violable constraints in language production: testing the ... · c. transitivity is a crucial...

Journal of Memory and Language42, 272–299 (2000)doi:10.1006/jmla.1999.2679, available online at http://www.idealibrary.com on

Violable Constraints in Language Production: Testing the TransitivityAssumption of Optimality Theory

Daniel J. Guest, Gary S. Dell, and Jennifer S. Cole

University of Illinois at Urbana-Champaign

Optimality Theory (OT) (Prince & Smolensky, 1993) characterizes linguistic knowledge as aranked set of constraints that select the best possible output form of a word given a particular input.OT assumes that constraints are ordered transitively with respect to their violability. An artificiallanguage learning paradigm was used to test this assumption by teaching participants to pronouncewords that provided evidence about three constraints affecting the stress patterns of words. The wordsdemonstrated that the first constraint outranked the second and the second outranked the third. Therelationship between the first and third could only be derived from the transitive nature of the system.Three experiments tested whether speakers could determine the stress patterns of words requiringknowledge of the relationship between these two constraints. Evidence was found for a transitivelyordered constraint system as well as a system that stores commonly heard stress patterns as metricaltemplates. © 2000 Academic Press

Key Words:Optimality Theory; stress; phonology; language acquisition; language production;
transitive inference.
y,in

e-arlrns

efu9797

nalarles

ofn-e ofn-thatam-Ynt–

sed.ni-m-

acehatlla-itialedatgeinoft

andente in

ruc-

nuagy,icasy-

tion98

191hery f

co

ntoisUr

Optimality Theory (Prince & Smolensk1993), or OT, is an important new paradigmlinguistics. In phonology OT has quickly bcome the dominant framework. Since the e1990s, numerous conference talks and jouarticles in this area have presented their ideaOT terms. OT has also proven to be a usframework for syntactic (e.g., Pesetsky, 19Speas, 1997) and morphological (Russell, 19analyses.

The OT approach diverges from traditiolinguistic analyses by proposing that grammconsist of violable constraints rather than ru

Portions of this work were presented at the 2nd anmeeting of the Mid-Continental Workshop on Phonolothe 69th Annual Meeting of the Midwestern PsychologAssociation, and the 38th Annual Meeting of the Pchonomics Society. The research was supported by NaScience Foundation Grants SBR 93-19368 and SBR73450 and National Institutes of Health Grants DC-00and HD-21011. The authors thank Kay Bock, Cindy FisStefan Frisch, Susan Garnsey, and Paul Smolenskhelpful discussions and Arva Bensaheb for assistancelecting data.

Correspondence and reprint requests should be seDaniel Guest at the Beckman Institute, University of Illinat Urbana-Champaign, 405 North Mathews Avenue,
bana, IL 61801. E-mail: [email protected].
2720749-596X/00 $35.00Copyright © 2000 by Academic PressAll rights of reproduction in any form reserved.

yalinl;)

s.

In OT, Universal Grammar defines a setuniversal constraints that function in all laguages to impose conditions on the structurlinguistic surface forms. In phonology, the costraints govern the features and structuresdetermine a language’s sound forms. For exple, the Weight to Stress Principle, or HEAVconstraint, requires that heavy consonavowel–consonant (CVC) syllables be stresA defining characteristic of OT is that the uversal constraints often impose mutually incopatible requirements on an individual surfform. For example, consider the conflict tcan arise in a language with heavy CVC sybles, and in which stress is located on the insyllable of a word, due to a constraint termALIGN-FOOT(left) that requires a stress footthe left edge of the word. If such a languaincludes a word with a heavy CVC syllablesecond position, the combined effectsALIGN-FOOT(left) and the HEAVY constrainwould be to place stress on both the initialsecond syllables. This pattern of adjacstresses violates another constraint operativstress systems, CLASH, which bans any st

l

l

al-

,orl-

to

-
ture that contains adjacent stressed syllables.

Y,ed i

OTn-

; inheofhendanyl--th

ngostheen

ntthaIf

ut-erou

deten-

fordeerafopaarerere

iatri-aptemlex

lestoh iin-gi-is

of),pact&

h ahatt ortrictiza-ys-

ro-usthe

ofon-ism

rmsandgi-ainten,

byust

ensalen-ge-as-hen-n

heiderbe--atCby

nn-int

A ce

273VIOLABLE CONSTRAINTS IN LANGUAGE PRODUCTION

Thus, there is a conflict among HEAVALIGN-FOOT(left), and CLASH: These threconstraints cannot be simultaneously satisfiea word with a heavy CVC second syllable.resolves constraint conflict by ranking all costraints in a transitive dominance hierarchythe case of conflict, priority is given to thigher ranked constraint, at the expenseviolation of the lower ranked constraint. In texample just mentioned, ranking CLASH aHEAVY higher in the constraint hierarchy ththe constraint requiring stress on the initial slable, ALIGN-FOOT(left), will result in a surface form with stress on the second but notfirst syllable of this word. The relative rankiof constraints in the hierarchy varies acrlanguages, which accounts for variation insurface phonological patterns found in differlanguages.

In this paper, we present three experimethat tested a key assumption of OT, namely,the constraint ranking relation is transitive:Constraint A outranks Constraint B, and B oranks C, then A also outranks C. In the expiments, participants learned constraints abhow to stress words in artificial languagesrived from OT grammars. They were then teson novel items requiring the application of trasitivity. As will be seen, we found evidencetransitivity and, thereby, for a system thattermines pronunciations from ranked genconstraints. But we also found evidenceanother process in learning words’ stressterns, a process in which novel wordsstressed by matching them to already stostress patterns or templates. Thus, oursearch—subject to the usual caveats assocwith artificial language learning expements—is consistent with a dual-systemproach to linguistic performance, one sysassociated with a grammar and one with aicon.

Our studies asked whether OT principhave utility in explaining how people learnproduce word forms. As such, the researcpsycholinguistic in the traditional sense: A lguistic theory is examined for its “psycholocal reality.” OT is particularly promising in th
vein. The theory has links to, and has evolved ac
n

a

e

s

t

st

-t

-d

-l

rt-

d-ed

-

-

s

least partially from, connectionist theoriescognition (Rumelhart & McClelland, 1986and these theories have had considerable imon psycholinguistics (see ChristiansenChater, in press, for review). OT is very mucpart of the Zeitgeist in cognitive science teschews hard-and-fast rules in favor of sofviolable constraints. At the same time, the sranking of constraints represents a crystalltion of a set of violable constraints into a stematic and learnable language.

Before describing the experiments, we pvide an overview of OT and review previopsycholinguistic research on OT and onpsychology of transitive inference.

OVERVIEW OF OPTIMALITY THEORY

The architecture of OT consists of a setuniversal constraints, a relation that ranks cstraints in a dominance hierarchy, a mechanfor generating candidate surface output fothat are evaluated against the lexical inputfor surface well-formedness by the phonolocal constraints, and a mechanism for constrevaluation (see Archangeli & Langendo1997, for an overview of the theory).

The claim that constraints are providedUniversal Grammar means that constraints min principle be violable; if they were not, thall languages would conform to the univerconstraints in a similar fashion and display idtical surface phonological patterns. Languaspecific constraint ranking is thus a criticalpect of the theory, since it provides tmechanism for determining which of two coflicting constraints will be violated in a givelanguage.

At this juncture it is useful to introduce tnotation that is adopted in this paper. Consagain the example from above of conflicttween the constraints HEAVY and ALIGNFOOT(left). We stated that the conflict tharises in the evaluation of a word with a CVsyllable in second position can be resolvedranking CLASH and HEAVY higher thaALIGN-FOOT(left). Ranking is noted as: Costraint A .. Constraint B, read as “Constra

outranks Constraint B.” Evaluation of surfa
tandidates for such a lexical input is illustrated

thergeateshe

gera-

inceo-

eft-theellhaw-intmaonAllrerelanthaonbo

e-ble“.”nd

a-irdsevi-rsits-

the2,

esig-ing

e,

Ta-al

hem-theat

tA inctc ints.N ol-u notr

usee is oft thew the

-c

274 GUEST, DELL, AND COLE

in Table 1, which presents an OT table informat proposed by Bernhardt and Stembe(1998) (this format diverges slightly from thof the mainstream OT literature, but providfor a potentially cleaner presentation. Treader is referred to Bernhardt & Stember1998, particularly Chapter 4, for the motivtions behind the changes).

The lexical input, or underlying, form isthe top, left corner, with candidate surfaforms following across the top row. The phnological constraints are listed down the lmost column, in a vertical order that reflectsdominance ranking. An asterisk (*) in a cmarks constraint violation for the candidate theads the column. An exclamation point folloing an asterisk (*!) indicates a “fatal” constraviolation that renders the candidate suboptirelative to other candidates that satisfy the cstraint evaluated in that, or higher, rows.cells in rows lower than the fatal violation ashaded, indicating that their marks are notevant to the determination of the optimal cdidate. The winning candidate is the onefares the best under the highly ranked cstraints and is indicated with the pointer sym(Z).

In the example given in Table 1, “CV” dnotes an open consonant–vowel sylla“CVC” denotes a closed heavy syllable, anddenotes a syllable boundary. The second cadate violates HEAVY, incurring a fatal violtion on this highly ranked constraint. The thcandidate, which stresses both the first andond syllables, violates CLASH, also a fatalolation. The winning candidate, then, is the ficandidate, which satisfies HEAVY, despiteviolation of the lower ranked ALIGN-FOOT

TABLE 1

OT Table Demonstrating Evaluation of SurfaceCandidates

cv.cvc.cv Zcv.cvc.cv cv.cvc.cv cv.cvc.cv

CLASH *!HEAVY *!ALIGN-FOOT(left) *

(left). Changing the domination order will

r

,

t

l-

--t-l

,

i-

c-

t

change which candidate is chosen as beingmost optimal. This is demonstrated in Tablewhere the form cv´.cvc.cv is determined to bthe optimal candidate, and therefore the denated surface form, under the new rankCLASH, ALIGN-FOOT(left) .. HEAVY forthe input.

THE ROLE OF TRANSITIVITY IN OT

The ranking relation in OT is transitivwhich means that if A.. B, and B.. C, thenA .. C. Transitivity is a crucial feature of Oconstraint ranking. Without transitivity, evalution can fail to identify any candidate as optimfor a given lexical input form. Consider tfollowing schematic example. Assume a gramar with three constraints, A, B, and C, andranking A.. B and B.. C. Now assume thhe grammar has the nontransitive ranking C... Table 3 shows the evaluation of three distandidate forms under these three constraote that constraint C is listed first in the cmn of constraints, although that order doeseflect its ranking beneath constraint B.

This table does not indicate a winner, becaach candidate must be rejected on the bas

he proposed ranking. Output-a cannot beinner because it loses to output-b under

TABLE 2

OT Table Demonstrating the Effects of ReorderingConstraints

cv.cvc.cv cv.cv´c.cv Zcv.cvc.cv cv.cvc.cv

CLASH *!ALIGN-FOOT(left) *!HEAVY *

TABLE 3

Evaluation of Candidates under a NontransitiveConstraint Ranking

Input Output-a Output-b Output

Constraint C *!Constraint A *!Constraint B *!

er thhede

setinerelidthicantion-itivbeor-

n-ar,wit

ittr a

o

ar&

7)

enryereAr-syndline-

caey

decalnoth

eadinfoighthe

ev-

ho-la-95,&the, &

p-sings as

oofted

t ni-m r, &Z 7;m 96;s rs,1 son,S 92)a 67;4 lds,L n,r hesei ple,s fer-e ula-t es(

ansi-t .g.,L dp ess .g.,B c-K n,1 ghto e,w so-c tov als”c ivei uch


ranking A .. B. Similarly, output-b cannot bthe winner because it loses to output-c underanking B.. C. Finally, output-c cannot be twinner either because it loses to output-a unthe ranking C.. A. Selecting any one of thecandidates as the winner entails contradicone of the assumed constraint rankings. Thfore, if all three constraint rankings are vathen there is no winner in a case such asand the grammar is incomplete because itnot guarantee a decisive outcome of evaluafor an individual input. This example demostrates that constraint ranking must be transin OT in order for any OT grammar tocomplete—a minimum requirement for a fmal theory.

Transitivity is also important for the learability of an OT grammar. To learn a grammone must learn how each constraint ranksrespect to all other constraints. In a system wn constraints, there aren! possible constrainrankings which, on the surface, makes fodaunting learning task. However, becausetransitivity, at a minimum, onlyn(n-1) informa-tive constraint interactions are needed to lethe correct ranking for the system (TesarSmolensky, 1993; Prince & Smolensky, 199

PSYCHOLINGUISTICS AND OT

Although a large amount of work has bedone in phonology utilizing Optimality Theo(at the time this article was written, there w319 papers listed in the Rutgers Optimalitychive alone), there has been relatively little pcholinguistic research on OT. Bernhardt aStemberger (1998), one of the few psychoguistic works employing OT, applied the framwork to a variety of phenomena in phonologiand prosodic development. In addition, thwere able to employ constraint rankings toscribe the deficits of children with phonologidelays. This is significant because they didneed to posit ad-hoc constraints to handlelanguage disorders of these children. Instlanguage performance was explained by findthe appropriate aberrant constraint rankingeach child and seeing how these rankings mchange through successful intervention. O
work on phonological development has alsov
e

r

g-

s,-n

e

hh

f

n

.

