Acquisition and Learnability in Phonology B. HayesLSA Linguistic Institute Summer 2003

Class 2.2, 7/9/03: Modeling Child-Generated Phonologies

1. Topics For Today

Relevance of OT: Pater Are child constraints the same as adult? Can the Gradual Learning Algorithm predict reranking order? Where do constraints come from?

OPTIMALITY THEORY AND THE CHILD’S OWN MAPPING

2. Pater’s Argument

The child’s output can be forced by a constraint that elsewhere is violated. This is what distinguishes OT from the constraints of phonology in the 1970’s and

1980’s

3. Example

In cases like banana [bæn], the ban on high-sonority onsets (like nasals) forces choice of [b] as the survivor consonant.

Yet nasal onsets die only when they compete with a better onset: Mommy comes out [mami], not *[ai].

/banana/ ALIGNFEETLEFT

FTBIN MAX-IO

*NASAL ONSET

*STOP ONSET

I-CONTIG

(bæn) ** * * **(næn) ** **!*(bnæn) *! ** **b(næn) *! ** *

/Mommy/ ALIGNFEETLEFT

FTBIN MAX-IO

*NASAL ONSET

*STOP ONSET

I-CONTIG

(mami) ***(ami) *! **(ai) *!* *

Class 2.2, 7/9/03 Child’s Own Mapping II p. 2

4. Hasse Diagram and Ranking Arguments

Requirements of utterly unmarked prosody force loss of a syllable—consonants of [bnæn] must now compete for space.

But in general, nasal onsets are kept, as in Mommy.

Better to keep a stop onset than a nasal onset, even if you have to “skip over” material in the input: [b n æn].

5. A Similar Data Pattern

[bfo], not *[blo], for buffalo

6. Larger Scale Analysis

I’ve done most of Amahl’s Stage 1; posted on course Web site.

ARE CHILD CONSTRAINTS THE SAME AS ADULT CONSTRAINTS? A SURVEY

7. Quick Answer

I think they’re strikingly similar, but it would be wrong to say that they’re identical. We need a theory that accounts for both similarities and differences.

8. Markedness Items Shared with Adult Phonology

Avoidance of consonant clusters Avoidance of non-prevocalic consonants—#CCV, VCCV, VC# Obstruents are voiced intervocalically, unvoiced initially and finally. A mild, characteristically violable preference for low-sonority onsets (Pater)

Comparison: onset-simplifying reduplication (Steriade, Phonology 1989)Ilokano ta-trabaho, not *ra-trabaho

Preference for harmony. The parallels with adults are vowel harmony, sibilant harmony, and nasality harmony—but children take this much further.

A dispreference, in strong-stress languages like English, for initial pretonic syllables (as in banana). Cf. diachronic effect in feckless, which was once effectless; also the

Class 2.2, 7/9/03 Child’s Own Mapping II p. 3

pervasive loss of pretonic function words in casual speech (Akmajian, Demers, Harnish and Farmer): Gotta go.

A strong dislike of [].

9. Vowel Harmony

Ingram, David (1974). Fronting in child phonology. Journal of Child Language, 1, 233-242.

Philip at age 1;9:

“If a word has more than one high vowel, the following constraint is operative:

0 0

or i-i; u-u, i-u but not *u-i.

10. Consonant-Vowel Affinities

The well-known affinities are:

labial consonants and rounded vowelscoronal consonants and front, esp. high vowels

Literature: Lyle Campbell (Language early 1970’s), Clements and Hume (1995) in Goldsmith Handbook of Phonological Theory, Edward Flemming (1995) UCLA diss.

For children: Clara C. Levelt (1996) “Consonant-vowel interactions in child language,” Proceedings of

the UBC Conference on Phonological Acquisition, Cascadilla Press, Somerville, MA.

11. CV Reduplication

Ingram, D. 1974. Phonological rules in young children. Journal of Child Language 1: 49-64, reporting work from 1899 by Roussey, on little Fernande learning French.

gloss word adult child age‘sit’ asseoir [aswa] [asss] 1;8‘play’ joue [u] [tutu] 1;8‘book’ livre [liv] [djidji] 1;9‘stick’ buchette [byt] [pkk] 1;6‘ball’ balle [bal] [babap] 1;8‘lady’ dame [dam] [dadap] 1;8‘nose’ nez [ne] [nene] 1;8

Cf. Samoan, also with pretonic CV reduplication. It’s not clear whether this child did CV reduplication to fulful a minimal word

requirement or just as a recreational activity; cf. 1;8 ‘coals’ charbon [ab] [b], with “upbeat” syllable truncation.

