Sie sind auf Seite 1von 23

Introducing Optimality Theory Author(s): D. B. Archangeli Source: Annual Review of Anthropology, Vol. 28 (1999), pp.

531-552 Published by: Annual Reviews Stable URL: http://www.jstor.org/stable/223405 Accessed: 28/11/2008 20:58
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=annrevs. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

Annual Reviews is collaborating with JSTOR to digitize, preserve and extend access to Annual Review of Anthropology.

http://www.jstor.org

Annu. Rev. Anthropol. 1999. 28:531-52 Copyright? 1999 by Annual Reviews. All rights reserved

INTRODUCING OPTIMALITY THEORY


D. B. Archangeli
Departmentof Linguistics,Douglass Building 200E, UniversityofArizona, Tucson, Arizona 85721-0028; e-mail: dba@u.arizona.edu

Key Words:

phonology, constraints, universal grammar, linguistics, Tibetan, Tonkawa

* Abstract Optimality theory was introduced in the early 1990s as an alternative model of the organization of naturalhuman language sound systems. This article provides an introduction to the model for the nonlinguist. The basic principles of optimality theory are introduced and explained (GEN, CON, and EVAL). Three important constraint families are explored (Faithfulness, Alignment, and Markedness). Illustrations are provided involving syllabification and vowel harmony in Tibetan and prosodic phonotactics in Tonkawa. The article closes with two general discussions. The first addresses recurring issues in phonological and linguistic analysis and sketches how optimality theory might account for these. The second points out how the explanations arrived at through optimality theory are providing new answers to familiar questions, as well as raising new questions for study.

CONTENTS Introduction ................................................... .................................................. Background The Structure of Optimality Theory ................................ Universal, Yet Violable, Constraints ............................... Faithfulness Constraints...................................... MarkednessConstraints............. .......................... Further Interactions of Faithfulness and Markedness ...................... GEN ............................................................. EVA ............................................................ L A Return to Tibetan ........... ................................. Consequences ..................................................... Other Domains of Explanation ................................... VowelHarmony ....................................................

.......

532 533 534 535 535 536 536 538 538 539 540 . 540 541

0084-6570/99/1015-0531$12.00

531

532

ARCHANGELI VowelHarmonyvia Alignment ........................................ VariableLow Vowelsvia Faithfulness and ConstraintRanking ............. Markednessand Feature Alignment .................................... Empirical Extensions ........................................... Conclusions ...................................................... Consequences ..................................................... Some MarkednessConstraints........................................ Phonological Issues and How OT Addresses Them ................... Phonotactics and Alternations ........................................ Inventories ....................................................... MorphemeStructureConstraints...................................... Contexts(or Environments).......................................... Exceptions ........................................................ ExtralinguisticPhenomena .......................................... OTBeyond Phonology .............................................. ................ New Questions; New Answers .................... ................................................... Representations Nature of Constraints............................................... Nature of Languages ................................................ Evaluation of Analyses .............................................. C onclusion ................................................... 541 542 543 543 546 546 547 547 547 547 548 548 548 549 549 549 549 550 550 550 550

INTRODUCTION
Optimality theory (OT), first introduced in Prince & Smolensky 1991, has swept through formal linguistic research. On encountering the formally simple constraint-based optimality model, phonologists abandoned rules with alacrity to explore the hypothesis that universal constraints alone, rather than a mix of constraints and rules, account for linguistic phenomena. Phonology suddenly emerged as focusing on universals, rather than on the specific language-particular details of rules. Research into phenomena in other domains-morphology, syntax, language acquisition, and change-has also been considered in terms of OT. A wide-spread change in paradigm happened with remarkable speed. In this article, I discuss the new paradigm and show how the model works and how it characterizes the nature of language. I hope that the examples used not only are clarifying but that they provide the reader with a sense of why this model has inspired such enthusiasm. I also aim to prepare the reader to understand more advanced work in the model and perhaps even to feel encouraged to experiment with the optimality paradigm in domains outside of theoretical linguistics. OT is a theory of linguistics. Before delving into the model itself, this review offers a brief and general discussion of formal linguistics. Most work in OT explores phonological phenomena and the interface between morphology and phonology; thus, the discussion focuses primarily on those domains. By necessity, I assume some prior understanding of principles of phonological analysis. The remainder of the article provides examples of the ways OT explains phenom-

THEORY OPTIMALITY

533

ena, reviews some of the nonphonologicalapplicationsof OT, and raises some of the issues OT faces as a theory of grammar.