-

-

l

-

te,

grtr

emphasized the use of violable constraints (Lelt, 1994).

OT has also been used to interpret psyclinguistic investigations of the role of the sylble in speech perception (Hammond’s, 19interpretation of Cutler, Mehler, Norris,Segui, 1983, 1986) and in investigations ofstructure of Hebrew words (Berent, EverettShimron, 1998).

TRANSITIVE INFERENCE

Transitive inference, in various forms, apears to be a pervasive information procesability. The essence of such an inference ifollows: The truth of a relation between twitems, A. C, can be inferred from the truthA . B and B. C. Research has demonstrahat transitive inference occurs in various aal species (e.g., pigeons, Steirn, Weaveentall, 1995; Weaver, Steirn, & Zentall, 199acaque monkeys, Treichler & Tilburg, 19

quirrel monkeys, McGonigle & Chalme977, 1992; chimpanzees, Boysen, Bernthreyer, & Quigly, 1993; and rats, Davis, 19nd in children (e.g., 7 year olds, Piaget, 19year olds, Trabasso, 1977; and 22 month oipkens, Hayes, & Hayes, 1993). In additioesearch has investigated mechanisms for tnferences in particular domains. For examtudies have examined whether transitive innces about ordered sets involve the manip

ion of propositions (Clark, 1969) or imagHuttenlocher, 1968).

There have also been suggestions that trive inferences can be made implicitly (eewicki, Hill, & Czyzewska, 1994). Mediateriming (lion primes tiger, and tiger primtripes, therefore lion primes stripes; ealota & Lorch, 1986; McNamara, 1992; Moon & Ratcliff, 1992; O’Seaghdha & Mari997; Shelton & Martin, 1992) can be thouf as a kind of implicit transitive inferenchere the relevant relationship is one of asiation. In a similar vein, people’s abilityerify statements such as “canaries are animan also be interpreted as implicit transitnference, provided that the mechanism for s
erification are the previously stored proposi-

are

ectharenel-re

lleracdgeersdsa

itivnseple

intere.Onkic-

( ntl eret ingt e.I icteb ah

P

dua edi irem and ndit ).

S

ons thag bicw

w ghts ando ov-e r ofc ofm twol andt om-p nts.T bledt rtic-i thel et oft s af toi r-e owl-e ntsw essp En-g

sp rmso

.le

o

llytressore

ss.the

bleR-

ern

uageh tressc f thec con-s temsi onH iledc hese


tions that “canaries are birds” and “birdsanimals” (Collins & Quillian, 1969).

For our purposes, the most important aspof this literature are the data suggestingtransitive inference occurs in young childand may occur implicitly. Phonological devopment largely occurs between birth (or befoand age 4 (Bernhart & Stemberger, 1998; O1980; Stark, 1980; Vihman, 1996) and thequisition process and the resulting knowleare generally considered implicit. If speakcompute the phonological forms of worthrough a system resembling an OT grammone might expect to see evidence for transinference as the system is acquired. Coquently, our studies investigated how peolearn the form of words.

We used a laboratory learning paradigmwhich adult speakers learned the stress patof syllable strings from an artificial languagThe stress patterns were derived from angrammar, and the training illustrated the raing of particular pairs of constraints, in partular that Constraint A.. B and that B.. C.The particular constraints varied in differeanguage conditions). The participants when asked to pronounce novel syllable strhat put A and C into conflict for the first timf speakers produce the stress pattern predy the grammar for these test items, they mave done so by using transitivity.

EXPERIMENT 1

Method

articipants

Twenty native English-speaking undergrates from the University of Illinois participat

n exchange for credit toward a course requent. Equal numbers of participants were romly assigned to one of two language co

ions, language A (LA) and language B (LB

timulus Materials

Two languages were created that demtrated the interaction among constraintsovern the location of stress in polysylla
ords. Words in each of the two languagesd
tst

),-

r,e-

ns

T-

s

dy

-

---

-t

ere built up from a lexicon of the same eiyllables, consisting of seven CV syllablesne CVC syllable. Stress placement was grned in each language by a small numbeonstraints adopted from the linguistic theoryetrical stress. The stress patterns of the

anguages were mirror images of one other,heir respective constraint systems were cosed of identical, but mirror-image constraihis opposing nature of the languages ena

hem to serve as controls for each other. Papants were trained on words from one ofanguages and then were given a common sest words. Differences on the test items aunction of training can therefore be usedndex learning. Importantly, any such diffences cannot be ascribed to participants’ kndge of English phonology since participaould be expected to produce similar stratterns given the same test words if theirlish phonology was at work.Language A stress constraints.The stres

atterns of LA words can be described in tef four basic properties:

1. The penultimate syllable of the word is stressed2. Stress is placed on every odd-numbered syllab

counting from the beginning of the word, exclud-ing the final syllable.

3. Every heavy syllable is stressed.4. Adjacent syllables may not both be stressed (n

stress clash).

Each property of LA stress is individuaattested in a number of natural language ssystems and can be attributed to one or mconstraints that govern the location of stre1

Properties 1 and 2 are attributed here toconstraints PENULT (the penultimate syllaof a word must be stressed) and LEFT-ALTENATING (there must be an alternating pattof stressed–unstressed stated below).

Both PENULT and LEFT-ALTERNATING

1 The constraints are described in nontechnical langere, to facilitate ready understanding of the essential sharacteristics of the two experimental languages. All oonstraints described here are modeled after actualtraints proposed for the analysis of metrical stress sysn OT by McCarthy and Prince (1993), drawing heavilyayes’ (1995) theory of metrical stress. A more detaharacterization of the metrical constraints relevant to t
ata is found in the Appendix.

ationtionat

ythn iir

ce-r aofre

Vof

isfy

onyLT

ssddsuisresonletedles

ssdd

oftisfse

cand

l-

e byrn ofs.ca-eavysideby

s isas

Y).

lesnul-rmisinton-nce,theeon-

tion

dusen-ion

edre-ordressr-m-


are composite constraints, built up from a smset of atomic OT constraints whose interacgoverns the location and internal composiof the metrical foot, the structural unit thunderlies stress marking in systems with rhmic stress. The atomic constraints are givethe Appendix, along with an illustration of theindividual functions in determining stress plament in LA. Leaving aside heavy syllables fomoment, in words with an even numbersyllables, the pattern of alternating stressesquired by LEFT-ALTERNATING will locate astress on the penultimate syllable, as inssssand ssssss, wheres refers to a single Csyllable. A form containing an even numbersyllables with this stress pattern would satboth the PENULT and LEFT-ALTERNATINGconstraints. For words with an odd numbersyllables, there is not, strictly speaking, astress pattern that could satisfy both PENUand LEFT-ALTERNATING. Locating a streon the penultimate syllable and on the onumbered syllable that precedes it would rein a stress clash, as insssss. Stress clashnot tolerated in many natural language stsystems; there are no words in the LA lexicthat present stress on two adjacent syllabThe avoidance of stress clash in LA is attributo the CLASH constraint (adjacent syllabmay not both be stressed).

In LA, CLASH is avoided by locating streon the penultimate syllable, but not on the onumbered syllable that precedes it, as insssssor sssssss. Words with an odd numbersyllables that display this stress pattern saboth CLASH and PENULT, but at the expenof a violation of LEFT-ALTERNATING. Theinteraction between these three constraintsbe modeled in OT by ranking CLASH a

TABLE 4

CLASH, PENULT.. LEFT-ALTERNATING

sssss Zsssss sssss sssss

CLASH *!PENULT *!LEFT-ALTERNATING * *

ll

-n

-

f

-lt

s

s.

-

y

n

PENULT above LEFT-ALTERNATING, as ilustrated in Table 4.

The three constraints discussed so far arthemselves responsible for the stress patteall words that contain only light syllableHeavy syllables introduce a further complition because they attract stress. That is, a hsyllable is stressed even when it occurs outof a position that is designated for stressPENULT and LEFT-ALTERNATING. Thestress-attracting behavior of heavy syllableillustrated by LA words with forms suchsssH and sssHss, wheresH refers to a CVCsyllable, and is attributed here to the HEAVconstraint (a CVC syllable must be stressed

The principle of stressing heavy syllaboverrides the placement of stress on the petimate syllable in LA, as examples of the fosssH or sssHss demonstrate. This resultachieved by ranking the HEAVY constraover PENULT. These examples also demstrate the principle of stress clash avoidawhich rules out the possibility of stressingheavy syllablein addition to the penultimatsyllable in these forms. The necessary cstraint rankings are illustrated in the evaluaof the example fromsssH in Table 5. TheLEFT-ALTERNATING constraint is includein Table 5 for completeness, although becaof its low ranking relative to all the other costraints, it plays no active role in this evaluatscheme.

The final property of LA to be discussconcerns exceptions to the principle thatquires stress on every heavy syllable. If a wcontains two adjacent heavy syllables, stwill fall on only one of them, due to the overiding effect of stress clash avoidance. Exa

TABLE 5

CLASH, HEAVY .. PENULT

sssH ZsssH sssH sssH

CLASH *!HEAVY *!PENULT *LEFT-ALTERNATING * *

ples are words of the formsHsHsss and

oa ineb resp SHo rm

te,g

c ofL to reso lsos le

arnsess

le

o

isesorss

ianhe-artsalandl-yll

ingv-n-ntal

heI-

-of

onas.d 7,-

-ges

e-lan-

8,re

talCV)so,avy

ith at po-s to-ko [do],[ Thus,t mosta wel–g con-t of a


ssHsHsss. The choice of which of the twdjacent heavy syllables to stress is determy the other ranked constraints. These statterns are accounted for by ranking CLAver HEAVY, as is demonstrated with the fo

sHsHsss in Table 6.Note that in Table 6, the third candida

sHsHsss, is eliminated in favor of the winninandidate on the basis of its two violationsEFT-ALTERNATING. While neither the firsr third candidates succeed in placing a stn the third syllable, the third candidate auffers from its failure to stress the first syllabLanguage B stress constraints.The words in

Language B (LB) exhibit stress patterns thatthe mirror images of the LA stress patterThus, the four basic properties of LB strpatterns are

1. The second (pen-initial) syllable of the word isstressed.

2. Stress is placed on every odd-numbered syllabcounting from the end of the word, excluding theinitial syllable.

3. Every heavy syllable is stressed.4. Adjacent syllables may not both be stressed (n

stress clash).

As with LA, each property of LB stressindividually attested in natural language strsystems and can be attributed to one or mconstraints that govern the location of streThis system can be accounted for with a varof the constraint set for LA. In particular, tPENULT and LEFT-ALTERNATING constraints that govern LA have their counterpin the constraints PEN-INITIAL (the pen-initisyllable of a word must be stressed)RIGHT-ALTERNATING (there must be an aternating pattern of stressed–unstressed s

TABLE 6

CLASH .. HEAVY

sHsHsss ZsHsHsss sHsHsss sHsHsss

CLASH *!HEAVY * *PENULTLEFT-

ALTERNATING * * **!
(
ds

s

.

e.

se.t

a-

bles starting the right edge of the word goleftward, excluding the initial syllable) that goern LB stress. The CLASH and HEAVY costraints play the same role in both experimelanguages.

The constraint interactions in LB are tsame as those in LA, substituting PEN-INTIAL and RIGHT-ALTERNATING for PE-NULT and LEFT-ALTERNATING, respectively. For example, in LB the rankingCLASH and PEN-INITIAL over RIGHT-AL-TERNATING accounts for the stress patternwords with an odd number of light syllables,in sssss or sssssss, as shown in Table 7

As can be seen by comparing Tables 4 anthe substitution of PENULT and LEFT-ALTERNATING in LA with PEN-INITIAL andRIGHT-ALTERNATING in LB produces mirror-image stress patterns in the two languawith identical constraint violations of their rspective constraints. This aspect of the twoguages can be further examined in Tablewhich lists all of the LA and LB forms that weused to train the participants.

Words.The words for the two experimenlanguages were created from the open (syllables of the musical scale: do, re, mi, fa,la, it, do. To this set we added a single he(CVC) syllable: “ton,” pronounced [to:n].2 The

2 When stressed, the open syllables are produced wense dipthongal rime, e.g., [dow], [rej]. In unstressedition, subjects varied in their productions, with someens produced with reduced vowels, e.g., [də], [rə], andther tokens produced with a tense vocalic rime, e.g.,

re], shorter in duration than the stressed counterparts.he open syllables of the experimental languages areccurately characterized as CV: (or consonant–volide), at least in stressed tokens. The open syllables

rast with the heavy syllable primarily in the absence

TABLE 7

CLASH, PEN-INITIAL .. RIGHT-ALTERNATING

sssss Zsssss sssss sssss

CLASH *!PEN-INITIAL *!RIGHT-ALTERNATING * *

nonvocalic) coda consonant, which is the basis for our

bederture

o

nth

oo”rdthein-ofldple

ithourdes,,”

to-on

res

E

the

r n-ss

othL het edo theC si-t glec calr H. -NdHPC thatn ela-t -NR

ngs s toa itht itht e.F fas lo nt

set, for

n-nts

forble


syllables of the musical scale were chosencause of their familiarity. The sequential orof the syllables was another important feathat was exploited in constructing the wordsthe experimental languages.