Class 2.2, 7/9/03 Child’s Own Mapping II p. 4

12. Templaticism

Priestly (1977; cited by Menn 1983; not yet seen by BH)

pillow [plo] [bijal]Brenda [bnd] [bajan]tiger [tag] [tajak]rabbit [æbt] [rajat]melon [mln] [majan]monster [mnst] [majos], later [mejan]dragon [æ [dajan], later [dajak]

This kind of process is known in Semitic and typologically similar languages.

cf. the 13th binyan of Arabic (McCarthy 1982, LI)

/ktb/ yields [ktawwab]

However, to my knowledge, templaticism is unknown as a straight phonological process.

I doubt this template is arbitrary; cf. a very young Peter Hayes [ ̍aja] ‘fan’

MORE ON THE RELATIONSHIP BETWEEN ADULT AND CHILD PHONOLOGY: SOME PUZZLING DIFFERENCES

13. Use of Adult Morphemes as Epenthetic Syllables

Upbeat syllable = initial and pretonicAmahl started using upbeat syllables around 2.134.Around 3.133, he partially standardized his upbeat syllable to [ri]

attack [ritæk] earlier [tæk]disturb [ristv] earlier [dv]design [ridzain] earlier [didain]giraffe [ridæf] earlier [diræf]Amita [rimit]thermometer [rimmt] earlier [mmt]

“For a brief period he experimented with /in-/ instead of /ri/ as the relevant prefix”

disgusting [igst]

See also Amalia Gnanadesikan’s daughter, with /fi/:

balloon [fibun]

Class 2.2, 7/9/03 Child’s Own Mapping II p. 5

Rebecca [fibk]

14. Fronting

Ingram, David (1974). Fronting in child phonology. Journal of Child Language, 1, 233-242.

In a word, labials must precede coronals, which must precede dorsals.

Philip at age 1;9:

alligator [ælge] [æge]animal [ænml] [mænu]candle [kændl] [nau]candy [kændi] [nai]Gumby [ cream [krim] [mi]hammer [hæm] [mænu]coffee [k{,}fi] [baki]

I haven’t ruled out the possibility that this will be found in an adult language, perhaps as a statistical effect or in a “Emergence of the Unmarked” (McCarthy/Prince 1995) context.

15. Sibilant Migration

Reference: Eric Hamp (1975) “Wortphonologie,” Journal of Child Language 1, 287-288.Alexander Hamp, aged 2 years 9 months.

string [trs] 2;9smell [meus] 2;9school [gus] 2;9spider [pats] 2;9squirm [kwms] 2;10

Other children who do this: Carol B. Farwell (1976) “Some strategies in the early production of fricatives,” Papers and Reports on Child Language Development 12, Stanford.

Is this English only? Cf. abundance of plurals, possessives, 3 sg. pres. verbs as models.

16. The /md/ Problem

When Amahl became capable of producing Nasal + Stop clusters, there first emerged [g], then [nd], then ... [md]!

number [nmd]camera [gmd]

“There was no hesitation or vacillation at all on A’s part. Even when asked to repeat an exaggeratedly stressed ‘number’, he consistently said [nmd].” Smith p. 106

Class 2.2, 7/9/03 Child’s Own Mapping II p. 6

This cannot be a lexical frequency effect: /md/ is overwhelmingly outnumbered by /nd/ in the adult lexicon: about 3000 to 30 in my electronic dictionary.

17. Upshot

Kid-constraints look far too much like adult constraints for it to be an accident, but one wonders if there is also not a certain amount of linguistic play going on.

WHAT CAUSES RANKING VALUES TO CHANGE?

18. A Brave Hypothesis

Paul Boersma and Clara Levelt (200) “Gradual Constraint-ranking Learning Algorithm predicts acquisition order,” with Clara Levelt. Proceedings of Child Language Research Forum 30, Stanford, California, pp. 229-237. ROA 361

Children start out with Faithfulness low (can’t say anything) Input data cause a reranking under the Gradual Learning Algorithm. Therefore, the reranking of constraints in the child’s production grammar will be tailored

to adult lexical frequency for marked structures.

