo
r
o
n
nsaxi
m
nn
l
mhtiei
e.s-if-llyentithircyg
e--
&e,nr-itbe
dsn-s-
tone
heusichd. toy-
inthrehbe
268 KLEIN AND
using different senses of the polysemous wothey were correct 87% of the time (SD5 10%).
The RT analysis also showed the consisteeffect, although it was only reliable in the iteanalysis. When the prime and target were csistent, sensicality was judged more rapid(792 ms;SD 5 96) than when prime and taget were inconsistent (859 ms;SD 5 118),F1(1,19)5 1.82,p . .10;F2(1,23)5 9.64,p 5.005. The lack of significance in the subjeanalysis is probably due to the small numberitems per subject, once trials with errors (primes or targets) were removed.
When the second occurrence of a word wconsistent with the first one, subjects wemore likely to judge the phrase as makisense, and they tended to make this decisfaster. This constitutes further evidence agaia single core meaning, using an online taOur results confirm those of Bainbridge et(1993), who had a similar design using a lecal decision task. But, as discussed earlBainbridge et al. used the exact same contein their consistent condition, and so their resucould have been an effect of receiving the sacontext on both trials. In our experiment, thcontexts were different in both conditions, awe still found an influence of same vs differesense.
The results of Experiments 1 and 2 use vedifferent techniques to converge on the concsion that the different senses of a polysemoword are functionally distinct. In Experiment 1we argued that there was little absolute comonality among the different senses of tsame word, because the inconsistent condiwas not different from chance. Such a clacould not be made in Experiment 2, howevbecause there was no lower baseline agawhich the inconsistent sense condition couldcompared. The next experiment provides suc
enue
realalfo-
baseline.
EXPERIMENT 3
Experiment 3 served both to replicate Expiment 2 and to gauge the size of the consisteeffect. The consistency effect arises becadifferent uses of a word involve different s
mantic properties. The stronger such an effeMURPHY
rd,
cy
n-ly-
ctofn
asregionstk.l.i-er,xtslts
eedt
ryu-us,-
eionmr,nstbeh a
r-cyse-
the less significant a core meaning could bHomonyms provide a useful benchmark for etimating the size of the effect, because the dferent meanings of a homonym are essentiaindependent. While a sense/nonsense judgmtask has not been done, to our knowledge, whomonyms, one would expect that since themeanings are far more distinct, the consisteneffect would be far larger. For example, readinthe word bank to mean a financial institutionwill have a severe cost when the word was prviously used to refer to a mound of earth (Swinney & Hakes, 1976; Tabossi, 1988; TabossiZardon, 1993; Van Petten & Kutas, 1987). Thprime will activate one lemma for this wordwhich will then be very strongly activated whethe word occurs again. When this is the incorect lemma (in the inconsistent condition),must be suppressed and another one mustselected. If the senses of polysemous woroverlap, then they should reveal a smaller cosistency effect. Furthermore, the linguistic asumption is that these senses are connectedthe same lemma, so selection of a new owould not be necessary.
In sum, homonyms provide a way to scale tsize of the consistency effect in polysemowords, since they represent the case in whdifferent meanings are completely unrelateThis experiment also serves as an attemptreplicate the consistency effect found with polsemous words in Experiment 2.
Method
The method was identical to that usedExperiment 2, with the one change that bohomonymous and polysemous stimuli weused. For the experimental conditions in whicthe homonyms appeared, a subject mightasked to judge the sensicality ofcommercialbank and thensavings bank, which use thesame meaning ofbank. However, another sub-ject might seecreek bankand thensavingsbank, which use different meanings ofbank.Polysemous and homonymous items weused in the same list. Half of the experimentitems a subject saw were consistent and hwere inconsistent. The subjects were 34 intr
ct,ductory psychology students who participated
O
f
sess
r
trrfuehd
fh
rim
phinew
cmetoe
he
forothts,
and
p-oly-
as.
of are-eenim- theglyameaseect07s),his
ngdif-di-24calo-x-t,
nm-ones,
hely-rorsis
theorthein
REPRESENTATI
in the experiment for partial fulfillment ocourse credit. They were all native speakersEnglish.
A slightly different set of polysemouphrases was used in Experiment 3; modifiwere replaced in some of the less succesitems from the prior experiments (e.g., thowith many errors). There were a total of 2polysemous words and 24 homonymous wo(see the Appendix). Each subject saw 48 posemous phrases, 48 homonymous phrases,288 control phrases, 96 in each of the conconditions. The foils from Experiment 2 weaugmented by new items to make up thecomplement. Four counterbalancing lists wused, as in the previous experiment, so teach critical phrase would appear in all contions across subjects. The procedure was idtical to that used in Experiment 2.
Results and Discussion
Two subjects were dropped from the studymaking errors on more than 20% of all tphrases or of the experimental trials. For themaining 32 subjects, RTs larger than 3 SDaboveeach subject’s mean were omitted, as were tin which subjects made an error on the prphrase. This resulted in the exclusion of 2%the data.
One assumption was that homonyms and ysemous words would differ in the size of tconsistency effect, with consistency bestronger for homonyms, due to the completseparate representations. Surprisingly, this not found, as shown in Table 1: There wereinteractions between consistency and word tyeither in RTs,F1(1,31) 5 1.69,p . .20; F2(1,46), 1; or errors,F1(1,31) 5 2.56, p . .10;F2(1,46) 5 2.62,p . .10.
Importantly, although there were no interations, consistency was still a reliable and iportant factor: when the prime and target wconsistent in the sense they referred to,target phrase was evaluated 85 ms mquickly and 12% more accurately than whthe prime and target were inconsistent in senF1(1,31)5 46.01,p , .001;F2(1,46)5 35.76,p , .001 for RTs; andF (1,31) 5 77.79,p ,
1.001; F2(1,46) 5 43.81, p , .001 for errors.
N OF POLYSEMY 269
of
rsfule4dsly-andolellreati-
en-
orere-
ialse
of
ol-eglyas
nope,
--
reherense,
There was no difference overall between thomonyms and polysemous words, allF’s , 1.
When results were analyzed separately the polysemous and homonymous words, bshowed the pattern found in prior experimenwith an advantage for consistency in errors RTs (see Table 1 for the means), all p’s , .005.If anything, the improvement of the stimuli apears to have strengthened the results for psemous items.
This pattern of results is surprising, sincelarger effect was expected for the homonymAs discussed earlier, the different meanings homonym are generally thought to be repsented distinctly, so there should have blarger consistency effects in these stimuli. Pring one meaning that is totally separate fromalternate meaning should have more stronaided understanding of a phrase using the smeaning and hurt understanding of a phrusing the other meaning. Although the effsize for the homonyms was slightly larger (1ms) than for the polysemous words (64 mthere was no interaction by word type, so tdifference is at most suggestive.