BACKGROUND
Linguistics, broadly speaking, explores all aspects of naturallanguage. Current linguistic theory takes the Chomskyanview that there is a genetic predisposition in humans to learn language, and that this predisposition results in languages being fundamentally similar (e.g. Chomsky 1957, 1975, 1986; Pinker 1994). Nonetheless, only a little experience with different languages makes it clear that languages are by no means identical to each other. The similarity, then, is at an abstractlevel. Linguists attemptto develop a formalmodel of the linguistic capabilities of humans, a model that characterizesthe fundamentalsimilaritieswhile permittingthe myriadvariationsthatappearin naturallanguage.The termuniverof refersto the innatelinguistic capabilitiesof humans.The grammar sal grammar each specific languageis one of the possibilities permittedby universalgrammar. The linguist attemptsto develop a theory of universal grammarthat clearly and correctly delineates possible and impossible grammars. In the domain of phonology, universal grammaris intendedto account for the regularities and variations found in the linguistic organization of sound. This includes means for expressing the following propertiesof the speaker's knowledge of his/her language:(a) phonotactics,the overall soundpatternof words in a variationsin the form of a morphemein differentsitualanguage;(b) alternations, tions; (c) segment inventories,the segments used by a language,both underlying and surface; (d) morpheme structureconstraints, the sound patterns of morphemes in a language;(e) contexts (or environments),the characterof the situations that condition alternations;and (f) exceptions to any of the above. In the linguistic domains,the ideal model will provide a means of characterizing the fundamentalsimilarities among languages, of expressing the variation that differentiateslanguages, and of predictingtypes of patternsthat should and shouldnot exist. It is always temptingto try extendinga model to include extralinguistic phenomena,propertiesof the linguistic organizationof sound that are not propertiesof the adultgrammar,for example first or second languageacquisition patterns,historical change, and dialectal variation. OT, proposed as a universal grammarfor phonology (Prince & Smolensky 1993, McCarthy& Prince 1993a), meets the challenges above more directlythan any other model of phonology. At the heart of OT is the idea of universal constraints, which are nevertheless violable. By being universal, the constraints themselves provide an explicit means of characterizing cross-linguistic simithe larities that exist. By being violable, there is a means of expressing language variation:the degrees of violation toleratedfor each constraintare unique to each language. OT proposes a single means of expressing which constraintsare violable, namely constraintranking-violations of lower-rankedconstraintsare tolerated in orderto satisfy higher-rankedconstraints.

534

ARCHANGELI

THE STRUCTURE OF OPTIMALITY THEORY


Grammar deceptively simple underOT. At the universallevel, there is a set of is constraintson phonological representations(CON). There is also a means for generatingrelationshipsbetween an actual inputand all potentialoutputs(GEN). Finally, there is a mechanismfor simultaneouslyevaluatingthe potential outputs against the set of rankedconstraintsin orderto select the optimal output for the level, each language needs input in question (EVAL). At the language-particular to identify its set of inputs and an appropriate rankingof CON, called the constrainthierarchy.The informationencoded for a specific language works with universalgrammar (Figure 1). On encounteringan input,GEN createsa candidate between elements in the candidatesand elements in set, showing correspondences the input.EVAL then takes over, using the constrainthierarchyof the languageto select the optimal outputfor that input. Exploringsome aspects of syllabificationwill make this model concrete. Phonologists generally assume that in all languagesthe sounds of words are incorporated into some syllable, the exhaustive parsing hypothesis (see Ito 1986): Consonants(Cs) andvowels (Vs) thatarenot syllabified arenot pronounced.This hypothesis claims thatthe syllable structuredeterminessurfacesequencingof Cs and Vs. For example, a language with CV syllable structureallows strings like CV, CVCV, CVCVCV, etc but does not allow stringslike VCV, CVC, CVCVC, etc. (for other advantagesto syllabification, see Clements & Keyser 1983, Blevins 1994). A general observation about syllable structureuniversals is that although many languages designate CV or CV(C) as the syllable template, sequences like by C, VC, and (C)VC are never so designated.The contrastis characterized four universal constraints (by convention given in small caps): ONSET, a syllableno initial consonant; PEAK,the most sonorous part of the syllable; NOCODA,

KEY of oval: grammar language (LG) box: Universal Grammar circle: derived interaction / by of UG andLG

at Figure 1 A schematicgrammar work.