In both LA and LB, the CV syllables withia word were kept in the sequence defined bymusical scale. This sequence defines a lwhere “ti” can be followed by “do,” “do re,and so on, continuing through the series. Wowere created by varying word length andstarting syllable of the word. Words had a mimum of three syllables and a maximumseven syllables. The CVC syllable “ton” coube placed anywhere in the order. Examwords (with no stress indication) arere mi fa sola, la ton ti do re,and ton ton mi fa so.

Words for LA and LB were constructed wthe purpose of inducing conflict among the fstress constraints. For example, LA incluwords with an odd number of CV syllablesuch as “do re mi fa so” or “mi fa so la ti do repatterns which were seen above to leadconflict between PENULT and LEFT-ALTERNATING. The stress pattern presentedwords such as these shows penultimate sti.e., “do re mi fa so” and “mıfa sola ti do re,”and demonstrates the ranking of CLASH, PNULT .. LEFT-ALTERNATING. Similarly,there were words in LA that demonstratedranking of HEAVY .. PENULT, such as “to´n

choice of labels: “CV” for the open syllables and “CVC”the closed syllables. For further discussion of sylla

TAB

LA and LB Training T

Language A

CLASH ..HEAVY

HEAVY ..PENULT

PENULT .. LEFT-ALTERNATING

sHsHsss (28) sHss (28) sssss (28)ssHsHsss (28) sssH (28) sssssss (28)

sssHss (28) sHssssss (6)sssssHss (6) sssHssss (28)sssssssH (28)

weight, see the Appendix.

-

f

ep,

s

s

a

s,

-

e mi” and “sola ton,” and words that demotrated the ranking of CLASH.. HEAVY,uch as “to´n ton mi faso.”Two sets of words were created, serving b

A and LB: training words and test words. Training words for the two languages differnly in that the placement of stress and ofVC syllables occurred in mirror-image po

ions. Each training word demonstrated a sinonstraint ranking from a set of three critiankings: for LA these rankings were CLAS. HEAVY, HEAVY .. PENULT, and PEULT .. LEFT-ALTERNATING; for LB theemonstrated rankings were CLASH..EAVY, HEAVY .. PEN-INITIAL, andEN-INITIAL .. RIGHT-ALTERNATING.rucial to the design of the training set wasone of the training words demonstrated a r

ionship between HEAVY and LEFT-ALTERATING (LA) or between HEAVY andIGHT-ALTERNATING (LB).There were 11 different types of traini

timuli for each language, where a type refergroup of seven words of the same length, w

he syllable “ton” in the same position, and whe words differing only in the initial syllablor example,do re ton, re mi ton, mi fa ton,o ton, so la ton, la ti ton,and ti do ton are alf the typesssH. Table 8 shows the differe

types of words that were used in the trainingto demonstrate each of the three rankingsboth LA and LB.

The test items probed transitivity in costraint ranking by putting the constrai

8

s (Number of Tokens)

Language B

LASH ..HEAVY

HEAVY ..PEN-INITIAL

PEN-INITIAL ..RIGHT-ALTERNATING

sHsH (28) sssH (28) sssss (28)sHsHs (28) sHss (28) sssssss (28)

sssHss (28) sssssssH (6)sssHssss (6) sssssHss (28)sHssssss (28)

LE

ype

C

ssssss

HEAVY and LEFT-ALTERNATING (RIGHT-

stth

t th-

ofth

n-lesnvitye-isinns

tleo

ataresanbuthir-

-thhe-

ening

A hetL t-t ini

G

oftem,,

etedes-of

eed1.ive

e-ici-LAIn

ereananhislesse

thegin

HR


ALTERNATING in LB) into direct conflict forthe first time. In producing forms from the teset, participants were required to determineplacement of stress and thus had to confronconflict between HEAVY and LEFT-ALTERNATING (RIGHT-ALTERNATING in LB).All of the test words were of three types. Twothese types had five syllables and testedcritical transitive relation for the trained laguage only. The third type had seven syllabwith the syllable “ton” in the fourth positio(e.g., so la ti ton do re mi), and tested transitiin both LA and LB. Table 9 illustrates thconflict between HEAVY and LEFT-ALTERNATING that arises in the evaluation of thform in LA. The first candidate succeedsstressing the heavy syllable, but at the expeof a LEFT-ALTERNATING violation thastems from its failure to stress the third syllabThis candidate incurs a second violationLEFT-ALTERNATING due to its failure tostress the fifth syllable, but the training dalready establish the dominance of penult stin this situation (see Table 4). The second cdidate fails to stress the heavy syllable,succeeds in placing stress on the first andsyllable, as required by LEFT-ALTERNATING. The single violation of LEFT-ALTERNATING is due to the absence of stress onfifth syllable, as with the other candidate. Tranking between HEAVY and LEFT-ALTERNATING will determine the choice betwethese two candidates. The independent rankHEAVY .. PENULT and PENULT.. LEFT-

LTERNATING have been established in training set, and transitivity gives HEAVY..EFT-ALTERNATING. Thus, the stress pa

ern expected on the basis of transitive ranks the first candidate in Table 9,ssssHsss.

TABLE 9

Seven-Syllable Test Item for LA: Conflict betweenHEAVY and LEFT-ALTERNATING

ssssHsss ZssssHsss ssssHsss

HEAVY *!LEFT-ALTERNATING ** *

The stress pattern of the second candidate im

ee

e

,

e

.f

s-td

e

s

g

plies the dominance of LEFT-ALTERNATINover HEAVY, defying transitivity.

Since the heavy syllable is in the middlethe seven-syllable test words, the same issssHsss, serves to test transitivity in LBthis time putting HEAVY into conflict withRIGHT-ALTERNATING.

The five-syllable test type with “ton” in thsecond position (e.g., mi ton fa so la) testransitivity in LA only. Table 10 demonstratthe conflict between HEAVY and LEFT-ALTERNATING that arises in the evaluationthis form in LA.

The five-syllable test type with “ton” in thfourth position (e.g., so la ti ton do) testtransitivity in LB only, as shown in Table 1The possible test words with their transitstress patterns are shown in Table 12.

Notice that if transitivity applies, the fivsyllable transitive test items require the partpant to begin with an unstressed syllable inor end with an unstressed syllable in LB.order to give transitivity a stringent test, thwere no training stimuli that began withunstressed syllable in LA or ended withunstressed syllable in LB (see Table 8). Tmeant that the word type of three CV syllab(sss) was not included in the training, becauboth LA and LB stress these words onmiddle syllable. Other word types that be

TABLE 10

Five-Syllable Test Item for LA: Conflict betweenHEAVY and LEFT-ALTERNATING

ssHsss ZssHsss ssHsss

HEAVY *!LEFT-ALTERNATING ** *

TABLE 11

Five-Syllable Test Item for LB: Conflict betweenHEAVY and RIGHT-ALTERNATING

ssssHs ZssssHs ssssHs

EAVY *!IGHT-ALTERNATING ** *
-

ithn-dthe

ica

ioning

pro

eyew

romaticath

ewld,s.uran

s aprse

itiosar itsybleblyse

xpe

rteritchwell.in-

f thelsoon-

ch ofwiceter-

uc-thetheyntertedor-ionov-ionthe

fromthatf thetici-ter-inses

rderher

s


with unstressed syllables in LA or end wunstressed syllables in LB involve HEAVY iteracting with LEFT-ALTERNATING anhence they were already excluded fromtraining stimuli because they place the crit(transitive) constraints in conflict.

Procedure

The experiment had four parts: a repetitphase and a production phase during trainfour internal tests spaced throughout theduction phase, and a final test.

Training. Participants were told that thwould be learning to pronounce words in a nlanguage and that this language differed fEnglish in two ways. First, the syllables thmade up the words were those from the musscale: do, re, mi, fa, so, la, and ti, as well assyllable “ton.” The second way that the nlanguage differed from English, they were towas where it accented or stressed the word

Participants were given instructions for all foparts of the experiment before they beganwere reminded of the appropriate instructionthey reached each section. The words weresented visually on a computer screen and thesions were tape recorded. During the repetphase of the experiment, the experimentereach word with its correct pronunciation afteappeared on the computer screen. Stressedbles were differentiated from unstressed syllaon this and all training items, most notathrough the use of vowel reduction in unstressyllables. Because the participants and the e

TAB

LA and LB Test Words with th

LA five-syllabletest words

LB five-syllabletest words

S

do ton re mı fa do remi ton fare ton mi fa so re mı´ fa ton somi ton fa sola mi fa so ton lafa ton so lati fa so la ton tiso ton la tı do so lati ton dola ton ti do re la tı do ton reti ton do remi ti do re ton mi

imenter were native English speakers, other mark

l

,-

le

dse-s-nid

lla-s

dr-

ers of stress typical of English such as a shovowel duration in unstressed syllables and a ppeak on stressed vowels were present asAfter seeing and hearing one example in thestruction phase, participants repeated each onext 125 words after the experimenter, while aviewing it on the computer screen. These csisted of all the words demonstrating HEAVY..PENULT (PEN-INITIAL), PENULT (PEN-INI-TIAL) .. LEFT-ALTERNATING (RIGHT-ALTERNATING), and CLASH.. HEAVY ex-cept those to be used in the internal tests. Eathe words was presented to the participants tduring the repetition phase at randomly demined positions.

The rest of the training consisted of a prodtion task, in which participants saw a word oncomputer screen and had to pronounce it. Ifpronounced the word correctly, the experimetold them that they said it correctly and presenthe next word. If they pronounced the word increctly, they were told the correct pronunciatand had to correctly produce it twice before ming to the next word. The words in the productphase consisted of all of the words used inrepetition phase, as well as three words eachtwo of the types used in the internal tests butwere not themselves selected for use in one ointernal tests. As in the repetition phase, parpants saw each word twice at randomly demined positions in the training. All of the wordsboth the repetition and production training phawere presented individually and in a random ofor each subject. None of the words in eit

12

ransitive Stress Patterns Indicated

n-syllable test wordsLA stress pattern

Seven-syllable test wordwith LB stress pattern

re mi ton fa sola do remi ton fa so lami fa ton so lati re mı fa ton so la tı´fa so ton la tı do mi fa so ton la ti do

so la ton ti do re fa sola ton ti do rela ti ton do remi so la ti ton do re mı´ti do ton re mı fa la tı do ton re mi fao re ton mi fa so ti do re ton mi fa so

LE

e T

evewith

doremıfasolatı d

-training phase directly demonstrated the domi-

VY

e-ainmaourdtic-haingn-eehatwo

L er

L -I

ranhech

rna1s

tiondiur

esypemsen), 5efo

s;tesfotehaortestoedrth

rds

inerds.or.ereanysh.sedlla-od-

othe3, aper-

val-60

het isures(Its

1).withimees).ereeri-t re-ningar-

ereothrst

neo kes ite eakersd g orv thes in-f ted,c lan-g ct tot fromE n oft in-


nance relation between the constraints HEAand LEFT-ALTERNATING (RIGHT-ALTER-NATING in LB).

Internal tests.Four internal tests were prsented during the production phase of the tring session to assess learning and to acclithe participants to producing the words withfeedback. Each test consisted of four wopresented visually for production by the paripant. None of the words in the internal testsbeen seen by the participants during trainTwo of the words were of familiar types, meaing that although the participants had not sthe exact words in the internal tests, theyseen words of that type during training, andwere of novel types (LA:sHssss, sssssH;

B: sssssH, sHssss). Each type tested eiththe HEAVY .. PENULT (PEN-INITIAL in

B) domination order or the PENULT (PENNITIAL in LB) .. LEFT-ALTERNATING(RIGHT-ALTERNATING in LB) dominationorder. The internal tests were produced bydomly selecting four words from each of tfour internal test types. One word from eatype was randomly presented in each intetest. Internal tests were placed after the 473rd, 105th, and 138th items in the producphase of the training session. Participantsnot receive feedback on their productions ding the internal tests.

Final test.The participants in both languagwere asked to pronounce the same word tduring the 15-item final test. Five of the itetested for transitivity in both languages (sevsyllable test words; e.g., fa so la ton ti do retested for transitivity in LA only (five-syllabltest words; e.g., re ton mi fa so), and 5 testedtransitivity in LB only (five-syllable test worde.g., do re mi ton fa). Five of seven possiblewords of each type were randomly selectedeach participant, such that each saw 10items (5 five-syllable and 5 seven-syllable) tput the two critical constraints into conflict fthe language they were trained on and 5items that put the two critical constraints inconflict for the language they were not trainon. The final test came directly after the fou
internal test, and participants were asked, simc
-tets

d.

nd

-

lt,

d-

s

-

r

trstt

t

larly to the internal tests, to produce the wowithout any feedback.

Results

The recordings were analyzed to determwhich syllables were stressed in the test woThe main coding was done by the first authSince all of the participants and coders wnative English speakers, coders listened forand all of the markers of stress in EngliThese include vowel reduction in unstressyllables, vowel shortening in unstressed sybles, and pitch peaks on stressed vowels. Cing reliability was fairly good. We did twreliability checks. The first used a subset ofdata from the final test phase of Experimentstudy that had the same test materials as Eximent 1. A second independent coder3 agreedwith the main coder’s assignment of stressues to syllables 93% of the time (based onwords containing a total of 340 syllables). Tsecond check involved data from a study thanot presented here, but used similar procedand the same materials as Experiment 1.findings were similar to those of ExperimentAnother independent second coder agreedthe main coder’s assignments 96% of the t(based on 336 words containing 1904 syllablIncorrect responses during training wtracked by the experimenter during the expmental session. The percentage of correcsponses during the production phase of traiwas 54% overall and 64% during the last quter of the production phase of training.