Because this finding is somewhat surprisiand because there is a hint of the expectedference, we replicated the experiment. In adtion to the main experiment just reported,subjects performed in an essentially identistudy, with the exception that some of the homnym items were different (as described in Eperiment 4A). In this version of the experimenwhich we will call Experiment 3A, we agaifound no important difference between the hoonym and polysemy conditions, the interactip’s all ..25. In fact, in this experiment, thpriming effect found for homonyms (25 m12% errors) was actuallysmallerthan that foundfor polysemous items (80 ms, 12% errors). Tpriming effect was significant in separate anases of both classes of words, in RTs and eranalyses, with the exception of the item analyof RTs for homonyms,F(1,23) , 1. Thus,across Experiments 3 and 3A, the size ofpriming effect was very close to being equal fhomonyms and polysemous words. In one,effect was slightly larger for homonyms, and
the other, the effect was slightly larger for poly-
eig
ho. bout
nh
a
t
e
a
tr
e
nd
orrde
oureri-of
polysemy and homonymy than did Williams’s
re-a
dif-ine
ousntichia ad-forn-
sis-e
hes,atrs
themselves might be responsible for the con-
and 4
M 579 95 579 94
semous words, but in neither case did the intaction of word class and priming approach snificance.
This pattern of results is consistent with tfinding in Experiment 1 that memory was ncued by using the word in a different senseboth cases, cross-sense performance was athe same as a baseline. That is, counteriitively, the overlap of different senses of tsame word is minimal—about the same as overlap of homonym meanings. In functionterms, this means that any core meaning shby these senses is also minimal. Such resrule out not only a full single-sense view balso any separate representation accounwhich a core meaning plays a significant roWe discuss possible reasons for the minimoverlap of polysemous senses in the GenDiscussion.
This finding is in direct contrast to that oWilliams (1992), who argued for representtional differences between homonyms and posemous words. In particular, he found thcontextually inappropriate senses of polysmous words were still activated for some timafter presentation, whereas the literaturehomonyms generally shows suppression ofincorrect meaning (see introduction to Expement 5 below). However, Williams did not actually include homonyms in his study, so he dnot directly reveal any differences betwethem and polysemous words. Also, his expements did not compare priming of relevant a
irrelevant senses, as the present experimer--
etIn
did. It is possible that the priming he found fthe contextually irrelevant sense of a wowould be significantly less than what would bfound for the relevant sense (analogous toconsistency effect). Thus, the present expments provide a more complete comparison
270 KLEIN AND MURPHY
TABLE 1
Mean Reaction Times (ms) with Standard Deviation in Parentheses and Percentage Correct for Experiments 3
Sense
Consistent Inconsistent
RT % correct RT % correct
Experiment 3Polysemous words 774 (118) 95 838 (156) 86Homonyms 743 (130) 97 850 (187) 83M 759 96 844 85
Experiment 4Polysemous modifiers 593 (108) 93 583 (112) 91Homonym modifiers 565 (111) 96 574 (115) 97
tu-e
thealredultsut inle.alral
f-
ly-ate-eonhei--idnri-
study.
EXPERIMENT 4
The first three experiments looked at thelation between phrases that used words inconsistent or inconsistent sense, finding aference between these conditions bothmemory and timed semantic judgments. Whave attributed these effects to the polysemword being used in the same or a differesense across trials. However, the way in whsense consistency was manipulated was vmodifying word, so it is possible that the moifiers themselves were partly responsiblethese effects. To illustrate this possibility, cosider the consistent phrase pairs,wrappingpaper and shredded paper. It is possible thatthe word wrapping was priming shredded,rather than the consistency of the use ofpapercausing the effect. In this case, “sense contency” results would be obtained, but for thwrong reasons—having nothing to do with tnoun, which was the word of interest. ThuExperiment 4 was a control experiment thinvestigated whether priming of the modifie
ntssistency effects obtained.
O
th
aadnaeo
shn
enelcx
lirr
da-e
nu
reik
noo
r
eirre-
told
eenairortdi-the
asa
sesforbyctsory
theyets.wning
y-eri-it.
0%For This
thetedse-nly
ao-ct)ct),
ays re-in-be-were
REPRESENTATI
Experiment 4 also raised the possibility of eplaining the surprising findings of Experimen3 and 3A. One hypothesis was that homonyms in that experiment would shostronger effects of sense consistency thanpolysemous words, since their meanings hlittle or no semantic overlap. In fact, there wno reliable difference between the two contions. One possible reason for these findicould be uncontrolled item differences—in pticular, the modifier–modifier priming describabove. Perhaps the relations between the mfiers was systematically different in the polymous and homonymous stimuli, explaining wthe predicted larger effect in homonyms was found.
To address these possibilities, it was necsary to perform a control experiment to esure that sense priming was not being drivby the modifiers—to test for any possibmodifier–modifier priming and compare supriming across word types. To this end, Eperiment 4 was run on the modifiers alonusing a lexical decision task. (The sensicajudgment task used in the previous expements could not be used with single-wostimuli.)
Method
Materials. The prior experiments usephrases that included the polysemous word a modifier. For example,paper was a polysemous word, and it was paired with the modifiwrapping, shredded, liberal, and daily. In thisexperiment, only the modifiers from Experime3 (from both polysemous and homonymoitems) were used. The comparison of intewas modifiers taken from consistent pairs (lwrapping and shredded) vs modifiers takenfrom inconsistent pairs (like liberal and shred-ded) in the previous experiments. Foils cosisted of words followed by nonwords. Prnounceable nonwords were constructed frwords matched for frequency (Francis Kucera, 1982) and length to the actual woused, with minimal changes (1–2 letters) transform them into nonwords.
Procedure. Subjects viewed the materials oa Macintosh Quadra 630 computer, which wa
re-
N OF POLYSEMY 271
x-tse
wthevesi-gsr-ddi-
e-yot
s--n
eh-
e,tyi-d
nd
rs
tsste
--m
&dsto
n
connected to a PsyScope button box. Thdominant hands were assigned to “WORD”sponses and their nondominant hands“NONWORD” responses. Subjects were tothat letter strings would appear on the scrone at a time, in pairs. The first string in a pwould be a word, and it would be up for a shwhile (500 ms) and then disappear. Immeately, a string of letters would appear onscreen. Their task was to decide as quicklypossible, without making errors, if it wasword or not. Feedback for correct responwas on the screen for 1 s, while feedbackincorrect responses was up for 2 s, followeda gap of 250 ms before the next pair. Subjewere also warned that there would be a memtest at the end on what they saw, to ensurepaid close attention to both primes and targThe memory test instructed them to write doall the words they remembered seeing durthe experiment.
Subjects. Subjects were 37 introductory pschology students who participated in the expment for partial fulfillment of course credThey were all native speakers of English.
Results and Discussion
Five subjects who made errors on over 2of the strings were omitted from analysis. the remaining 32 subjects, RTs larger than 3SDabove each subject’s mean were omitted. resulted in the exclusion of 3% of the data.