THEORY OPTIMALITY

535

syllable margins(onsets, codas) contain syllable-final consonant;and COMPLEX, at most one consonant. These four constraintsdo not conflict with each other: every CV syllable satisfies them all, e.g. [mi.co] "mico," [ma.ri.ne] "marina" (the period separates two syllables). Furthermore,these constraints are good candidatesfor being universals, for two reasons. First, togetherthey characterize the most common type of syllable, CV: Every language allows this syllable regardless of the other types of syllables that occur in the language. Second, if a language allows only one type of syllable, that syllable is CV (e.g. Blevins 1994). in However, these constraintsarenot paramount all languages.Numerouslanguages, including English, allow syllables with codas (e.g. [lan] "lawn"),syllables without onsets ([an] "on"),or even syllables without vocalic peaks (e.g. the final sound in [dap,l] "dapple").The existence of such languages demand that these constraintsbe violable.

UNIVERSAL, YET VIOLABLE, CONSTRAINTS This section explores the central OT hypothesis: Universal constraints are violable; they do not hold absolutelytruein all languages.First,two generalfamilies of constraints, faithfulness and markedness, are discussed. These, along with alignment constraints, form the core constraints in OT. Second, the model sketched in Figure 1, showing the effects of GEN and EVAL, is detailed.Finally, a fragmentof the Tibetangrammar, relying primarilyon Dawson (1980) for data, is analyzed.

Faithfulness Constraints
One general propertyof phonological systems is that the input, or mental representation, and the output, or surface representation,are largely and essentially identical. With an input of/fals/ ("false"),we expect (and are rewardedwith) an outputthat is similar, i.e. [falts], ratherthan somethingbearinglittle resemblance or to the input(such as [kaet] [tru]).The similaritiesareexpressedin OT via a famof faithfulness(or correspondence)constraints,constraintsthat requirecorreily spondence between the input and the output.Although this might in principlebe viewed as a symmetricrelation (input and outputare identical), there is substantial evidence supportingasymmetric correspondencerelations. The example of /fals/ X [falts] is illustrative. Every input sound has an output correspondent (f,a,l,s), but there is an outputsound (t) thatdoes not have an inputcorrespondent. Examples illustratingthe opposite asymmetryexist as well. These involve input sounds with no outputcorrespondent,illustratedby the V/0 alternationin the two pronunciationsof "separate,"[sspeart/[s?pret]. These correspondences are formally characterizedby the faithfulness constraints. The class of faithfulness constraints that insists that properties of the inputcorrespondto propertiesof the outputare called MAX (maximize the input)

536

ARCHANGELI constraints.Those demandingthat the outputcorrespondto the inputare dubbed DEP(outputdepends on input)constraints.The MAXand DEPconstraintfamilies are relativized to every type of phonological structure-features, segments, and prosody.1 As shown in particularby the alternate pronunciations [ssperat/[sEpr9t] "separate,"faithfulness is violable. In fact, were faithfulness inviolable, there would be no phonology: input and output would always be identical. However, alternationsabound,in English and elsewhere. Additionaltypes of constraintsare necessary in orderto account for the types of alternationsfound. We explore two additionalconstraintfamilies in this article.The first is the markednessconstraint family, a set of constraintsthat preferunmarkedconfigurations.

MarkednessConstraints
Markednessconstraintseither demandunmarkedconfigurations,such as ONSET and PEAK,or prohibit marked configurations, such as NOCODA COMPLEX and These constraintsresult in a preferencefor CV syllables (Table (defined above). 1). Any other type of syllable violates at least one of these constraints. Table 1 shows that,with respect to this model, English is lenient in its syllable structure: Violations of each of these markednessconstraintsare tolerated.As is well-known, this is not universal. Many languages permit only limited types of syllables (see Blevins 1994 for a typology). UnderOT, these differences arecharacterizedby the relativeimportanceof the variousmarkednessconstraintsandthe different faithfulness constraints [this typology is due to Prince & Smolensky and NOCODA, faith(1993)]. Table 2 illustrateshow differentrankingsof ONSET, fulness result in different syllable types.