Responses for the critical test items weach put into one of seven categories for bfive- and seven-syllable test words. The fi

3 All coding, including the reliability checks, was don responses that were blocked by speakers. This maasier to hear the intended stress patterns, because spiffer in speech rate and other factors that impact timinowel reduction. A consequence of this blocking is thatecond coders, even though they were not explicitlyormed of the condition in which each speaker participaan easily hear differences that indicate the traininguage. Thus, the reliability check is not blind with respe

raining language. However, the check on the codingxperiment 3 was blind with respect to the key conditio

hat experiment, whether seven-syllable items were
i-luded or not in the training set.

-r oineoryon

e-s av

anatete

ateeth

Thbere

serinpaev-uldbltere

t thp n-i

ted

ernson

B

ses.be-

Btestsressthatnalr-entord

)

thetran-ghlygnalfs aswasthatonseage

4)

1)

age sincest itemsuse theose items, opposite

re n es.


category,transitive responses,were those responses that perfectly followed the grammathe language the participants had been train. Collapsing over languages, this categcontained 29% of the responses. The seccategory, opposite language transitive rponses,consisted of responses that would h

been transitive responses had the participbeen trained in the other language. This cgory contained 6% of the responses on theitems allowing such responses. The third cgory, CLASH violations,containing 13% of thresponses, consisted of all of the responsesstressed two or more contiguous syllables.next three categories consist of what willreferred to as “template” responses. Thesesponses had stress patterns identical to thowords of the same syllable length seen dutraining. There were three possible stressterns seen during training for five- and for sen-syllable words from which templates cobe derived. For both the five- and seven-syllatest words, there were three possible templaA, B, and C. Intemplate-Aresponses, there awo adjacent unstressed syllables beforeenultimate syllable (in LA) or after the pe

nitial syllable (in LB) (LA: ´ ´ ,´ ´ ´ ; LB: ´ ´ ,

´ ´ ´ ). These patterns accoun

TAB

Mean Percentages of Response Types Pr

Transitive

Template A(critical

nontransitive)Tem

B

LAFive-syllable 36 (28) 32 (36)Seven-syllable 24 (26) 34 (25) 1

LBFive-syllable 24 (28) 16 (16) 1Seven-syllable 32 (19) 24 (32)

a Opposite language Transitive responses are not pothe five-syllable items only test transitivity for one langfor LA had the formssHsss. This cannot be given a tratest item’s “ton” is in the wrong place. Opposite languathat tested transitivity in the opposite language. Sincelanguage transitive responses for five-syllable items a

for 26.5% of the responses.Template-Bre-

fd

d

ets-

st-

ate

-of

gt-

es:

e

sponses had strictly alternating stress pattbeginning on the first syllable and endingthe final syllable in both LA and L(LA: ´ ´ ´ , ´ ´ ´ ´ ; LB:´ ´ ´ , ´ ´ ´ ´ ). Template-Bpatterns occurred on 7.5% of the responThey may have been relatively infrequentcause five-syllable items with the Templatestress pattern were used only in the internaland therefore could be used only to model stpatterns during the later tests to the degreethey were produced correctly during the intertests. Finally,template-Cresponses were chaacterized by alternating stress with two adjacunstressed syllables at either the end of the w(LA) or the beginning of the word (LB(LA: ´ ´ , ´ ´ ´ ; LB:

´ ´ , ´ ´ ´ ). Template-Caccounts for 14.5% of the responses. All oftemplate responses are suboptimal to thesitive response since each violates the hiranked constraint HEAVY in favor of satisfyinlower ranked alignment constraints. The ficategory,Other responses(9%), consisted oany other responses to the critical test itemwell as responses for which the main coderunsure of the stress pattern produced fortoken. The percentages of each type of respare given in Table 13 broken down by langu

13

ced for Five- and Seven-Syllable Test Items (SD)

te TemplateC

Oppositelanguagetransitive

CLASHviolations Other

) 2 (6) NAa 16 (30) 12 (19)17) 8 (10) 2 (6) 10 (17) 10 (1

7) 38 (29) NA 8 (10) 4 (8)10) 10 (14) 0 (0) 18 (20) 10 (1

le on the five-syllable critical test items for a given langue or the other. For example, the five-syllable critical teive pattern for LB regardless of how it is stressed becaresponses were seen on five syllable items only for thponses to those items were not included in Table 13ot shown even though they occurred in both languag

LE

odu

pla

2 (62 (

0 (16 (

ssibuagnsitgeres

and test type.

ityenthe

howredt,irem-ntsverag

, aastivee rwitth

6%

diat-y tt orib-ny

m-reemisose

tthens-r-re

-atst

i-

in

foreakveforoneitheend bye re-antsdgethe

esem.m-eastips

PNl em-o itiver ansi-t ela-t itiver VYaN ef us2 ef sus2 edt llylLI -IH per-c ent-a her on-s ed,p on-


All statistical tests evaluating the transitivhypotheses in this and subsequent experimconsisted of planned comparisons involvingnumber of transitive responses given andoften some other kind of response occurBecause these numbers are not independenused a nonparametric test appropriate for padata, the Wilcoxon. To gain power, we coputed the following tests including participatrained in both languages. We do, howepresent the means broken down by langugroup.

Two tests of transitivity were carried outweak test and a strong test. The weak test wcomparison between the number of transiresponses and opposite language transitivsponses. More test items were producedthe transitive stress pattern (29%) than withopposite language transitive stress pattern (This difference was significant (Z 5 23.64,p ,.05). The weak test shows that participantslearn something from their training and whever they learned increased their tendencgive the correct stress pattern on the testransitivity. Any such tendency cannot be attuted solely to their knowledge of English or aother previous knowledge.

The strong test for transitivity was a coparison between the number of transitivesponses produced and the number of tplate-A responses produced. This comparis crucial because template-A respon(LA: ´ ´ , ´ ´ ´ ; LB:

´ ´ , ´ ´ ´ ) are whawould be expected on the test items ifexplicitly demonstrated constraint relatioare learned, yet LEFT-ALTERNATING(RIGHT-ALTERNATING) was satisfied oveHEAVY, thus violating transitivity. Template-A responses will, therefore, also beferred to as thecritical nontransitive responses.This test did not reach significancethe .05 level for either the five-syllable tewords (transitive5 30%, critical nontranstive 5 24%), Z 5 20.60,p . .05, or for theseven-syllable test words (transitive5 28%,critical nontransitive5 29%), Z 5 20.04,p . .05. Finally, no differences were found
the number of transitive responses produces
ts

.wed

,e

a

e-he).

d

of

--ns

-

by the two language groups,t(18) 5 .21, p ..05.

Discussion

The results of Experiment 1 give supporta transitive constraint system under the wtest of transitivity. Participants were nearly fitimes as likely to give the transitive responsethe language they were trained in than thethey were not trained in. This, combined wthe fact that there were no differences betwthe number of transitive responses produceeach language group, demonstrates that thsponses were due to the training the participreceived rather than their previous knowleof English or some idiosyncrasy of one oflanguages.

The strong test of transitivity, however, donot support a strictly ranked constraint systIn order for a transitive relationship to be deonstrated, participants must show that, at lon a given trial, the trained relationshHEAVY .. PENULT(PEN-INITIAL) and

ENULT(PEN-INITIAL) .. LEFT-ALTER-ATING(RIGHT-ALTERNATING) were fol-

owed. The only responses that correctly dnstrate both of these relations are the transesponses themselves and the critical nontrive responses (template-A). When these rionships were satisfied, however, the transelationship between the constraints HEAnd LEFT-ALTERNATING(RIGHT-ALTER-ATING) was satisfied only 56% of the tim

or five-syllable words (30% transitive vers4% critical nontransitive) and 49% of the tim

or seven-syllable words (28% transitive ver9% critical nontransitive). If participants fail

o apply transitivity, they would be equaikely to produce a form showing HEAVY..EFT-ALTERNATING(RIGHT-ALTERNAT-

NG) and one showing LEFT-ALTERNATNG(RIGHT-ALTERNATING) .. HEAVY.ence, by chance, one would expect theseentages to be 50%, and the obtained percges do not differ significantly from this. Tesults suggest that, when faced with a ctraint conflict that they had not experiencarticipants chose to satisfy each of the c

dtraints an equal number of times.

itynkeciaf

opa

s o

ere

ichorerpeysr tordsnttimpe

se.r 4ecthird

oentesthet 1sitvensuit

a

then beo esw syll -t

fo

i-

notes.

ex-tivensesas

tivee-ll theringofdueaint

n-tion.thehadameng.. In

ofhaveh atheied

oredtiongueer-

ip-hasutenre-ort

andernbe

uc-notV8).


The failure of the strong test of transitivcould indicate that speakers do not use a raconstraint system to determine the pronuntion of novel forms. Alternatively, 30 min otraining may not have been enough to develsolid constraint system. A small experiment wtherefore carried out to assess the effectextended training in the two languages.

Performance under Extended Training

To determine if insufficient learning of thartificial languages was responsible for thesults of Experiment 1, an experiment in whparticipants were trained in the languages fdays before final testing was carried out. Thwere internal tests analogous to those of Eximent 1 after the first, second, and third daThe experiment had a procedure very similathat of Experiment 1. The same training woand constraint systems used in Experimewere used here. The only difference in the suli was the addition of a six-syllable test tyfor each language (LA: ssHssss; LB:sssssHs). Because of the addition of thetest types, the final test contained 25 items

Four participants were each trained fodays. All four participants were 100% corron the internal test that occurred after the tday of training and on the training sessionday 4, thus achieving the goal of excelllearning of the trained stimulus types. Thewords were presented for the first time atend of the fourth session. As in Experimenthere were more transitive (22%) than oppolanguage transitive responses (0%). Ethough there were no opposite language trative responses, this comparison did not qreach statistical significance due to the smnumber of participants,Z 5 21.84,p 5 .07.

The strong tests of transitivity, comparingumber of transitive responses with the numf critical nontransitive (template-A) responsere, as before, unsuccessful for both five-

able test items (transitive5 5%, critical nonransitive 5 40%), Z 5 1.63, p . .05, andseven-syllable test items (transitive5 45%,critical nontransitive5 35%),Z 5 .00,p . .05.The strong test also showed no preference
transitive responses for the new six-syllable tes
d-

asf

-

4er-.

1-

ftt

,eni-ell

r,-

r

items (transitive 5 15%, critical nontranstive 5 35%),Z 5 1.11,p . .05.

Increasing the amount of training thus didincrease the likelihood of transitive responsIn fact, only 22% of the responses in thetended training experiment had the transistress pattern compared to 29% of the respoin Experiment 1. As in Experiment 1, there wno difference between the number of transiand critical nontransitive (template-A) rsponses. On the responses that satisfied aconstraint relationships demonstrated dutraining, transitivity was followed only 37%the time. As mentioned earlier, this could beto the absence of a strictly ranked constrsystem.

However, examination of all of the nontrasitive responses suggests another interpretaIn the extended training experiment, 62% ofresponses on items that test for transitivityidentical stress patterns to words of the ssyllable length that were seen during trainiThat is, they fit one of the template patternsExperiment 1, such patterns comprised 49%the responses. Thus, the participants maydeveloped metrical templates for words witcertain number of syllables (regardless ofCV structure of the syllable) and then applthese templates to the novel forms.

Evidence that stress patterns can be stseparately from a word’s segmental informacan be seen in studies of the tip-of-the-tonphenomenon (Brown & McNeill, 1966; Mey& Bock, 1992; Rubin, 1975) and implicit priming studies (Roelofs & Meyer, 1998). The tof-the-tongue state is one in which a speakeran intense feeling of knowing a word, withobeing able to produce it. Quite often, evthough the word’s segments can not betrieved, the speaker is able to correctly repthe number of syllables the word containsgive evidence of knowing word’s stress pattas well. The basis of this knowledge couldmetrical templates.

Further evidence that a word’s metrical strture may be stored in a form independentonly of segmental information but also of Cstructure is given by Roelofs and Meyer (199
tIn a series of implicit priming experiments,

ithgetithsyureklyairyesetandhentlto

CVst

theet-Vb

t 19)yere

nlyult

owte

theThrdbebleth

te

ro

heput

dput

tteesavwit

telyisingrit-cedro-andf am-ine

r-nd aofout-hethisd inveringanytic-

n-ey

ran-blemstest

atei-

d aeretwoof

nd

useded


Roelofs and Meyer presented participants wDutch word pairs in sets in which the tarwords were either held constant or varied wregard to segmental overlap, the number oflables, main stress position, and CV structThe participants were then required to quicrecall the second word of each of the word pfrom the set they saw. Roelofs and Meshowed that priming due to shared word onoccurred only when the number of syllablestheir stress patterns were held constant. Tinterpreted this as evidence of independestored metrical frames, or what is referredhere as metrical templates. Interestingly,structure did not play a role in priming, suggeing that CV structure is not encoded inmetrical frames. This could explain why a mrical template from a word with a different Cstructure but same number of syllables couldassociated to the test words in ExperimenRecently, Levelt, Roelofs, and Meyer (199amended the proposal of Roelofs and Me(1998). They kept the idea that there are stometrical frames that lack CV structure, but ofor lexical items that do not follow the defastress pattern for the language.