Analyses performed on the RTs found thatmodifiers of the homonyms were evaluafaster (569 ms) than the modifiers of the polymous words (588 ms), but this was reliable oin the subject analysis,F1(1,31) 5 4.25, p ,.05; F2(1,46) 5 1.22,p . .25. There was alsoreliable difference in accuracy, with the homnym modifiers more accurate (96% correthan the polysemous modifiers (92% correF1(1,31) 5 10.45,p , .005; F2(1,46) 5 5.31,p , .05. One possible reason for this effect mbe that the senses of the polysemous wordquired more complex modifiers to be distguished. However, as simple differences tween homonyms and polysemous phrases not of interest in previous experiments, this
ssult is not very revealing.
leonbfolnle
eem
pmte
nm
te
a re
oil
o
aeto3rinnin
o
ae--
erol-d inthee-cya
nton-
sthe
ope
r-rd
ef-re
gerhentdso3Alsordd-
nees
a- ofnt, in-tep-ctedewenag-
272 KLEIN AND
More importantly, there was no reliabpriming of consistent items (we will refer tthe modifiers by the names of the conditiothey served in in Experiment 3), as wouldexpected if the modifiers were responsiblethe priming of the previous experiments, aF’s , 1. However, the interaction betweeword type and consistency, while not reliabin the RT analysis,F1(1,31) 5 1.40, p . .20;F2(1,46) 5 1.35, p . .25, was reliable in theerror analysis F1(1,31) 5 4.96, p , .05;F2(1,46)5 4.16,p , .05.
Because of this reliable interaction and bcause there was a hint of a difference betwthe results for homonyms and polysemous itein Experiment 3, the modifiers of homonymand polysemous items were examined serately (see Table 1). The polysemous iteshowed no differences between the consisand inconsistent modifiers in RTs,F’s , 1,while in the error analysis, there was a nosignificant trend toward the consistent itebeing more accurate,F1(1,31) 5 3.14,p 5 .09;F2(1,23) 5 4.02, p 5 .06. The homonymsshowed no differences between the consisand inconsistent modifiers in either RTs,F’s ,1, or errors,F1(1,31) 5 1.00,p . .30; F2(1,23), 1. The lack of significant effects and the smsize of the trends make it hard to attribute theliable consistency results in the previous expments to modifier priming.
A similar experiment was also carried out the modifiers used in Experiment 3A. (We wrefer to this as Experiment 4A.) In the compleanalysis, we found no consistency effect,p’s ..10, and no interaction of consistency and wtype,p’s . .25. However, visual examination othe results did seem to suggest that there wpossible difference between the word typSeparate analyses revealed no priming whaever for polysemous items. Therefore, the csistency effect found in Experiments 2 and cannot be attributed to modifier effects. Surpingly, a sort of reverse priming effect was fouin the homonym stimuli, in which consisteitems were verified 24 ms more slowly than consistent items,F1(1,27) 5 5.01, p , .05;F2(1,23) 5 4.66,p , .05, although there was n
effect in errors.MURPHY
serl
-ens
sa-snt
-s
nt
lle-ri-
nl
In order to eliminate modifier priming aspossible explanation for the results, we rmoved eight items that contributed to this priming difference, so that the amount of modifipriming was equated in the homonyms and pysemous items. (These phrases were not useExperiments 3 and 4.) We then reanalyzedresults of Experiment 3A with those items rmoved, to reevaluate the effect of consistenfor these equated modifiers. We again foundreliable effect of consistency, with consistephrases evaluated 60 ms faster than their incsistent counterparts,F1(1,23) 5 7.38,p , .05;F2(1,38)5 6.06,p , .05. The consistent itemwere also 12% more accurately judged thaninconsistent ones,F1(1,23)5 32.85,p , .001;F2(1,38)5 22.39,p , .001. There was again ninteraction between consistency and word ty(all F’s , 1).
In short, although there were some diffeences between the modifiers of the two wotypes, they could not explain the consistencyfects found earlier. Consistency effects wefound for stimuli that had no modifier primin(Experiment 3), and the effects did not diffwhen the modifier priming was equated for ttwo word types (the reanalysis of Experime3A). Note that the interaction of priming anword type was not reliable in any experiment,the reanalysis of the results of Experiment was a conservative step. Nonetheless, it afound strong consistency effects for both wotypes. Thus, the results show very similar fin
terdfs as.so-n-As-dt-
ings for homonyms and polysemous words.
EXPERIMENT 5
This experiment was conducted to examiwhether the priming of polysemous word sensis due to inhibitory or facilitory processes. Fcilitation could come about through a numbermeans. One obvious way is that if differeword senses have separate representationsterpreting a word in one sense might activathat sense for later trials. When the word apears again, the correct sense is already seleand ready to be used in interpreting the nphrase. Thus, continued activation of a givsense could speed processing. One could im
ine a rather different facilitative process that
O
bioim
fp
nohueihT
t.
wocn
sh
al
sc
hnliasn
-n
tow-n-tiony,ctgsherbyc-ndbybeto
n-chindes
m-
l-eed, two
ndrts,ldenses
fa-er.
or-
toed ofis-
ta-on-ral-
tral:
-tic
REPRESENTATI
would also lead to priming, however. It may simply that the consistency of the interpretatof the entire phrase is responsible for the pring. For example,wrapping paperand shreddedpaperrefer to similar kinds of things (i.e., havshared semantic properties), whereas wrappingpaperand daily paperrefer to different kinds othings. It could be the shared semantic comnents of the whole phrase that speed subjedecision that the consistent phrase makes se
Inhibition of incorrect senses is another psible cause of the priming results. Gernsbac(1990) has emphasized the importance of spressing incorrect meanings of words and stences as part of a fluent comprehensprocess. However, if senses inhibit one anotthey must have separate representations. newspaper sense of papercannot be suppressewhile the sheets of material sense is sparedless the two senses are functionally distincjudging wrapping paperactually made it moredifficult to interpret daily paper (relative to aneutral condition), this would suggest that tdifferent senses are stored. The second accgiven above for facilitation would not prediinhibition of senses. That is, if consistephrases are easier to judge because they semantic components, it does not follow tphrases using different word senses wouldmutually inhibitory, merely that there would bless facilitation, because they do not sharemany semantic properties.
The prior experiments do not distinguish fcilitation or inhibition of senses. Because ontwo conditions were tested, it is not possiblesay whether consistent phrases received anvantage, inconsistent phrases suffered a divantage, or some combination of the two. Sua conclusion requires a neutral condition tthe others can be compared to: Faster respothan the neutral condition would indicate facitation, and slower responses would indicinhibition. In studies of homonyms, Gernbacher and Faust (1991; also see discussioGernsbacher, 1990) showed that there was bfacilitation and inhibition involved in the comprehension of ambiguous words in sentecontexts. When reading words likebank, sub-
jects showed increased activation to words rN OF POLYSEMY 273
en-
e
o-cts’se.
s-erp-n-
oner,he
dun- If
ounttt
hareatbeeas
-ytoad-ad-h
atses-te-in
oth
ce
lated to the intended meaning as well aswords related to the unintended meaning. Hoever, after a delay, the activation to the unitended meaning decreased, whereas activato intended meanings did not. Significantlwhen the context did not pick out the corremeaning of an ambiguous word, both meaninremained active after the delay. Gernsbacand Faust explained this pattern of resultsarguing that both meanings initially receive ativation during the comprehension process, athe incorrect meaning is then suppressedcontextual processes. Such a pattern mightexpected when a single word is associatedtwo very different semantic representations. Ihibition is possible and indeed desirable in sua case, because one can fish from only one kof bank and withdraw money from only thother kind: identifying one of the meanings acorrect indicates that the other meaning is copletely incorrect and so should be inhibited.