19. Gradual Learning Algorithm

Invented: Boersma 19971

Applications to various problems in phonology: Boersma and Hayes 20012

learning optionality matching corpus frequency learning gradient well-formedness

Constraints have ranking values Constraints with higher ranking values more frequently outranked those with lower Hence, the OT grammar is stochastic, generating different outputs on different speaking

occasions GLA is error driven and gradual: when its current guess is wrong, it slightly demotes

constraints that favor the wrong guess, slightly demotes the constraints that favor the current learning datum.

20. Applications: Dutch Syllable Structure

Frequencies from a Dutch corpus:

1 How we learn variation, optionality, and probability. IFA Proceedings 21: 43-582 Empirical tests of the Gradual Learning Algorithm, Linguistic Inquiry 32: 45-86.

Class 2.2, 7/9/03 Child’s Own Mapping II p. 7

Constraints

*ONSET, *CODA, *COMPLEX ONSET, *COMPLEX CODAFAITH = MAX(C) + DEP(C)

History of simulation:

This may be interpreted as follows:

Stage 1: Markedness on top, only CV can be derived.Stage 2: FAITH crosses *CODA: codaful syllables become (gradually) possible.Stage 3: FAITH crosses ONSET: onsetless syllables become (gradually) possible.Stage 4: FAITH crosses the two *COMPLEX constraints in (probably) fairly random order;

complex onsets and codas become possible.

Running various underlying (adult) forms through the grammar, Boersma/Levelt find that they have characteristic “histories” of acquisition:

This looks fairly good matched up with Levelt’s data on acquisition order:

Dutch children do master codas before they master null onsets

Class 2.2, 7/9/03 Child’s Own Mapping II p. 8

Complex onsets and codas come in different orders for different children. Likewise different simulation runs give different outcomes, because they are so close.

21. Why I’m Skeptical

This is a tiny simulation, so the resemblance could be accidental--note that forms in adult language tend to occur in frequencies matching their markedness in any event; this provides an alternative explanation for any loose correlation.

Interchild variation. Boersma and Levelt’s scheme relies entirely on accidents of adult input to produce

differences between children learning the same language but the literature emphasizes a certain degree of creativity and problem-solving

among toddlers. Menn (readings): Some children render all their [l]’s as [j], some render all their

[j]’s as [l]’s. Her idea: stumble on the ability to same a few things, stick them for awhile for simplicity.

Gradualness issue. Boersma and Levelt rely entirely on the schedule of adult inputs to produce gradual change, yet change is not always gradual. Smith reports changes that took place between a morning and an afternoon, suggesting outright inspiration

22. A General Doctrinal Objection

Output phonology is not learned! It is invented. Previous work using the GLA has used it actually to learn; i.e. to detect and analyze

patterns in the ambient data. The GLA approach does force the rankings to evolve in a way that will maximally match

adult speech; thus I think it could be part of the answer, but not all of it.

LEARNING CONSTRAINTS

23. Preamble

I don’t think that output phonology is learned--but the actual constraints might be.

24. Where Do Constraints Come From?

I. Constraints are innate (McCarthy text 2002, Tesar/Smolensky 2000 book) II. Constraints are discovered; they are grammatical responses to patterns of phonetic

difficulty; both in perception and production--difficulty is part of experience, and thus accessible in principle.

Oeuvre of Donca Steriade (http://mit.edu/linguistics/www/steriade.home.html) Hayes (1999)—algorithm for projecting constraints from the child’s phonetic

experience (http://www.linguistics.ucla.edu/people/hayes/phonet.htm#pdp) Forthcoming book (Hayes, Kirchner, and Steriade, eds., Phonetically Based

Phonology; http://www.linguistics.ucla.edu/people/hayes/PBP/). Intro is posted on Web, and can be read for background on the more phonetically-based material below.

Class 2.2, 7/9/03 Child’s Own Mapping II p. 9

Why impose perceptually-grounded constraints? Makes sense if child is attempting to achieve intelligibility.

25. Maps

Idea is that phonetic experience is stored systematically, indicating what individual sounds and sound sequences are articulatorily difficult what sound pairs are perceptually difficult

Scanning such a map, the child could arrive at constraints that are phonetically effective phonological simple

26. A Plausibly Perceptually-Based System

Amahl: /nt/ [t] but /nd/ [d], and similarly for other places of articulation.

tent [dɛt]mend [mɛn]

Inspection of durations in spectrograms in adult speech shows a very short nasal in [nt], a very short stop in [nd].

Amahl is doing what is Articulatorily easier (no cluster, rapid velum transition) Less likely in each case to produce a mutilation that will render his speech

unintelligible.