These studies support the idea that knwords are associated with metrical templa(lacking CV structure) and that words usingsame pattern activate a common template.existence of shared templates, across wosuggests that templatic information cantransfered to novel items that match on syllalength. Thus, a speaker knows how to stressnew string fa so la ti because the templa´ ´ is stored.

These stored templates may have beensponsible for the failure of the strong teststransitivity in Experiment 1. Since none of ttraining words in these experimentsHEAVY and LEFT-ALTERNATING (RIGHT-ALTERNATING) into conflict, any five- anseven-syllable training words that did notHEAVY and PENULT (PEN-INITIAL) intoconflict necessarily had the same stress paas that of the critical nontransitive responsThe critical nontransitive responses thus hthe unfair advantage of being associated
frequently used templates, whereas the trans
l-.

srs

yy

-

e.

rd

ns

es,

e

e-f

rn.eh

tive responses, by design, were complenovel patterns and thus lacked templates arfrom training. Therefore, the transitive and cical nontransitive responses may be produby two different processes. The former is pduced by a system of violable constraintsthe later by a system associating words ocertain syllable length to stored metrical teplates. Experiment 2 was carried out to examthis hypothesis.

EXPERIMENT 2

If output forms are determined by two diffeent processes, a ranked constraint system asystem of metrical templates, then if onethese systems can be rendered ineffectual,put forms will have to be determined by tother system. Experiment 2 attempts to doby adding a condition to the procedures useExperiment 1 in which participants are nepresented with seven-syllable words durtraining. Because they never see or hearseven-syllable words during training, the paripants in thisholdout condition will not haveany words of this length in the artificial laguages from which to build templates. Thshould, therefore, show more evidence of tsitivity on the strong test with the seven-syllatest items than with the five-syllable test iteor than the seven-syllable and five-syllableitems for the participants in thefull, or non-holdout, condition.

Method

Participants

Forty native English-speaking undergradustudents from the University of Illinois particpated in exchange for either credit towarcourse requirement or $5. All participants wrandomly and equally assigned to one oflanguage conditions (LA and LB) and onetwo training conditions (holdout condition afull condition).

Stimulus Materials

The same constraint systems and wordsduring the training of Experiment 1 were us
i-here. The main difference between Experiments

di-ouingmin-

atoips

P

LT,

tivethtoen

esnt

blein

estof

on-hafindu

ionastn-oneA

,as

ame

ringer-ro-

ms,bertwo,on-

(full

. nsesg utedt s ofr

ri-m m-p ical


1 and 2 was the addition of the holdout contion within each language. Since the holdcondition eliminated a number of the trainitems, two internal test types held out frotraining in Experiment 1 were used during traing in Experiment 2 so that there would begreat enough variety in the training stimulidemonstrate the critical constraint relationshOne of these demonstrated the HEAVY..

ENULT relationship in LA,sssssH, and theHEAVY .. PEN-INITIAL relationship in LB,sHssss. The other demonstrated the PENU.. LEFT-ALTERNATING relationship in LAsHssss, and the PEN-INITIAL.. RIGHT-ALTERNATING relationship in LB,sssssH.In addition, in order to keep the same relanumber of training types that demonstratetwo critical constraint relationships andmaintain the same number of five- and sevsyllable training types, the LA typessHss andsHsssss and the LB types sssH andssssssH were not used. The training typand the number of tokens used in Experimeare given in Table 14.

To keep Experiments 1 and 2 comparaand to accustom the participants to readthese words without any feedback, internal twere performed in a manner similar to thatExperiment 1. The internal test items, in ctrast to Experiment 1, were training items tparticipants had seen previously. The sametest items used in Experiment 1 were used

TAB

Experiment 2, LA

Language A

CLASH ..HEAVY

HEAVY ..PENULT


sHsHsss (22/38) sssH (21/38) sssss (22/38)ssHsHsss (21/36) sssssH (22/38) sHssss (22/38)

sssHss (23/38) sssssss (28/0)sssssHss (27/0)sssssssH (28/0)

sssHssss (28/0)

Note.Number of tokens in full condition/number of to

ing the final test of Experiment 2. n

t

.

e

-

2

,gs

talr-

Procedure

Experiment 2 was run in an identical fashto Experiment 1 with one exception. In contrto Experiment 1, 10 final test lists were radomly generated, and each was given toparticipant in each of the four conditions (Lfull, LA holdout, LB full, LB holdout). Thusone participant from each of the groups wtested with the same 15 test items in the srandom pattern.

Results

The percentage of correct responses duthe production phase of training was 69% ovall and 72% during the last quarter of the pduction phase of training. For the test itethere were no differences in the overall numof transitive responses produced by thelanguage groups,t (38) 5 0.65,p . .05. Againthere were more transitive responses (full cdition 5 22%; holdout condition5 31%) thanopposite language transitive responsescondition5 3%; holdout condition5 3%), full:Z 5 23.16,p , .05; holdout:Z 5 23.46,p ,05, demonstrating that the transitive respoiven by both language groups can be attrib

o some aspect of the training. Percentageesponse types are given in Table 15.

The major comparison of interest in Expeent 2 was the strong test of transitivity, coaring the transitive responses with the crit

14

LB Training Types

Language B

CLASH ..HEAVY

HEAVY ..PEN-INITIAL

PEN-INITIAL ..RIGHT-

ALTERNATING

ssssHsH (22/38) sHss (21/38) sssss (22/38)ssssHsHs (21/36) sHssss (22/38) sssssH (22/38)

sssHss (23/38) sssssss (28/0)sssHssss (27/0)sHssssss (28/0)

sssssHss (28/0)

ns in holdout condition.

LE

and

ke

ontransitive (template-A) responses. If the hy-

latorentothdi

perestoren-ceofAe

hathe

w eet anc ens0 esw re-s nsit anf 1,w

i en,n edsf e-

:

di-othsed

ute-old-sedde-m 6

s istiveereuton

lla-to

ven-

)27)

0)25)

5)23)

1)24)


pothesis of competing constraint and tempsystems is correct, we would expect to see mevidence of transitivity on the strong test whthere were no training words from whichdevelop templates, as was the case withseven-syllable test items in the holdout contion. However, no more success than in Eximent 1 should be expected for the strong tof transitivity on the five-syllable test items fthe holdout condition or for the five- and sevsyllable test items for the full condition sinwords were available during training in allthese cases from which to build templates.expected, more transitive responses (36%) wproduced for the seven-syllable test items tcritical nontransitive responses (8%) inholdout conditionZ 5 22.37, p , .05. There

as, however, no significant difference betwhe number of transitive responses (20%)ritical nontransitive responses for the sevyllable items in the full condition (14%)Z 5.71,p . .05. The other comparisons of interere between the five-syllable transitiveponses and the five-syllable critical nontraive responses produced in both the holdoutull conditions. Similarly to Experiment

TAB

Percentages of Response Types Produ

Transitive

Template A(critical

nontransitive)Te

LAFull

Five-syllable 32 (37) 24 (31)Seven-syllable 6 (14) 22 (27)

HoldoutFive-syllable 36 (37) 18 (14)Seven-syllable 20 (42) 14 (21)

LBFull



here five-syllable training items with the crit-

ee

e--s

sren

nd-

t

-d

cal nontransitive stress patterns were giveither the holdout or full conditions producignificantly more transitive (holdout5 26%;ull 5 23%) than critical nontransitive rsponses (holdout5 25%; full 5 25%), holdoutZ 5 0.05,p . .05; full: Z 5 .22, p . .05. Theevidence for transitivity in the holdout contion for seven-syllable items came from blanguages. In LA, transitive responses increafrom 6 to 20% from the full to the holdoconditions, while critical nontransitive rsponses decreased from 22 to 14%. In LB, hing out seven-syllable training items increatransitive responses from 34 to 52% andcreased critical nontransitive responses froto 2%.

Discussion

When the number of transitive responsecompared to the number of critical nontransiresponses for the five-syllable test items, thare no differences for either the full or holdoconditions. This might be expected basedExperiment 1 since both groups saw five-syble training words and therefore were ableform templates for these items. When the se

15

d for Five- and Seven-Syllable Test Items (SD)

late TemplateC



(30) 12 (32) NA 10 (17) 2 (6(31) 0 (0) 2 (6) 30 (41) 14 (

(28) 2 (6) NA 16 (18) 8 (1(29) 2 (6) 0 (0) 26 (25) 20 (

(8) 38 (27) NA 2 (6) 16 (2(10) 16 (18) 0 (0) 4 (13) 34 (

(10) 22 (27) NA 10 (14) 12 (2(0) 16 (16) 0 (0) 8 (10) 22 (

LE

ce

mpB

2026

2018

46

80

syllable test items are considered, there was

thesint

iveoshisingomar

thatheteds

susn-ering

s oofont

si-thas. Asigenne

ingageinghe

ishisse

ci-ar

duedrsean

nt 2

the

s sedf LT.N ofw thisr us-i in-d blelI -t ou-b rdsd hisp thePA. nlyfiTsb en-c

P

ono ed-b thep y tob egint tenc -p tain-i Thew ex-a ei thel nlyd Bw thep re


again no difference for the full condition, witransitivity applying on 59% of the responsfor which all of the demonstrated constrarelationships were followed (20% transitversus 14% critical nontransitive). This is mlikely due to the fact that the participants in tgroup saw seven-syllable words during trainHowever, when the seven-syllable items frthe holdout condition are considered, theremany more transitive responses producedcritical nontransitive responses. When all ofconstraint relationships directly demonstraduring training were followed, transitivity waapplied 82% of the time (36% transitive ver8% critical nontransitive). This difference coforms to the expected results since there wno seven-syllable words seen during trainfrom which to develop templates. The resultExperiment 2 thus support the hypothesistwo-process system, one using ranked cstraints and the other associating inputsstored templates.

EXPERIMENT 3

Experiment 2 provided evidence of a trantive constraint system as well as a systemmatches input to previously stored templatethird experiment using the same basic dewas run to replicate the findings of Experim2. Experiment 3 differed from the previous oby eliminating the production phase of trainso participants learned the artificial languonly by hearing and repeating the trainitems. This greatly simplified training from tparticipant’s perspective. Repeating whatheard can be done largely without error. Tnearly error-free training procedure was choin an effort to reduce the likelihood of partipants employing conscious strategies to lethe experimental languages.

Method

Participants

Twenty native English-speaking undergraates from the University of Illinois participatin exchange for either credit toward a courequirement or $5. Five participants were r
domly assigned to each of the four experimentaw
t

.

en

e

fa-

o

t

nt

n

n

-

-

conditions (LA full, LA holdout, LB full, LBholdout).

Stimulus Materials

The same training types used in Experimewere used, as well as the LA typesHss and theLB type sssH. These types demonstrateHEAVY .. PENULT(PEN-INITIAL) relation-hip and were added to partly offset an increarequency of tokens demonstrating the PENU. LEFT-ALTERNATING (RIGHT-ALTER-ATING) constraint relationship. The numberords given during training that demonstrate

elationship were increased after a pilot studyng the procedure employed in Experiment 3icated that participants had slightly more trou

earning the PENULT.. LEFT-ALTERNAT-NG (RIGHT-ALTERNATING) constraint relaionship than the others. To remedy this, we dled the number of times each of the woemonstrating this relationship was shown. Trovided only slightly more demonstrations ofENULT .. LEFT-ALTERNATING (RIGHT-LTERNATING) relationship than the HEAVY. PENULT relationship since there were ove types demonstrating PENULT.. LEFT-AL-ERNATING (RIGHT-ALTERNATING) andix that demonstrated HEAVY.. PENULT. Ta-le 16 shows the training types and their frequies.

rocedure

All materials, including auditory presentatif the instructions, training words, and feack, were presented by a computer. Whenarticipant indicated that she or he was readegin, the experimenter pressed a key to b

he instructions. The direction “please lisarefully to the following instructions” apeared on the screen, and an audio file con

ng the prerecorded instructions began.ords on the screen were replaced by themple word “fa so la ti do” at a point in th

nstructions demonstrating what a word inanguage would look and sound like. The oifference in the instructions for LA and Las the sound file used for demonstration ofronunciation for the example word. The
l ere no differences between the instructions for

theasex

rain

rol,waterdi-in-ndthe

theanthtoerthm. IthutforIf

erifil

enheain

om

ardingualssi-d 2ch

thatntedted

ac-ere

re-ps,i-

m ses(

re--:heagef the

are

ofm-ical


the holdout and non-holdout conditions. Atend of the instructions, the participant wgiven the opportunity to ask questions. Theperimenter then pressed a key to begin the ting phase of the experiment.

Training.To add an extra measure of contone token of each word for both languagesdigitally recorded and stored on the compuAlthough the participants in the holdout contion did not encounter all of the possible traing words, the words that both the holdout anon-holdout conditions did hear had exactlysame pronunciation.