This situation is not exactly the same for poysemy, since the two senses are related. Indin some cases, the same word can be used insenses at once, as in Your book is not only badlywritten, it is too heavy; or The factory fired itsworkers and then was burned to the grou(Fauconnier, 1985; Cruse, 1986; Geerae1993). Thus, it is not clear that inhibition woube as desirable as it is for homonymy—or evpossible. To the degree that polysemous senshare some meaning, it should be harder tocilitate one sense while suppressing the othThus, finding results parallel to those found fhomonyms would indicate fairly distinct representations.
The present experiment therefore aimeddiscover whether the priming effects observfor polysemous senses were due to inhibitionthe inconsistent sense, facilitation of the constent sense, or both. Since inhibition and facilition are both measured relative to a neutral cdition, the selection of an appropriate neutitem is obviously crucial. It would not be appropriate to use an unrepeated condition as neuOnce one has read the word paper, it would beeasier to read it again in the next trial completely independent of any issues of seman
e-representation, so an unrepeated condition could
D
te r
hx
elaioineuner
tr
as6)haraesk-
mbpr
tosntwthr
ysm
-esas
thee in
edsere-n-tar-the
me to),
lete(seeen-, 60
ndd toereal one
y-ri-
t.
ac-the
ereei-lted
nn-ntns
e9%emson-
274 KLEIN AN
not meaningfully be compared to the consisand inconsistent trials in which the word waspeated.
Our neutral baseline used a prime in whicblank line took the place of the modifier. For eample, ______papermight be followed bylib-eral paper. This neutral condition primes thword paper but does not select any particusense. This was to be compared to the conditin which the prime selected the consistent orconsistent sense of the noun, as in prior expments. Because of the form of the target stimwhich consisted of multiple words, a blank liwas deemed more appropriate than simply psenting the single word. Other possible neumodifiers would have been the use of a stringletters (XXXX) or an unrelated word (suchBLANK). Den Heyer, Taylor, and Abate (198found that unrelated words are more neutral tthe strings of X’s. However, in the current padigm, it was impossible to find completely unrlated words that would not influence the taFor example,BLANK paperis inappropriate because it suggests a particular sense ofpaper, soit would hardly be neutral. This sort of probleseems likely for any neutral word that wouldused. For this reason, we used a blank lineceding the word in the neutral primes.
Method
Design. As in Experiments 2 and 3, subjecwere asked to read phrases containing the psemous word and then make judgments atheir sensicality. Phrases occurred in sequepairs. The prime could select one of the tsenses of the polysemous word or be inneutral, blank line condition. The targets wethe same as before. Therefore, each polmous word had associated with it three priphrases (e.g.,wrapping paper, daily paper, and______paper) and two target phrases (shred-ded paperandliberal paper), so six counterbalanced lists were necessary. One-third of theperimental items in each list were senconsistent, one-third sense inconsistent,one-third neutral. Subjects saw each polymous word in only one pair.
As before, we constructed foils in which thfirst, second, or both phrases did not ma
MURPHY
nte-
a-
rns-
ri-li,
e-alof
n--.
ee-
sly-toialoeee-e
x-ende-
e
sense. Then we replaced one-third of each offoil types with pairs of phrases in which thprime was neutral, to match the frequenciesthe experimental stimuli. Although this resultin some foil primes that now did make sen(when the initial nonsensical phrase was placed by the neutral line modifier), it maitained the pattern that the sensicality of the get phrase was statistically unrelated to condition of the initial phrase.
Materials. The polysemous materials froExperiment 3 were used here as well, but duthe addition of a new condition (neutral prime6 more items were created so that the compset consisted of 30 polysemous word quads the Appendix). Each subject saw 60 experimtal polysemous phrases and 180 foil phrasesof each of the types of foils.
Procedure. Using the same equipment aprocedure as previously, subjects were askejudge the sensicality of each phrase. They wexplicitly instructed that all the blank, neutrphrases made sense (since they were justword).
Subjects. Subjects were 60 introductory pschology students who participated in the expement for partial fulfillment of course crediThey were all native speakers of English.
Results and Discussion
Performance in this experiment was very curate, and no subjects were dropped from study. However, RTs greater than 3 SD aboveeach subject’s mean were omitted. Trials womitted in which subjects made an error on ther the prime or the target phrase. This resuin the exclusion of less than 1% of the data.
An ANOVA found a clear difference betweethe consistency conditions, with the neutral codition (879 ms) falling between the consiste(832 ms) and inconsistent (938 ms) conditioin RTs,F1(2,118) 5 6.33,p , .005; F2(2,58) 511.16,p , .001. Error means followed the sampattern as the RTs, with the neutral items (8correct) less accurate than the consistent it(96% correct) and more accurate than the incsistent items (84% correct),F1(2,118) 5 19.21,p , .001; F2(2,58) 5 14.26,p , .001. More de-
ketailed analyses found evidence for both facilita-
ne
i
os
s
tewIte
os
av e
tto
nii
hh
i
i
n
n-t.lean-nsw-ly-ir
ghtarei-
se.
dt.re
p- ofl-
is-eeve inetct,edd-ost
pa-
uldg
nr,
ri-ire-besly
dynalar-ueity
REPRESENTATIO
tion of consistent items and inhibition of incosistent items. The consistent condition was rably better than the neutral condition in boRTs (a 47-ms difference),t(59) 5 2.26,p , .05;t(29) 5 2.30,p , .05; and errors,t(59) 5 4.39,p , .001; t(29) 5 3.26,p , .01. The compari-son of the inconsistent and neutral conditiowas fully reliable over errors,t(59) 5 2.12,p ,.05; t(29) 5 2.66, p , .05, and the 59-ms RTdifference was reliable in the item analyst(59) 5 1.77,p 5 .08; t(29) 5 2.32,p , .05.This suggests that for polysemous words, csistency facilitates comprehension and incontency inhibits comprehension.
These data indicate that multiple procescombine to give sense-consistent phrases anvantage. First, it appears that the consissense is being activated, which effectively loers the threshold for subsequent activation. not clear from these data how long this efflasts, but it is present at least long enoughhave an effect on the next phrase read. Secthe activation of one sense appears to caudampening of activation for different senses.
The facilitation of consistent uses of a wordeasily explained as repetition priming of semtic features or an entire word sense. Howethis mechanism cannot explain the inhibitionthe inconsistent sense. If inconsistency wworse than consistency only because the content sense was primed, then the inconsiscondition would be about as fast as the neucondition, since neither one would have the crect sense primed (but both would have had geral priming from reading the word, etc.). Istead, we found that the inconsistent conditwas reliably harder than the neutral one, whis analogous to Gernsbacher and Faust’s (19finding that the inappropriate meaning of thomonym showed less priming relative to tneutral condition after a delay. This findincould be caused by an active process of inhtion (as Gernsbacher and Faust propose),though other explanations may also be poss(see the General Discussion).