27. Analysis

Inner arrows: *STRUC forces deletion, at cost of loss of poorly perceivable segment. Outer arrows: force the loss of the less perceivable segment, given a choice.

Class 2.2, 7/9/03 Child’s Own Mapping II p. 10

/tent/ MAX(LONG N)

MAX(LONG ALV STOP)

*STRUC MAX(SHORT N)

MAX(SHORT ALV STOP)

det *** **dent ****!*den *! ***

/mend/ MAX(LONG N)

MAX(LONG ALV STOP)

*STRUC MAX(SHORT N)

MAX(SHORT ALV STOP)

men *** **mend ****!*met *! ***

28. A Plausibly Articulatory System

“vegetative voicing”: obstruents are voiced intervocalically, voiceless peripherally. see Westbury and Keating 1986, Journal of Linguistics, for a modeling study that

explains why this is the articulatorily-vegetative outcome

29. The Consonant Harmony Mystery

Most toddlers display consonant harmony (defined as assimilation across intervening vowels), at a full-scale level involving major places (labial, coronal, velar).

This is very rare in adult language, which instead shows “antigemination” (McCarthy, LI 1986)

Proposal: Assimilation is ordinarily “no gap”; i.e. articulatorily continuous (Stampe,

Flemming, Gafos, Padgett) Children use jaw movement to produce consonant-vowel alternation. Hence the active articulator may be “deployed” throughout the word. Adults, with low jaw movement, can’t do this. Adults also have a faster speech tempo, where articulatory overlap implies greater

ease in producing non-harmonized sequences. Ideally, this sort of strategy should play a role in explaining why kid constraints are not

the same as adult constraints--the answer should emerge from differences between kid and adult articulation and perception.

AVOIDANCE

30. Avoidances: [p]

It is claimed by various people that in earliest stages, children simply refuse to say words that violate certain phonotactic constraints.

Ferguson and Farwell, Language 1975: “[Subject] T simply does not attempt an adult p word until [stage] VI, at which point she has four words beginning with [p].”

Class 2.2, 7/9/03 Child’s Own Mapping II p. 11

31. Avoidance: Fricatives

Reference: Carol B. Farwell (1976) “Some strategies in the early production of fricatives,” Papers and Reports on Child Language Development 12, Stanford.

“For two months, JB said virtually no fricative words. For another two months, he produced only one fricative word with regularity. But then, he began to learn fricative words and within a month 50-60% of the words he produced contained fricatives (this high percentage not due to frequent use of the plural morpheme). Almost all these fricative words were indeed produced with fricatives, and fricatives were overextended—boo was produced for some time as boof.

32. Semi-Avoidance

Jacob Hankamer, studied by Menn, converted /b/ to [d] in bye-bye, avoided all other /b/ words.

Mechanism of avoidance is in principle not hard, if there is a Lexicon of Me. Just don’t install any of the forbidden type of word. Hence, self-monitoring to avoid saying hard words is not necessary.

33. Avoidance in OT

There is a small literature on this.

Avoid the illegal by arranging the grammar so that is becomes something else--can’t help here.

Add a special CONTROL component to filter the output of OT, with undominated, *-assigning constraints. Orgun, Orhan and Ronald L. Sprouse. 1999. From MPARSE to CONTROL: deriving ungrammaticality.

Phonology 16:191-224.

Let the “null parse” (Prince and Smolensky 1993) be a candidate. In the view of McCarthy (in press),3 It obeys all Markedness constraints, violates all Faithfulness constraints.

34. Null Parse Example

For labial-avoiding Jacob, we would have:

/boi/ *LABIAL IDENT(PLACE) *NULL PARSE

null **doi *!*boi *!

3 “Against Gradience”, no longer Web-posted, but in press in Phonology.

Class 2.2, 7/9/03 Child’s Own Mapping II p. 12

35. General Patterning of Null Parse

Any Markedness constraint dominating *NULL PARSE will be respected. It will be respected by avoidance if the Faithfulness constraints for any possible “fix”

dominate *NULL PARSE. Else the form will be fixed, by whatever candidate violates the lowest-ranking

Faithfulness constraint. Example: [f] in Tagalog words, not rejected but fixed optionally as [p].

36. Try It…

McCarthy’s version of the null parse theory is different from the earlier Prince/Smolensky version and worth testing further, including, I think, for gradient well-formedness (null parse wins only stochastically).