At the beginning of the training session,participants saw on the computer screenheard the computer “say” “please repeatfollowing words.” The participants then hadrepeat the next 224 words. The words wpresented visually for 1000 ms, and thenwords were presented auditorily by the coputer while the word remained on the screenthe participant repeated the word correctly,experimenter pressed a button and the compsaid “correct.” The screen then went blank1000 ms and the next word was presented.word was pronounced incorrectly, the expmenter pressed a key that played an audiosaying “the correct pronunciation is” and ththe word’s sound file was played again. Tparticipants then had to repeat the word agAll training items were presented in rand

TAB

Experiment 3, LA

Language A

CLASH ..HEAVY

HEAVY ..PENULT


sHsHsss (14/14) sHss (14/14) sssss (28/56)ssHsHsss (14/14) sssH (14/14) sHssss (28/56)

sssssH (14/28) sssssss (28/0)sssHss (14/28) sHssssss (28/0)sssssHss (14/0)sssssssH (14/0)

Note.Number of tokens in full condition/number of to

order.

--

s.

de

ee-feer

a-e

.

Test

After the training phase, participants heand saw the words “please read the followwords.” They then were presented with visbut not auditory representations of all 21 poble test words used in Experiments 1 an(Table 6). After the participant produced eaword, the experimenter pressed a keycleared the screen for 1000 ms and presethe next test item. All test items were presenin random order.

Results

Training words were produced with 99%curacy. For the test items, no differences wfound in the overall number of transitivesponses between the two language grout(18) 5 2.71, p . .05. As in the other exper

ents, there were more transitive responfull condition 5 16%; holdout condition5

37%) than opposite language transitivesponses (full condition5 1%; holdout condition 5 1%), full: Z 5 23.11,p , .05; holdoutZ 5 23.55, p , .05, demonstrating that ttransitive responses given by both langugroups can be attributed to some aspect otraining. Percentages of response typesgiven in Table 17.

As in Experiment 2, the major comparisoninterest was the strong test of transitivity, coparing the transitive responses with the crit

16

LB Training Types

Language B

CLASH ..HEAVY

HEAVY ..PEN-INITIAL

PEN-INITIAL ..RIGHT-

ALTERNATING

ssssHsH (14/14) sssH (14/14) sssss (28/56)ssssHsHs (14/14) sHss (14/14) sssssH (28/56)

sHssss (14/28) sssssss (28/0)sssHss (14/28) sssssssH (28/0)sssHssss (14/0)sHssssss (14/0)

ns in holdout condition.

LE

and

ke

nontransitive (template-A) responses. Again,

ceicaou

E er-e ns( fot on( nso ans ican thh ert ifi-c2 ses(0g ncf rs geI to4 n,w sef la-b

f an-s

ins om-p henp ingf itht ld-o s int asl s at sfieda m-o now er,a outc ereg la-t ndn ctlya sus0 n-s in

0)42)

8)13)

)13)

9)26)


more transitive responses (53%) were produfor the seven-syllable test items than critnontransitive responses (0%) in the holdcondition, Z 5 22.82, p , .05. Also as in

xperiment 2, there was no significant diffnce between the number of transitive respo11%) and critical nontransitive responseshe seven-syllable items in the full conditi14%),Z 5 0.0,p . .05. The other comparisof interest were between the five-syllable tritive responses and the five-syllable critontransitive responses produced in botholdout and full conditions. As before, neith

he holdout or full conditions produced signantly more transitive (holdout5 20%; full 51%) than critical nontransitive responholdout 5 26%; full 5 20%), holdout:Z 5.539,p . .05; full: Z 5 20.18,p . .05. To areater degree than in Experiment 2, evide

or transitivity in the holdout condition foeven-syllable items came from both languan LA, transitive responses increased from 90% from the full to the holdout conditiohile critical nontransitive responses decrea

rom 20 to 0%. In LB, holding out seven-syl

TAB

Percentages of Response Types Produ

Transitive

Template A(critical

nontransitive)Te

LAFull



LBFull



le training items increased transitive responset

dlt

esr

-le

e

s.

d

rom 14 to 66% and decreased critical nontritive responses from 9 to 0%.

Discussion

Experiment 3 provides strong evidenceupport of the dual-system hypothesis of ceting template and constraint systems. Warticipants were given words during train

rom which they could derive templates, as whe five-syllable words in both the full and hout conditions and the seven-syllable word

he full condition, participants were justikely to give a transitive response pattern aemplate response pattern when they satill of the constraint relationships directly denstrated during training. When there wereords from which to build templates, howevs with the seven-syllable words in the holdondition, the transitive stress patterns wiven 100% of the time that the constraint re

ionships demonstrated during training, aeeded for the transitive system, were correpplied in the test words (53% transitive ver% critical nontransitive). The evidence of traitivity was stronger in this experiment than

17

d for Five- and Seven-Syllable Test Items (SD)

late TemplateC



(16) 9 (19) NA 17 (31) 14 (2(13) 17 (26) 0 (0) 14 (18) 31 (

(26) 17 (26) NA 11 (12) 39 (1(13) 6 (8) 0 (0) 26 (31) 20 (

(8) 23 (16) NA 29 (20) 9 (8(6) 17 (16) 0 (0) 23 (16) 34 (

(0) 14 (25) NA 20 (22) 11 (1(0) 11 (19) 0 (0) 11 (19) 11 (

LE

ce

mpB

119

269

63

00

she previous one. We believe that the changes in

anlan

erehetiomesysistitoar

ly.tedsorlysyyllvertic, 3dsisr osi

whrehoroesonghmbho

stets.

f aly,ont 1didn-ers tesofain

ow-oseofhe-ro-raintored

oth-en-m aed,hesyl-thationsrdingtran-ressofev-

tedom

acs.s toer,At-urtheat-

en-in

notn’sre-al-

theforo-cesiza-;

hat


the procedures were responsible. Participlearned the sound patterns in the artificialguage simply by listening and repeating.

Although the results of this experiment wstrongly in line with the expectations of tdual-system hypothesis, one may queswhether our participants were learning sothing that corresponds to a linguistic stresstem, as opposed to some sort of nonlingurhythmic pattern. One way to test this isexamine the data of those participants whomost likely to be treating the task linguisticalIf these participants do not show the predicpattern of results, then we would have reafor concern. The participants who most cleatreated the task as that of learning a stresstem were those who reduced stressless sbles, importing this property from their natilanguage, English. The data of the seven paipants who did reduce most clearly (2 LA fullLB full, 0 LA holdout, 2 LB holdout) providestrong support for the dual-system hypotheThere was no difference between the numbeseven-syllable transitive and critical nontrantive responses produced by the participantsreduced in the full condition. Both types weproduced at a rate of 14%. Participants wreduced in the holdout condition, however, pduced 79% of the seven-syllable critical titems with the transitive stress pattern and nwith the critical nontransitive pattern. Althouthese percentages do not reflect a large nuof trials, they show that the speakers wmarked stress most naturally had data consiwith that shown by the full set of participan

GENERAL DISCUSSION

Our three experiments tested the validity okey assumption of Optimality Theory, namethat the relation of dominance between cstraints is subject to transitivity. Experimenshowed that responses reflecting transitivityoccur as a result of learning an artificial laguage. The strong test of transitivity, howevwhich directly compared transitive responsethe alternative critical nontransitive responsdid not yield evidence for the applicationtransitivity. Responses showing that Constr
A .. Constraint C were just as likely as re-
ts-

n--

c

e

n

s-a-

-

.f

-o

-te

er

nt

-

,o,

t

sponses showing that C.. A. Most of theerroneous responses given to test items, hever, exhibited stress patterns identical to thof training words with the same numbersyllables. This led to our dual-system hypotsis: Words are produced by two competing pcesses, one governed by a transitive constsystem and the other consisting of a set of stmetrical templates.

Experiment 2 tested the dual-system hypesis by creating a condition in which sevsyllable test items could not be stressed frotemplate derived during training. As expectevidence of transitivity was then found with tstrong test. Specifically, 82% of the seven-lable test responses in the holdout conditionexhibited the demonstrated constraint relatwere consistent with transitivity. In the thiexperiment, which used a simplified trainprocedure, 100% of these responses weresitive. When items could be assigned stfrom a template derived from training in allthe experiments, there was no unambiguousidence for transitivity.

The idea that language forms are compufrom a grammatical system as well as frsome kind of memory system or lexicon iscommon theoretical claim in psycholinguistiExamples include dual-system approacheregular and irregular morphology (e.g., Pink1991), word reading (e.g., Coltheart, Curtis,kins, & Haller, 1993), and, most relevant to oresearch, Levelt et al.’s (1999) account ofproduction of regular and irregular metrical pterns.

The claim that output can come from a geral rule-like process acting in concert orcompetition with a memory-based system isrestricted to the language domain. Loga(1988) theory of automaticity proposes thatsponses in many tasks arise from either angorithm (analogous to a grammar) or fromretrieval of an instance (a specific memoryprior performance of that specific task). In Lgan’s theory, the algorithm and the instancompete. Similarly, some models of categortion (Nosofsky, Palmeri, & McKinley, 1994Erickson & Kruschke, 1998) hypothesize t
learners both store exemplars and induce rules.

bm

ourtiest isons,ted-

le,

nsondeeasheerricas ohebythbledeuc-th,rode

rvathefren-

vin

gwhngn tialinkyefuuctatich

peri-by

ruc-fre-ableh toveand

, as

emstic

aninen-

is noave

oci-aveh of

ctlytionts.en-

ss”)ec-

t ofanyghnypre-t theictsic-tsdole,any-sso-am-nofithdi-

heof


The two systems compete, but output canbased on computations from both syste(Erickson & Kruschke, 1998).

In our view, the transitive responses instudy arise from the generalization capabiliof the induced grammar, a grammar thadefined by the relative rank or strength of cstraints. The alternative template responsecontrast, reflect the memorization of templaand their application to novel words. Our finings suggest that when a template it availabis often used.

The hypothesized template system is conant with much psycholinguistic theory adata. Evidence for metrical templates has bprovided for adults (Roelofs & Meyer, 1998)well as for children (Gerken, 1991, 1994). Timplicit priming studies of Roelofs and Mey(1998) suggest that stress is stored in metframes independently of the word’s segmenteven CV structure. This is very similar to tproposal of metrical templates being offeredthis study. Gerken (1991, 1994) accounts forwidely observed phenomenon of weak syllaomissions in child language through a moincorporating both rules for generating strtures, and metrical templates representingmost frequent pattern seen in the languagewhich structures can be fit. Gerken (1991) pposes that English-speaking children mayvelop trochaic templates due to their obsetion that this is the most frequent foot type inlanguage. Children’s utterances are thenquently fit to the templates by omission of ustressed syllables of right headed feet, leathe canonical trochaic pattern.

Although this theory fits well with that beinproposed here, the question arises as tochildren omit syllables rather than shiftistress to match the templates, as the adults icurrent study did. One possibility is the initlow ranking of Faithfulness constraintschildren’s constraint systems (Smolens1996a,b). Faithfulness constraints comprisclass of OT constraints that require a faithreproduction of input structures to output strtures. Smolensky proposes that the initial sof the child’s constraint system is one in wh
these constraints are ranked below structura
es

-ins

it

-

n

lr

e

l

eto---

-

g

y

he

,al-e

constraints such as those used in our exments. If the metrical templates proposedGerken (1991, 1994) can be thought of as sttural constraints requiring utterances to fitquently heard patterns in the language, syllomissions would be a reasonable approacsuch a fitting. Adult grammars, however, hathe Faithfulness constraints ranked higherwould thus prohibit such syllable omissionsis indeed the case (for English, at least).

While the hypothesized template systmeshes with much of current psycholinguidata and theory, it remains unclear what roleOT-like constraint system plays during onlproduction or perception. Our finding that trasitive responses are in evidence when thereavailable template tells us that learners hpicked up information that is, at least, assated with the OT constraints and that they hdetermined something of the relative strengtthese information sources.

We speculate that the constraints are direrepresented as weighted sources of activathat bias for particular phonological outpuFor example, any heavy syllable (or some geral symbol or node representing “heavinewould be associated with a very strong conntion weight between it and the assignmenstress. Similarly, the assignment of stress tosyllable would compete, for example, thoumutual inhibitory connections, with that to aadjacent syllable, thus creating a bias thatvents clash. The learning process would seweights of these connections. Direct conflbetween CLASH and HEAVY would be partularly informative, of course. But the weighwould also be influenced by examples thatnot put these into direct conflict. For examplearning that CLASH is more important thsome constraint,X, could strengthen the hpothesized inhibitory connections that are aciated with preventing stress clash. Other exples showing thatX is more important thaHEAVY would then lead to the weakeningthe weights associating heavy syllables wstress. Thus, even if the examples do notrectly place two constraints into conflict, tweight modifications support the learning
ltheir ranking, or transitivity. Beyond these spec-

imss

teswethele,lan

uldarendn-dinarnuldionrnes

onon)tioro-inre

rkenlanren95,

&thynsr &eri-keuclin-

LAn thdic

lysitionsic

everytheared (no

termsress.llus-Ales.

ic)fs rthya

bep

ngm

syl-l N-P

isa icalf

bereds nse-q T-( dicw

fte

nte or as ord,w

cea -F hatg arsee ra,w veo a de-g tresss ed int h canb int.A n oft twoa


ulations, though, we cannot make any claabout the role of the grammar in online proceing. In fact, given the strong effect of templafound in our study and in other studies,believe that the templatic system is by farmajor determinant of behavior in, for examplanguage production by normal speakers ofguages like English.