These results give even stronger evideabout the separate representation of polysemsenses. If one sense can be activated and
other inhibited as a result of an earlier encounN OF POLYSEMY 275
-li-th
ns
s,
n-is-
es ad-nt-
isct tond,e a
isn-er,ofresis-entralr-
en--onch91)ee
gbi-al-
ble
ceous the
with the word, these senses must be functioally and therefore representationally distincClearly, the notion that the word has a singcore meaning that is the basis for every use cnot accommodate such results. The implicatioof these results go beyond that conclusion, hoever, in that they suggest that the senses of posemous words are quite distinct in spite of theapparent relatedness. For example, one mihave thought that the fact that newspapers printed on paper would make it difficult to actvate the newspaper sense of paperwhile simul-taneously inhibiting the sheets-of-paper senHowever, this was not the case.
We argued earlier that such inhibition ruleout one explanation of the earlier priming effecPerhaps a liberal paper is just generally mosimilar to a daily paper than it is to some wraping paper. This kind of semantic consistencythe whole phrases—without any mention of poysemy—might explain the advantage of constent over inconsistent word uses in the first threxperiments, as similar concepts would hasome of the same properties, which might aidmaking a sensicality judgment to the targphrase. (However, we should note that, in fawrapping paper is not that similar to shreddpaper: One cannot wrap something with shreded paper, and wrapping paper is not the mtypical shredded paper. Similarly,daily paperemphasizes a very different aspect of newspers than does liberal paper. We selected modi-fiers just so that the same-sense phrases wonot describe very similar things, avoidinphrase pairs like weekly paperand daily paper.)This explanation would not predict suppressioof an inconsistent use of the word, howevewhich is what we found in the present expement. Such suppression would seem to requthat a word like paperhave different representations for different senses, one of which could suppressed while the other was simultaneoufacilitated.
We should emphasize that the present stuwas designed to investigate the representatioissues involving polysemous senses, and in pticular it did not address the processing issthat has been of so much interest in ambigu
terresearch more generally, namely, the question of
xn
dfg
ni
h
c
t
hatstedonheificro-
n in-atepro-anyrateouldthe
icply
276 KLEIN AND
whether the initial activation process is contesensitive from the start or only after initial cotext-free activation of all meanings (i.ewhether the selection process is or is not molar). We mention this because such studies ouse a priming technique similar to ours (thouusually with a lexical task rather than a semaone). Given the difficulty in answering thquestion in the homonymy literature (see Simson, 1994, for a review), it seems likely that tanswer will be no easier to find for polysemOur results suggest that selective access is pos-sible, since different senses seem to be repsented separately. If the senses shared a representation that included most of easense’s meaning, selective access could noachieved. Whether selective access in fact
une
a
lo
e
t
n
reonicaln-y,
utde-tedhenthex-ter-s,ear-s-iswe,ificsup-hyforit
yse-redofheyterntsle
curs is a question we leave to future research
GENERAL DISCUSSION
Taken as a whole, these experiments provevidence for separate representations of senses of polysemous words. The studies foevidence for sense priming in polysemy. Usia word in a specific sense facilitated comprehsion for a phrase that used the word in the sasense and inhibited comprehension for a phrthat used the word in a different sense. In adtion, we found that the effects of sense constency in polysemy are similar to the effects meaning consistency in homonyms. Finalusing a word in the same sense was a gmemory cue, but using the word in a differesense was not.
A straightforward core meaning view is eaily disconfirmed by our data, because it wounot predict any difference between using a woin the same or a different sense. And this viwould certainly predict that there should bequalitative difference between homonym(which have no core meaning) and polysemowords (which do), yet we found no such diffeence. In short, words like papercannot be repre-sented by a single semantic description thaaccessed every time (“flat sheets made of wopulp” or the like).
One can elaborate a core meaning view inumber of ways that might make it more cons
tent with the present results. Caramazza aMURPHY
t--
.,u-
tenh
ticsp-ey.
re-ore
ch beoc-.
idethendgn-mese
di-is-ofy,od
nt
s-ldrdw
asusr-
isod
ais-
Grober (1976) argued for a core meaning trepresents polysemous words but also suggethat the lexicon contains a set of “instructirules,” analogous to derivation rules in tgrammar, that could produce more specsenses from the core. If one interprets this pposal to mean that every distinct sense has astruction rule that is stored in order to generthat sense from the core meaning, then this posal is actually not a core meaning account more, since the instructions would be separepresentations of each sense. Such rules whave to be quite specific in order to derive senses they found for line, such as “rope” or“line of business,” from the core geometrmeaning, so they could not be rules that apwidely across the lexicon.
An alternative more consistent with the comeaning view might instead make a distinctibetween the permanent representation in lexmemory and more temporary, episodic costructions of word senses (Anderson & Orton1975). That is, perhaps a word likepaperhas asingle (core) representation in the lexicon, bwhen it appears in a specific context, a moretailed context-appropriate sense is construcin the sentence representation. One could targue that this sense is what is influencinginterpretation of later uses of the word. For eample, once one has used the context to inpretpaperto mean a daily publication of newone can then interpret the subsequent appance ofpaperin the same way much more eaily. There are four major problems with thproposal. First, it is not clear how this viecould explain the inhibition of a different sensas was found in Experiment 5, since specsenses must be represented in order to bepressed. Second, this view cannot explain wthe consistency effect was at least as strongpolysemous and homonymous words. Sinceassumes that the specific senses of a polmous word are created on the fly from a shacore meaning, but different meaningshomonyms are stored separately (since tcannot be derived), it should predict a greaconsistency effect for homonyms. Experime3 and 3A did not find this result. Third, the so
ndrepresentation of a core meaning seems psycho-
O
eif
der
thblnal
wn
ntttdthct
ais,so
, oss
tass
sidsho
isna5and
iss
saso
enutit-sesnehee
-of
en-la-ds.r-n-
ctsntry
es,se-
reectsto-t,r
lit-t
tlyted benti-s,
isrdsrgu-ped,;ec-
REPRESENTATI
logically implausible if specific senses are frquent. That is, it would be very surprisingpeople often used the wordpaper to refer tonewspapers and to published articles, yetnot represent this fact in the lexicon, but drived it from general principles on every occurence. Finally, linguists have argued thatdetails of word usage are not in fact derivafrom a core meaning and that conventiosenses must therefore be represented in theicon (Lehrer, 1990; Rice, 1992). In sum, vieing senses as episodic constructions doessave the core-meaning theory.