To see the effect of the grammar, one wohave to look at situations in which templatesrelatively unavailable, as in Experiments 2 a3. However, our use of patently artificial laguages is, of course, less than ideal. Extenour conclusions to natural languages and leing situations would be a challenge. It worequire finding languages, people, or situatin which novel, but grammatical, stress patteneed to be correctly processed. Clearly, thare likely to be found early in the acquisitiprocess (either native or nonnative acquisitiAlso, one may be able to test the assumpusing forms in languages with extremely pductive morphologies, by devising situationswhich novel, or at least rare, outputs arequired.

CONCLUSIONS

OT provides a unique linguistic framewothat has many potential applications to differaspects of language, including the study ofguage learning and processing (e.g., BeEverett, & Shimron, 1998; Hammond, 19Levelt, 1994), linguistic analyses (e.g., ColeKisseberth, 1994; Hammond, 1997; McCar& Prince, 1993), and clinical applicatio(Bernhardt & Stemberger, 1998; StembergeBernhardt, 1997). The results of our expments add to these efforts by testing aassumption of OT and by suggesting how sa grammar might coexist with other psychoguistic mechanisms.

APPENDIX

This appendix details the metrical analysis of thestress system in terms of a ranked set of constraints ostructure and location of metrical feet within the prosoword. The stress system of LB is subject to an anaparallel to the one presented here for LA, with substituof a few mirror-image metrical constraints. The four ba
properties of the LA stress system are 1. The penultimat
s-

-

g-

sse

.n

-

t-t,

yh

e

s

syllable of the word is stressed. 2. Stress is placed onodd-numbered syllable counting from the beginning ofword, excluding the final syllable. 3. Heavy syllablesstressed. 4. Adjacent syllables may not both be stressestress clash).

The stress system can be characterized in metricalas a trochaic, quantity-sensitive system of rhythmic stThe metrical structures that underlie LA stress are itrated in the following example of a word from the Ltraining set, with stress on the first and fourth light syllabMetrical feet are parsed with parentheses notation.

(i) (ss) s (ss)

The basic metrical unit in LA is a left-headed (trochaoot, (ss) or (sH), which results from the following con-traints governing metrical parsing, adapted from McCand Prince (1995).

(ii) PARSE(mora): Every mora of the input mustarsed in the prosodic structure of the output.(iii) FOOT-FORM(trochaic): Feet are left-headed.(iv) BINARITY(moraic): Feet must be binary (parsioras).

The regular occurrence of stress on the penultimateable in LA is accomplished through the ALIGrWd(right) constraint (v).

(v) ALIGN-PrWd(right): Every Prosodic Wordligned at its right edge with the right edge of a metr

oot.

The sequence of alternating stresses on odd-numyllables starting from the left edge of a word is the couence of another alignment constraint, ALIGN-FOOleft) (vi), this time requiring that every foot in the prosoord be aligned at the left edge of the word.

(vi) ALIGN-FOOT(left): Every foot is aligned at its ledge with the left edge of the Prosodic Word.

The effect of ALIGN-FOOT(left), due to its gradievaluation (see McCarthy and Prince, 1995), is to ancheries of alternating stresses at the left edge of the with no unparsed syllables separating the feet.LA words with an odd number of light syllables introduconflict between ALIGN-PrWd(right) and ALIGN

OOT(left). The source of the conflict lies in the fact tiven an odd number of moras it is not possible to pvery mora into a binary foot. Barring deletion of a mohich never occurs in LA, all candidates will either leane mora (and thus one syllable) unparsed or parseenerate, unary foot in such words. Like most natural systems, LA does not tolerate unary feet (a fact reflecthe total absence of adjacent stressed syllables), whice attributed to the highly ranked BINARITY constramong the candidates with only binary feet, the locatio

he unparsed mora is determined by the ranking of thelignment constraints, as shown in Table A1.

e In LA the placement of stress on the penultimate syllable

e od-

he

ies,

a truct on-s anl Bd e“ sc pers theL ls.C icat s im isr lla-b hiles tedT -waww ablep pers ne

t st int eenl ho-n callyl e[ itsc andL ctiveh ofs f thel ng at

ifiedb ng N-P nt ture

es a

ints.esablesint

f , hasa Thes tern,( IN on-s ofi

iss

avorad-

lash

e t

-ndthet thinglly


takes priority over the left alignment of the sequencalternating stresses, modeled by ranking ALIGN-PrW(right) over ALIGN-FOOT(left). This ranking underlies tranking of PENULT.. LEFT-ALTERNATING describedn earlier sections of this paper.

Most of the word types of LA contain heavy syllablnd their analysis in terms of constraints on metrical s

ure will require the introduction of several additional ctraints. First, a few remarks about the nature of heavyight syllables in LA and LB are in order. Both LA and Listinguish the “heavy” CVVC syllable [to:n] from thlight” CVV syllables [do:], [re:], etc. Both syllable typeontain long vowels, at least when viewed from thepective of English, which contrasts the vowels ofA/LB syllabary with a set of short, nondipthongal voweontrastively long vowels are analyzed in phonolog

heory as bimoraic (bearing two weight units), and thuay be argued that the weight distinction in LA and LB

eally a distinction between heavy (bimoraic) CVV syles and the superheavy (trimoraic) CVVC syllable. Wuch a distinction is typologically rare, it is not unatteshe stress system of Hindi, for example, makes a threeeight distinction: CVVC, CVCC. CVV, CVC . CV,ith stress assigned to the rightmost, heaviest syllrovided it is nonfinal (Hayes 1995:162–167). Anotherpective on the LA/LB weight distinction is possible if o

TAB

BINARITY, ALIGN-PrWd(

sssss Z(ss)s(ss)

BINARITYALIGN-PrWd(right)ALIGN-FOOT(left) !

TABLE A2

CLASH, WSP.. ALIGN-PrWd(right)

sHss Z(sH)ssa sH(ss) (sH)(ss)

CLASH *!WSP *!ALIGN-PrWd(right) * *

a There is another metrical structure that can give risthe surface stress pattern ofsHss, and that is (sHs)s, withan initial bisyllabic foot. This foot violates the BINARITY(moraic) constraint because it contains three moras aexcluded from the inventory of trochaic feet underanalyses of Kager (1995) and Hayes (1995). We adopbinarity restriction here, although we note that introducthe possibility of a (HL) trochee would not substantia
alter our analysis of the LA stress system.
f

-

d

-

lt

.y

,-

akes into account the absence of a vowel weight contrahe LA/LB syllabary. In the absence of a contrast betwong and short vowels, it is consistent with standard pological theory to assign a single mora to the (phoneti

ong) vowel in all of the LA/LB syllables. Then, only thto:n] syllable would qualify as bimoraic, on the basis ofoda consonant, which makes the weight contrast in LAB of the standard variety. We adopt the latter perspeere and find support for this view from the tendencyome subjects to reduce the vowel length and quality oight syllables in unstressed positions, thereby producirue CV syllable.

The quantity-sensitive nature of LA stress is exemply words of the typesssH, among others. Without disti-uishing heavy syllables from light, the ranking of ALIGrWd(right).. ALIGN-FOOT(left) would locate stress o

he penultimate syllable, yielding the incorrect strucs (ssH). The LA stress pattern for this sequence requirmetrical structure like (ss)(sH). The incorrect s(ssH)structure can be eliminated on the basis of two constraBINARITY rejects the (ssH) foot on the basis that it parsthree moras rather than two (assuming that heavy syllare bimoraic). The PEAK-PROMINENCE constra(Prince & Smolensky, 1993:39 OT) also rejects the (ssH)oot on the basis that the head element, a light syllable

lower weight prominence than the non-head syllable.tructure that correctly determines the LA stress patss)(sH), satisfies both BINARITY and PEAK-PROM-ENCE. Thus, without appeal to the foot-alignment ctraints the surface patternsssH is selected on the basis

ts optimal metrical structure.Another manifestation of quantity sensitivity in LA

een in forms of the typesHss. Like the sssH type, thistype fails to locate stress on the penultimate syllable in fof stressing the heavy syllable. Stressing the penult indition to the heavy syllable would result in a stress c

A1

t) .. ALIGN-FOOT(left)

)(ss)s (ss)(s)(ss) (ss)(ss)(s)

*! *!*!

o

is

e

TABLE A3

FOOT-FORM(trochaic), BINARITY(moraic).. PARSE

sHss Z(sH)ss (sH)s(s) (sH)(ss)

FOOT-FORM *!BINARITY *!PARSE ** *

LE

righ

(ss

LAtosyl-

, anSS

hat

esst b

d

fo

inconnul

two

h iofT-

he

SHrdica

I-

N-ard

est-ordsvyaliza-iola-tress(i.e.,g theisfies

p-of itsThes farked

arthylign-

shed1.

s to

e liza-t bles,e of


which, as already noted, is not a permissible structure inThe absolute absence of stress clash can be attributedCLASH constraint, which prohibits stressing adjacentlables.

To guarantee that stress will go on the heavy syllablenot on the penult, we appeal to the WEIGHT-TO-STREconstraint (viii) from Prince (1990), the OT constraint tunderlies our HEAVY constraint from earlier sections.

(viii) WEIGHT-TO-STRESS (WSP): Heavy syllablare prominent in foot structure (i.e., heavy syllables muheads).

Ranking ALIGN-PrWd(right) below both CLASH anWSP results in the observed surface stress pattern forsHsswords, as shown in Table A2. This ranking is the basisour earlier ranking HEAVY.. PENULT.

The final two light syllables of (sH)ss are unparsedthe above table. The highly ranked CLASH and WSPstraints succeed in eliminating candidate forms with petimate stress. Other candidates for asHss input includeforms with a final stress, which could arise only underconditions: a final unary foot, (sH) s (s), or a final right-

eaded (iambic) foot, (sH) (s s). These competing cand-dates for ansHss input are eliminated under our analysisthe LA system by ranking the constraints FOOFORM(trochaic) and BINARITY(moraic) above tPARSE constraint, as shown in Table A3.

Other examples illustrating the dominance of CLAand BINARITY over PARSE include the following wotypes from the training set, shown here with their metrstructures:

(ix) Word types illustrating the ranking CLASH, BNARITY .. PARSE

TAB

CLASH .. WSP

sHsHsss Z(sH)sHs(ss)

CLASHWSP *ALIGN-FOOT(left) Optimal

TAB

BINARITY

ssHsHsss (ssH)(sH)s(ss)

BINARITY *!WSP *PARSE *ALIGN-FOOT(left) Optimal

.the

d

e

r

--

l

(ss) (sH) s s

(ss) (sH) s s (ss)(ss) (ss) (sH) s s

(sH) s (ss) s (ss).

The last example also illustrates the role of ALIGFOOT(left), since the medial foot is placed as far leftwas possible, subject to the prohibition on clash.

The data in the training set exhibit some further intering consequences of the prohibition of stress clash. Win LA of the type sHsHsss present two adjacent heasyllables. The ban on adjacent stresses prohibits the retion of stress on both of the heavy syllables. Thus, a vtion of WSP is unavoidable as long as the ban on sclash is upheld and input syllable weight is maintainedno syllable reduction). Among the alternatives, stressinfirst heavy syllable is preferred because it best satALIGN-FOOT(left), as shown in Table A4.

The details of evaluation for ALIGN-FOOT(left) are supressed here, but include an evaluation for each footdistance from the left edge of the prosodic word.optimal candidate is the one in which all the feet are aleft as possible, subject to satisfaction of more highly ranconstraints. The interested reader may consult McCand Prince (1995) for a more detailed discussion of ament theory.

A summary of the ranked constraint grammar establithus far for the LA stress system is presented in Fig. A

There is one word type in the LA training that remainbe discussed, and that is the typessHsHsss. Like the

xample shown in Table A4, CLASH prevents the reaion of stress on the both of the adjacent heavy sylla.g., s(sH)(sH)s(ss). Table A5 shows the evaluation

A4

LIGN-FOOT(left)

(sH)(sH)s(ss) sH(sH)s(ss)

*!*

* *!

A5

. PARSE

Zs(sH)sHs(ss) ssH(sH)s(ss)

* ***** ****

* *!

LE

, A

LE

vs

e toleftth

ld bITY

thiatteith

o tyt

ther th

cn i

e inoras obleavoN-un

ity atedts in

is o

ctsd

mst of

of then ine

c andA nots

n-d onP avest ate3 eavys en 1a N-F ly int eret theta -sF ichc atedi

ondL en

lightt

s er the

.


several clash-avoiding parsed candidates for assHsHsssinput.

The second and third candidates fare equally on ththree constraints in the table and differ only in thealignment of their feet. The first candidate presentssurface stress pattern, but under this evaluation it shoueliminated because it violates the top-ranked BINARconstraint with its parse of an initial trimoraic foot, (ssH).Based on the table above, the optimal candidate forinput should be the second one. The surface stress pfor this input,ssHsHsss, can, however, be generated w-

ut a BINARITY violation once we allow for the possibilihat the (ssH) foot, itself a violation of PEAK-PROMI-NENCE, can be improved by reducing the weight ofheavy syllable. Trochaic shortening is the term used foreduction of a unbalanced trochee to a balanced (ss) tro-hee. The surface stress pattern ofssHsHsss in LA is

consistent with a trochaic shortening analysis, as showTable A6.

Shortening incurs a violation of PARSE(mora), sincshortening the heavy syllable one of its underlying mfails to be parsed. For this reason the first candidate hamore PARSE violation in this table than it shows in TaA5. The second and third candidates are eliminated in fof the first by their greater number of PARSE or ALIGFOOT(left) violations. Since these two constraints areranked with respect to one another, there is an ambiguto which constraint marks the fatal violation, which is noin the table by enclosing fatal marks for both constrainparentheses.