As mentioned in the introduction, there aredetailed psychological models that attemptexplain how polysemous words are represenand processed. The present results speaknumber of the possible models one couldvelop. Perhaps the main empirical result isfinding that different senses have little funtional overlap—about the same as the unrelameanings of homonyms. (We attempt to explthis finding in the next section.) Obviously, thcreates problems for a core view. Howeveralso greatly limits all the possible core-plusenses views. If one believes that there is a cmeaning that is common to all uses of a wordat least a prototypical meaning that is accesfirst, then processing the word in one senshould provide some benefit to processing ianother sense. However, in both a memory t(Experiment 1) and a semantic judgment ta(Experiments 3 and 3A), we found that crossense priming was minimal. In the memory tathe different-sense use of a word did not prova reliable memory cue. In the sensicality tathe effect of switching senses was virtually tsame as the effect of switching meaningshomonyms. These results argue that if therecommon core to words, it has little content alittle effect on processing. The finding thcross-sense priming is inhibitory (Experimentalso provides evidence that sharing a core meing is not sufficient to benefit comprehensioAs a matter of parsimony, the present resultsnot give us any reason to believe that therecore meaning, so separate representationeach sense appear to be a better bet than
core-plus-senses view.N OF POLYSEMY 277
-
id--eel
ex--ot
ooedo ae-e-edin
it-rer
ede
inskk-
k,e
k,efa
dt)n-.oaof
the
Reasons for Minimal Sense Overlap
As we noted in the Introduction, polysemousenses are (by definition) related, wherehomonyms are not. Why, then, did we find ndifference in the consistency effect betwepolysemous words and homonyms? Or, pmore generally, why do our results show so ltle overlap in polysemous senses if the senare related, often by productive relations? Omight wonder whether there is some error in tmethods that have given this counterintuitivresult.
In her thesis, Klein (2000) explored the relations between polysemous senses in a varietyconceptual tasks (rather than online comprehsion tasks) in order to better understand the retions among the senses of polysemous worHer results confirmed the low overlap of diffeent senses in a number of different tests of coceptual coherence. For example, subjestrongly preferred to sort together differewords from the same superordinate categoover the same word used in two different sensand the induction of a property from one senof a word to a different sense was very low. Interestingly, the scores for homonyms weslightly, but significantly, lower in the samtasks. For example, she found that subjesorted different senses of a polysemous wordgether only 14% of the time in one experimenbut this was reliably higher than the 6.6% fothe different meanings of homonyms.
These results show that naive subjects seetle conceptual commonality in the differensenses of a word, although it is very slighmore than that seen in the completely unrelahomonym meanings. How can such resultsreconciled with the fact that senses are semacally and historically related? To explain thiwe need to distinguish semantic relatednessfrom semantic overlap, which we have beenusing somewhat interchangeably prior to thpoint. The different senses of polysemous woare clearly related. In some cases, a strong ament can be made for how one sense develohistorically from another one (Clark & Clark1979; Malt et al., 1999; Murphy, 1997Sweetser, 1990). Although such cases must n
essarily be “related,” a relation does not make
D
tehpt
o
ru rlufa-na
ui
.eodtns.lsihinsroo
rovrelt n on
tec
m
s isare
farwn.is-
th aofs. thens al
en-e
calilarntals asu-
esarr,
beer, assess,neone notntdelesen a
ionf-ately.nt init
ep-yet
278 KLEIN AN
senses similar. For example, if one uses word paper to refer to sheets of writing matrial, one might find it very natural to extend tword to refer to the content written on the pa(His paper was boring.). However, the contenis not similar to the sheets: One is informatithe other is made of wood pulp; one has semtic content, the other has molecular structuone has an author, the other has a manufactand so on. As a result, wrapping paper and aeral paper do not have much semantic oveeven if the relation between the two is easily derstood. When the word is extended even ther (e.g.,paper meaning the publisher of newspaper, as in The paper fired half its reporters), the overlap is even less. If priming amemory cueing in these tasks (and sorting induction in Klein, 2000) primarily result fromshared semantic features, then even obviorelated senses may not overlap in meanenough to produce performance advantages
If this argument is correct, then it convergon our previous conclusions about what a cmeaning might be, for those views that inclusome kind of core representation. Clearly,core cannot be a basic component of meathat is shared by all or most of the sensepapercan refer to a kind of a material but athe content printed on that material—and,deed, the content minus the material as wone presents a paper verbally, or perhaps uscomputer projection system—then the senare related without having overlapping repsentations. So, if there is a core sense, it wprobably be a prototypical or perhaps histcally early sense that is not sharedby the othersenses but instead is the original basis fwhich the other senses were historically deri(although it seems likely that some senses aturn derived from other derived senses; Maal., 1999; Murphy, 1997). Such a core would play a critical part in the online processingthe word, although it might be important in liguistic analysis.
To summarize, our proposal is that linguisanalyses arguing that polysemic senses arlated (e.g., Nunberg, 1979) are correNonetheless, the different senses of a word
not be very similar, and it is this variable thaMURPHY
he-eer
n,an-e;rer;lib-ap,n-ur-
dnd
slyng
sree
heing Ifon-eng aes
e-uldri-
med inetotf
-
ic re-t.ay
determines priming in the present tasks. If thicorrect, it is not surprising that the senses stored separately in the mental lexicon.
Limitations on the Present Results
One question about these results is howdiscriminations between senses can be draWe chose senses of words that were fairly dtinct, so that we could select each sense wisingle-word modifier. However, some forms polysemy involve very subtle distinctionFor example, Nunberg (1979) discusses type/token polysemy found in almost all nouin which a word can refer to an individual orclass. The word dog can refer to an individua(The dog dug up my flowers) or to the entireclass (The dog has been domesticated for cturies). Since the individual is a member of thclass, generally having all the properties typiof the class, these two senses have very simcontent. Although there may be an importaontological difference between an individuand the class it is in, when describing dogs aclass (domesticated four-legged mammals, ually with fur, etc.), the description also applito a typical individual in that class (a particuldomesticated four-legged mammal, with fuetc.). Given this overlap in meaning, it might quite difficult, if not impossible, to facilitate onof these senses while suppressing the othewe found in Experiment 5. In contrast, the uof dogas a verb, meaning to follow and haraseems quite different from the individual canisense, so it should be possible to facilitate while suppressing the other. This issue doesarise for homonyms, because their differemeanings are unrelated. But a complete moof polysemy will have to explain which sensare enhanced and which are suppressed whgiven sense is selected.
This issue leads to another critical questabout polysemy, namely, how to tell if two diferent senses are indeed represented separIn the limit, “the meaning of any word form is isome sense different in every distinct contexwhich it occurs” (Cruse, 1986, p. 51). Clearly,is impossible (and probably undesirable) to rresent each such distinction in the lexicon;
tthe present data suggest that at least some mean-
O
seee
g
no
acarhe
ffot
4onr
rusnhee
wuoip9s
verybeeven sohus,eon-thene
aller issetenentseshativenofre-reehein
ome-lishon
thatlso
otedtooddi-is
toets.ntedan
isesat a
The development of a model of sense creat-
REPRESENTATI
ingful distinctions are indeed represented. Thit is likely that the mental lexicon represensome reasonable number of senses, ratherthe extremes of one core sense or every possemantic distinction. If two senses are only vsubtly different, it seems unlikely that speakwill develop separate entries for them, sincsingle entry will suffice to specify most of thmeaning for both. If two senses are strikindifferent, then a single entry will probably bunsuccessful at representing both meaniwhich will presumably lead to the formation separate entries. Although this “in-betweeproposal seems eminently reasonable, it is rather vague. What is needed is a more spemodel of what causes a sense to be separrepresented, from which one could derive pdictions about which uses would involve tsame senses and which would involve differsenses. Such a proposal is not yet to be founthe literature (though see Murphy, 1997,some discussion) and is beyond the scope opresent article. However, one empirical demstration that diverse uses could be reduced manageable number of senses can be founGibbs, Beitel, Harrington, and Sanders (199who examined 35 different senses of the wstand. Using hierarchical clustering, they fouthat three main categories could adequately resent the similarity among the 35 uses (i.e.,count for substantial variance of the similaties). Thus, one might argue that people wobe likely to represent three general sensestand in order to make the distinctions amodifferent uses. It would be interesting to test tgrouping with the priming technique used hor the conceptual methods used by Kl(2000).
Some linguistic analyses of polysemy haproposed that there is no firm line to be drabetween polysemy and more usual ambig(see Geeraerts, 1993, and Tuggy, 1993, for such discussion). Furthermore, some linguhave argued against the more basic assumof separable senses. For example, Cruse (1pp. 71 ff.) argued that some words had rangesenses that shaded imperceptibly from oneanother, forming a sense spectrum. The difficult
aspect of such spectra, he claimed, was tN OF POLYSEMY 279
us,tsthanibleryrs aelyegs,f
n”lsoifictelye-ent
d inor then-o ad in),rddep-ac-i-ld ofgatrein
ven
ityne
ststion86, of
to
sometimes the endpoints of a spectrum had different meanings and could not plausibly counted as being part of the same sense,though intermediate “adjacent” senses wereclose as to not require separate entries. Tone is faced with arbitrarily dividing up thspectrum into distinct senses, in spite of the ctinuum of meanings, or conversely treating very diverse meanings as all falling under osense.
This awkward situation is a methodologicpuzzle, but it may not be so much of a puzzfor speakers to represent. For example, colophysically instantiated through a continuousof dimensions, with no natural break betwethe colors we perceive. In order to represcolor differences, the human visual system ua few landmarks—three photoreceptors tdiffer in the frequencies they are most sensitto—and any given color is identified in relatioto those landmarks. The particular locationthose landmarks (i.e., the frequencies theceptors are most sensitive to) is to some degarbitrary, so long as they can triangulate tspace that needs to be covered. Similarly,order to represent a word likepaper that has awide range of possible senses (ranging frthings made out of wood pulp to the managment of a newspaper), one needs to estabenough specific senses so that most commmeanings are covered—the specific sensesare represented may not be critical (see aNunberg, 1979). If one first learnspaper torefer to unlined sheets for writing, one may nneed to develop a new sense referring to linsheets, since such uses are easily undersby the pre-existing sense, and contextual adtion of the lines is simple. But the reversealso true: If one first learnedpaper to refer tolined sheets, one would probably not needform a new sense to refer to the unlined sheThus, the specific senses that are represemay be somewhat arbitrary, dependent onindividual speaker’s learning history. Whatprincipled is the process by which new sensare formed, based on the range of senses thword expresses.
hating would require more detailed description of
wo
s
3fme
otly
r
o
e
c-. Ith afor we
an-ngnhi-renticof ate, toicalodel to
that paper, be it a talk, a journal article, or a
and
280 KLEIN AND
the principles involved in establishing nesenses. It is possible that the connectionist ccept of local attractors may be helpful in deveoping such principles. That is, so long asgiven use is sufficiently close enough to a loc(semantic) attractor, there will be no needform a new sense. But if a given sense is notmantically close enough to an establishedtractor, a new set of distinctions will have to blearned with experience. Indeed, in his articprimarily about homonymy, Kawamoto (199pp. 511–512) makes just such a proposalpolysemy. On his account, a sense will becoseparately represented when it is frequenough or different enough from existing sensto develop its own local minimum in the semantic space. If there are a number of abequally frequent senses that are only slighdifferent from one another, this will not resuin separate representations but rather a fawide “basin” in semantic space. All such uswould be quickly understood, as listenewould not be attempting to distinguish differesenses.
Kawamoto’s (1993) proposal does not haexplicit inhibition of meanings, so how couldexplain the interference found in our Expement 5? The negative effects of using a worda different sense could be attributed to the dculty of having to climb out of the old (incorect) local minimum and into the correct onThat is, if one reads the word paperused to referto a newspaper, connection weights wouldmodified so that subsequent encounters ofword would be more likely to go to that locatioin semantic space. When the word is read agwith a different sense, contextual informatiwill have to overcome this tendency, thereby sulting in longer comprehension times (i.longer times to reach the correct local mimum). Such an explanation implies that tsenses used in our studies were in fact resented as separate local minima rather thanwide basin, so it is also consistent with our coclusion that separate senses are representthe lexicon. However, one advantage of tview, which Kawamoto points out, is that it donot require us to specify in advance whi
senses are separate, as the system discoversMURPHY
n-l-aaltoe-
at-ele,ore
ntes-ut
itself by virtue of the semantic similarity struture of the entire set of uses of the wordwould be particularly interesting to see if sucmodel could account for the similar results polysemous senses and homonym meaningfound.
In conclusion, polysemous senses are semtically related but are not very similar, resultiin same-sense facilitation and cross-sense ibition. If polysemous words do have a comeaning, it cannot be a substantial semacomponent that is common to all the senses word. It remains for a future paper to illuminathe exact form of the sense representationsexplain how new senses are added to a lexrepresentation, and to construct a process mof understanding polysemy. We look forward
ltirlyesrsnt
veitri- in
iffi--e.
bethenainn
re-e.,ni-hepre-in an-d in
hisesch
news story.
APPENDIX: STIMULI
Polysemous Homonym Polysemousitems in all items in Experi- items added inexperiments ments 3 and 4 Experiment 5
book arms glassestin ball classrun bat boxsee bridge signpaper calf trunkchicken date drivetelevision pokersheet fancorn pool*oak post*shower bankdrinker penatmosphere ringcoat bandcold mintfortune moldhall pitcherletter plotnail pitnavy matchorange sealproduction secondcotton tickfilling toast
bail**net**
Single asterisks indicate items used in Experiments 3
this4 only. Double asterisks indicate items used in Experiments3A and 4A only.
O
le
3)c
th.n
s a
l
Jo
e
-
a
ee-
n
f
s
ups
ce
tio
e
Ed
n
f
-
.-
a-
f
-
lu-
t
n:
-
-
m-
-
REPRESENTATI
REFERENCES
Anderson, R. C., & Ortony, A. (1975). On putting appinto bottles—A problem of polysemy. Cognitive Psy-chology, 7, 167–180.
Bainbridge, J. V., Lewandowsky, S., & Kirsner, K. (199Context effects in repetition priming are sense effeMemory & Cognition, 21, 619–626.