We find evidence for the Trochaic Shortening analys

FIG. A1. Summary of Language A constraint ranking

TAB

Trochaic

ssHsHsss Z(ss)(sH)s(ss)

BINARITYWSP *PARSE **ALIGN-FOOT(left) Optimal

p

ee

srn

e

n

sne

r

-s

f

ssHsHsss input in the observation that many subjeproduced a reduced vowel quality ([ə]) for the unstresseheavy syllable in words such as DO [tən] TON re MI fa.

We turn now to the metrical analysis of the two test itein LA. These input forms can be evaluated by the seconstraints under the ranking established on the basistraining items. The three best competitors are showTable A7 for the first test item,ssHsss. Since these threandidates equally satisfy the FOOT-FORM, CLASH,LIGN-PrWd(right) constraints, these constraints arehown in the table.Candidate 2 fails on BINARITY, known to be an u

ominated constraint in LA. Candidates 1 and 3 tieARSE, though for different reasons: candidate 1 le

wo light syllables out of the metrical parse, while candidleaves one light syllable unparsed and shortens a h

yllable to a light parsed syllable. The decision betwend 3 boils down to the ranking between WSP and ALIGOOT(left). These two constraints do not interact direct

he analysis of any of the training items, and this is whransitivity comes into play. The training items establishwo independent rankings WSP.. ALIGN-PrWd(right)nd ALIGN-PrWd(right).. ALIGN-FOOT(left), the tranitive closure of which yields the ranking WSP.. ALIGN-OOT(left). This transitive ranking is the grounds on whandidate 3, the trochaic shortening candidate, is eliminn favor of candidate 1.

A similar situation obtains in the evaluation of the secA test item,ssssHsss. In fact, the only difference betwe

the two test items is in the presence of two additionalsyllables at the beginning ofssssHsss. These two lighyllables form a perfect trochaic foot, (ss), and since themainder of the string is identical to the first test item,

A6

ortening

s(sH)sHs(ss) ssH(sH)s(ss)

* ****(!)* ***(!)*

*(!) *(!)

TABLE A7

Test Item:ssHsss

ssHsss Zs(sH)s(ss) (ssH)s(ss) (ss)s(ss)

BINARITY *!WSP * *!PARSE ** * **ALIGN-

FOOT(left) * Optimal Optimal

LE

Sh

atioalu-he

rs

.B tic

ro--og-

-ncet

fcon-

B . Sby-

B ”al

C d to

reey-

inses,

eg

).dis-,

hech

nor-

em-l

n’s-

ofers’d

d

D.

-

f

: A

n J.ryck-

lis-

A-

n--

itu-


evaluation of the candidate forms is parallel to the evalushown in Table A7. Table A8 spells out the details of evation for the test itemssssHsss and illustrates once again tole of the transitive ranking WSP.. ALIGN-FOOT(left) inelecting the optimal candidate, candidate 1.

REFERENCES

Archangeli, D., & Langendoen, D. T. (Eds.) (1997).Opti-mality Theory: An overview.Malden, MA: Blackwell

alota, D. A., & Lorch, R. F. (1986). Depth of automaspreading activation: Mediated priming effects in pnunciation but not in lexical decision.Journal of Experimental Psychology: Learning, Memory, and Cnition, 12, 336–345.

Berent, I., Everett, D. L., & Shimron, J. (1998).Do phonological representations specify variables? Evidefrom the obligatory contour principle.[manuscripsubmitted for publication]

Bernhardt, B. H., & Stemberger, J. P. (1998).Handbook ophonological development from the perspective ofstraint-based nonlinear phonology.San Diego: Academic Press.

oysen, S. T., Berntson, G. G., Shreyer, T. A., Quigly, K(1993). Processing of ordinality and transitivitychimpanzees (Pan troglodytes).Journal of Comparitive Psychology,107,208–215.

rown, R., & McNeill, D. (1966). The “tip of the tonguephenomenon.Journal of Verbal Learning and VerbBehavior,5, 325–337.

hristiansen, M., & Chater, N. Special issue devoteconnectionist models of language processing.Cogni-tive Science.[in press]

Clark, H. H. (1969). Influence of language on solving thterm series problems.Journal of Experimental Pschology,82, 205–215.

Cole, J., & Kisseberth, C. W. (1994). An optimal domatheory of harmony.Studies in the Linguistic Scienc24, 101–114.

Collins, A. M., & Quillian, M. R. (1969). Retrieval timfrom semantic memory.Journal of Verbal Learninand Verbal Behavior,8, 240–247.

Coltheart, M., Curtis, B., Atkins, P., & Haller, M. (1993Models of reading aloud: Dual-route and parallel-tributed-processing approaches.Psychological Review100,589–608.

TAB

Test Item

ssssHsss Z(ss)s(sH)s(ss)

BINARITYWSPPARSE **ALIGN-FOOT(left) *

Cutler, A., Mehler, J., Norris, D., & Segui, J. (1983). A

n

-

.

-

language specific comprehension strategy,Nature,304,159–160.

Cutler, A., Mehler, J., Norris, D., & Segui, J. (1986). Tsyllable’s differing role in the segmentation of Frenand English,Journal of Memory and Language,25,385–400.

Davis, H. (1992). Transitive inference in rats (Rattusvegicus). Journal of Comparitive Psychology,106,342–349.

Erickson, M. A., Kruschke, J. K. (1998). Rules and explars in category learning.Journal of ExperimentaPsychology: General,127,107–140.

Gerken, L. A. (1991). The metrical basis for childresubjectless sentences.Journal of Memory and Language,30, 431–451.

Gerken, L. A. (1994). Young children’s representationprosodic phonology: Evidence from English-speakweak syllable productions.Journal of Memory anLanguage,33, 19–38.

Hammond, M. (1995).Syllable parsing in English anFrench.[unpublished manuscript]

Hammond, M. (1997). Optimality theory and prosody. InArchangeli & D. T. Langendoen (Eds.),OptimalityTheory: An overview(pp. 33–58). Malden, MA: Blackwell.

Hayes, B. (1995).Metrical Stress Theory.Chicago: Univ. oChicago Press.

Huttenlocher, J. (1968). Constructing spatial imagesstrategy in reasoning.Psychological Review,75, 550–560.

Kager, R. (1995). The metrical theory of word stress. IGoldsmith (Ed.),Handbook of Phonological Theo(pp. 367–402). Oxford and Cambridge, MA: Blawell.

Levelt, C. C. (1994).On the acquisition of place.Doctoradissertation, Holland Institute of Generative Lingutics.

Levelt, W. J. M., Roelofs, A., & Meyer, A. S. (1999).theory of lexical access in speech production.Behavioral and Brain Sciences,22, 1–75.

Lewicki, P., Hill, T., & Czyzewska, M. (1994). Noncoscious indirect inferences in encoding.Journal of Experimental Psychology: General,123,257–263.

Lipkens, R., Hayes, S. C., & Hayes, L. J. (1993). Long

A8

ssHsss

(ss)(ssH)s(ss) (ss)(ss)s(ss)

*!* *!* **

Optimal Optimal

LE

:ss

dinal study of the development of derived relations in

y,

om

M ign-f

M ys

M ra-gy.

M er-rimy:

M tivey:

M ue

N 4).ing.

O eecgh,.ss.

O sees.

P ve-T.w

P

P .P mic

ticso-

P ry:catnte

P om

R in

nd

R of-g

Ruc-

R . In

:

S isl

S f theU-art-

S ction

S ro-n-

S n thea-,ic

S an-andd

S lity

T ofm-e-

.T m in

.of

T n-by-

V he:

W an-ial

(


an infant.Journal of Experimental Child Psycholog56, 201–239.

Logan, G. D. (1988). Toward an instance theory of autatization.Psychological Review,95, 492–527.

cCarthy, J. J., & Prince, A. S. (1995). Generalized alment. In G. Booij & J. van Marle (Eds.),Yearbook omorphology(pp. 79–153). Dordrecht: Kluwer.

cGonigle, B. O., Chalmers, M. (1977). Are monkelogical?Nature,267,694–696.

cGonigle, B. O., Chalmers, M. (1992). Monkeys aretional! Quarterly Journal of Experimental PsycholoB, Comparative and Physiological Psychology,45B,189–228.

cKoon, G., & Ratcliff, R. (1992). Spreading activation vsus compound cue accounts of priming: Mediated ping revisited. Journal of Experimental PsychologLearning, Memory, and Cognition,18, 1155–1172.

cNamara, T. P. (1992). Theories of priming: I. Associadistance and lag.Journal of Experimental PsychologLearning, Memory, and Cognition,18, 1173–1190.

eyer, A. S., & Bock, K. (1992). The tip-of-the-tongphenomenon: Blocking or partial activation?Memory& Cognition, 20, 715–726.

osofsky, R. M., Palmeri, T. J., & McKinley, S. C. (199Rule-plus-exception model of classification learnPsychological Review,101,53–79.

ller, D. K. (1980). The emergence of the sounds of spin infancy. In G. H. Yeni-Komshian, J. F., Kavana& C. A. Ferguson (Eds.),Child phonology, Vol. 1Production(pp. 93–110). New York: Academic Pre

’Seaghdha, P. G., & Marin, J. W. (1997). Mediatedmantic-phonological priming: Calling distant relativJournal of Memory and Language,36, 226–252.

esetsky, D. (1997). Optimality theory and syntax: Moment and pronunciation. In D. Archangeli & D.Langendoen (Eds.),Optimality Theory: An overvie(pp. 134–170). Malden, MA: Blackwell.

iaget, J. (1967).Six psychological studies.New York:Random House.

inker, S. (1991). Rules of language.Science,253,530–535rince, A. S. (1990).Quantitative consequences of rhyth

organization: Parasession on the syllable in phoneand phonology(pp. 355–98). Chicago Linguistic Sciety.

rince, A. S., & Smolensky, P. (1993). Optimality theoConstraint interaction in generative grammar. Pisaway, NJ: Rutgers Univ. Cognitive Sciences CeTechnical Report 2.

rince, A. S., & Smolensky, P. (1997). Optimality: Frneural networks to universal grammar.Science,275,1604–1610.

oelofs, A., & Meyer, A. S. (1998). Metrical structureplanning the production of spoken words.Journal ofExperimental Psychology: Learning, Memory, aCognition,24, 922–939.

ubin, D. C. (1975). Within word structure in the tip-the-tongue phenomenon.Journal of Verbal Learnin
and Verbal Behavior,14, 392–397. (
-

-

h

-

-r

umelhart, D. E., & McClelland, J. L. (1986).Paralleldistributed processing. Explorations in the microstrture of cognition.Cambridge, MA: MIT Press.

ussell, K. (1997). Optimality Theory and morphologyD. Archangeli & D. T. Langendoen (Eds.),OptimalityTheory: An overview(pp. 102–133). Malden, MABlackwell.

helton, J. R., & Martin, R. C. (1992). How semanticautomatic semantic priming?Journal of ExperimentaPsychology: Learning, Memory, and Cognition,18,1191–1210.

molensky, P. (1996a). The initial state and ‘richness obase’ in Optimality Theory. Technical Report JHCogSci-96-4. Baltimore: Johns Hopkins Univ. Depment of Cognitive Science.

molensky, P. (1996b). On the comprehension/produdilemma in child language.Linguistic Inquiry, 27,720–731.

peas, M. (1997). Optimality Theory and syntax: Null pnouns and control. In D. Archangeli & D. T. Langedoen (Eds.),Optimality Theory: An overview(pp. 171–199). Malden, MA: Blackwell.

tark, R. E. (1980). Stages of speech development ifirst year of life. In G. H. Yeni-Komshian, J. F. Kvanagh, & C. A. Ferguson (Eds.),Child phonologyVol. 1, Production(pp. 73–90). New York: AcademPress.

teirn, J. N., Weaver, J. E., & Zentall, T. R. (1995). Trsitive inference in pigeons: Simplified proceduresa test of value transfer theory.Animal Learning anBehavior,23, 76–82.

temberger, J. P., & Bernhardt, B. H. (1997). OptimaTheory. In M. J. Ball & R. D. Kent (Eds.),The newphonologies: Developments in clinical linguistics(pp.211–245). San Diego: Singular.

esar, B., & Smolensky, P. (1993). The learnabilityOptimality Theory: An algorithm and some basic coplexity results. Boulder, CO: Technical Report, Dpartment of Computer Science, Univ. of Colorado

rabasso, T. (1977). The role of memory as a systemaking transitive inferences? In R. V. Kail & J. WHagen (Eds.),Perspectives on the developmentmemory and cognition.Hillsdale, NJ: Erlbaum.

reichler, F. R., & Tilburg, D. V. (1996). Concurrent coditional discrimination tests of transitive inferencemacaque monkeys: List linking.Journal of Experimental Psychology: Animal Behavior Processes,22, 105–117.

ihman, M. M. (1996).Phonological development. Torigins of language in the child.Cambridge, MABlackwell.

eaver, J. E., Steirn, J. N., & Zentall, T. R. (1997). Trsitive inference in pigeons: Control for differentvalue transfer.Psychonomic Bulletin and Review,4,113–117.

Received February 17, 1999)
Revision received September 1, 1999)

violable constraints in language production: testing the ... · c. transitivity is a crucial...

Documents