Caramazza, A., & Grober, E. (1976). Polysemy andstructure of the subjective lexicon. In C. Rameh (EdGeorgetown University roundtable on languages alinguistics. Semantics: Theory and application(pp.181–206). Washington, DC: Georgetown Univ. Pres
Clark, E. V., & Clark, H. H. (1979). When nouns surfaceverbs. Language, 55, 767–811.
Clark, H. H., & Gerrig, R. J. (1983). Understanding owords with new meanings. Journal of Verbal Learningand Verbal Behavior, 22, 591–608.
Cohen J. D., MacWhinney, B., Flatt, M., & Provost,(1993). PsyScope: A new graphic interactive envirment for designing psychology experiments. Behav-ioral Research Methods, Instruments, and Comput,25, 257–271.
Cruse, D. A. (1986). Lexical semantics. Cambridge: Cambridge Univ. Press.
den Heyer, K., Taylor, M. A., & Abate, A. (1986). On the nture of neutral primes in a lexical decision task. Psy-chological Research, 48, 161–168.
Deane, P. D. (1988). Polysemy and cognition. Lingua, 75,325–361.
Durkin, K., & Manning, J. (1989). Polysemy and the subjtive lexicon: Semantic relatedness and the saliencintraword senses. Journal of Psycholinguistic Research, 18, 577–612.
Fauconnier, G. (1985). Mental spaces: Aspects of meaniconstruction in natural language. Cambridge, MA:MIT Press.
Francis, W. N., & Kucera, H. (1982). Frequency analysis oEnglish usage: Lexicon and grammar. Boston:Houghton Mifflin.
Geeraerts, D. (1993). Vagueness’s puzzles, polysemy’sgaries. Cognitive Linguistics, 4, 223–272.
Gernsbacher, M. A. (1990). Language comprehension astructure building. Hillsdale, NJ: Erlbaum.
Gernsbacher, M. A., & Faust, M. (1991). The role of spression in sentence comprehension. In G. B. Simp(Ed.),Understanding word and sentence(pp. 97–128).Amsterdam: North-Holland.
Gerrig, R. J. (1986). Process and products of lexical acLanguage and Cognitive Processes, 1, 187–195.
Gerrig, R. J. (1989). The time course of sense creaMemory & Cognition, 15, 55–71.
Gerrig, R. J., & Murphy, G. L. (1992). Contextual influencon the comprehension of complex concepts. Languageand Cognitive Processes, 7, 205–230.
Gibbs, R. W., Beitel, D. A., Harrington, M., & Sanders, P.(1994). Taking a stand on the meanings of “stanBodily experience as a motivation for polysemy. Jour-
nal of Semantics, 11, 231–251.N OF POLYSEMY 281
s
.ts.
e),d
.s
d
.n-
rs
-
c- of
g
va-
-on
ss.
n.
s
.”:
Heine, B. (1992). Grammaticalization chains. Studies inLanguage, 16, 335–368.
Hino, Y., & Lupker, S. J. (1996). Effects of polysemy inlexical decision and naming: An alternative to lexi-cal access accounts.Journal of Experimental Psy-chology: Human Perception and Performance, 22,1331–1356.
Kawamoto, A. (1993). Nonlinear dynamics in the resolutioof lexical ambiguity: A parallel distributed processingaccount. Journal of Memory and Language, 32,474–516.
Klein, D. E. (2000). “Paper has been my ruin”: On the or-ganization of polysemous senses. Unpublished Ph.D.dissertation, Psychology Department, University oIllinois at Urbana-Champaign.
Lakoff, G. (1987). Women, fire, and dangerous things.Chicago: Univ. of Chicago Press.
Langacker, R. W. (1987). Foundations of cognitive grammar(Vols. 1 & 2). Stanford: Stanford Univ. Press.
Lehrer, A. (1990). Polysemy, conventionality, and the structure of the lexicon.Cognitive Linguistics, 1, 207–246.
Levelt, W. J. M. (1989). Speaking. Cambridge, MA: MITPress.
Light, L. L., & Carter-Sobell, L. (1970). Effects of changedsemantic context on recognition behavior. Journal ofVerbal Learning and Verbal Behavior, 9, 1–11.
Malt, B. C., Sloman, S. A., Gennari, S., Shi, M., & Wang, Y(1999). Knowing versus naming: Similarity and the linguistic categorization of artifacts. Journal of Memoryand Language, 40, 230–262.
Murphy, G. L. (1991). Meanings and concepts. In P. Schwnenflugel (Ed.),The psychology of word meanings(pp.11–35). Hillsdale, NJ: Erlbaum.
Murphy, G. L. (1997). Polysemy and the creation onovel word meanings. In T. B. Ward, S. Smith, & J.Vaid (Eds.), Creative thought: An investigation ofconceptual structures and processes(pp. 235–265).Washington, DC: American Psychological Association.
Nunberg, G. (1979). The non-uniqueness of semantic sotions: Polysemy. Linguistics and Philosophy, 3,143–184.
Perfetti, C. A., & Goodman, D. (1970). Semantic constrainon the decoding of ambiguous words. Journal of Ex-perimental Psychology, 86, 420–427.
Rice, S. A. (1992). Polysemy and lexical representatioThe case of three English prepositions. InProceed-ings of the Fourteenth Annual Conference of the Cognitive Science Society(pp. 89–94). Hillsdale, NJ:Erlbaum.
Ruhl, C. (1989). On monosemy: A study in linguistic semantics. Albany: State Univ. of New York Press.
Simpson, G. B. (1994). Context and the processing of abiguous words. In M. A. Gernsbacher (Ed.),Handbookof psycholinguistics(pp. 359–374). San Diego: Acade-mic Press.
Sweetser, E. E. (1990). From etymology to pragmatics:Metaphorical and cultural aspects of semantic struc
ture. Cambridge: Cambridge Univ. Press.
282 KLEIN AND MURPHY
Swinney, D. A., & Hakes, D. T. (1976). Effects of prior con-text upon lexical access during sentence comprehen-
r
e
o
s
Van Petten, C., & Kutas, M. (1987). Ambiguous words incontext: An event-related potential analysis of the time
in
sion. Journal of Verbal Learning and Verbal Behavio,15, 681–689.
Tabossi, P. (1988). Accessing lexical ambiguity in differtypes of sentential context. Journal of Memory andLanguage, 27, 324–340.
Tabossi, P., & Zardon, F. (1993). Processing ambiguwords in context. Journal of Memory and Language,32, 359–372.
Tuggy, D. (1993). Ambiguity, polysemy, and vagueneCognitive Linguistics, 4, 273–290.
nt
us
s.
course of meaning activation. Journal of Memory andLanguage, 26, 188–208.
Williams, J. N. (1992). Processing polysemous words context: Evidence for interrelated meanings. Journal ofPsycholinguistic Research, 21, 193–218.
Zgusta, L. (1971). Manual of lexicography. The Hague:Mouton.
(Received July 28, 1999)(Revision received November 25, 2000)Published Online June 21, 2001