FurtherInteractionsof Faithfulnessand Markedness


constraintsof the lanWhen inputstringsdo not conformto the syllable structure the differentrankingsof the two classes of faithfulnessconstraints,MAX guage, and DEP,have different effects. In addition, it is importantto understandhow GEN and EVAL each work. ConsiderTibetansyllabification(Dawson 1980).
1MAX DEPwere introducedin McCarthy& Prince (1995) to replace the faithfulness and and constraintsPARSE FILLfrom earlierwork (Prince & Smolensky 1993, McCarthy& model uses the same Prince 1993a). There are two chief differences. First, the MAX/DEP means to relate inputs and outputsthat it uses to relate reduplicantsand bases. Thus, the MAX/DEP model unifies the formal treatmentof correspondingrelations in phonology, ratherthan requiringone type of relationfor input-output correspondenceand a different model abandonsthe confor base-reduplicant Second, the MAX/DEP type correspondence. tainmenthypothesis,thatevery candidatefor a given inputliterallycontainsthatinput,a hypothesis that led to a variety of very complex mechanismsthat served only to satisfy the hypothesis, as well as failing utterlyto explain metathesis(see McCarthy1995). Underthe MAX/DEP model, relationsbetween inputand outputare indicatedfor every candidate(the convention is to use subscripts;matchingsubscriptsindicatecorrespondingelements).

OPTIMALITYTHEORY TABLE 1 The universalpreference for CV syllables


ONSET PEAK NOCODA COMPLEX

537

CV [ra] "raw" V [ae.pl]"apple" C [ae.pl]"apple" CVC [kaet]"cat" CCV [spa] "spa" VC [aet]"at CCVC [flaep]"flap" CVCC [taesk]"task"
VCC [ask] *
*

* *
*

* *
*

CCVCC [flaesk]"flask"

**

TABLE 2

Syllable typology under optimalitytheorya


ONSET >> FAITH FAITH >> ONSET

NOCODA FAITH >> FAITH NOCODA >>

CV (Hua) CV, CVC (Cairene)

CV, V (Hawaiian) CV, V, CVC, VC (Mokilese)

aMAXand DEPare collapsed together as a single constraintFAITH. Example languages are taken from Blevins 1994.

Tibetan syllables allow single onsets and at most a single coda, which suggests that ONSET, COMPLEX outrank faithfulness, which in turn outranks NOCODA.

When numberscombine to formteens andtens, an "intrusiveconsonant"is found at the boundarybetween the two morphemes(Table 3). The variety of types of intrusiveconsonants argues that they are in fact partof the input representation. The datahere are compatiblewith an analysis in which the intrusiveconsonant is the initial consonantof the second memberof the compound(e.g. /pu/ -> [cu],
"ten"; /ija-pcu/

essence of the analysis is thatthe initial /p/ of/pcu/ survives only when it can syllabify-and it can only do so when it follows a vowel. Undermorphemeconcatenation, the CC sequence survives if a vowel-final morpheme comes first (/ia-pcu/ -- [iapcu]). However, when the CC sequence is word-initial,syllabification fails to rescue /p/, so it does not surface (/pcu/ -> [cu]).

-> [rapcu], "fifty"; but /pcu-i-a/

-- [cuia], "fifty"). The

538

ARCHANGELI TABLE 3 Tibetannumbersshowing the "intrusive"consonantappearing between two numberswhen they are concatenatedwith eath othera Simplex cu cik Isi Gloss Teens (10-#) Gloss Tens (#-10) Gloss

Ten One Four Nine Five cukcik cupsi curqu cuia Eleven Fourteen Nineteen Fifteen r sipcu sipsi qupcu apcu Forty Forty-four Ninety Fifty

qu ia

aSimplex numberssurface as CV or CVC. When numbersconcatenate to form a complex number, a consonant often appearsbetween the two morphemes. In the "tens,"this consonant is always [p], but in the teens, the consonantvaries dependingon the second number([k] with "one,"[p] with "four,"[r] with "nine,"and nothing at all with "five"). Data are from Dawson 1980 and Halle & Clements 1982; the Halle/Clements transcriptionhas been modifid to match that of Dawson.

GEN
In OT analysis of the surfaceformsof/pcu/, (as depictedin Figure 1), the first step is the input,/pcu/, "ten."GEN matchesthis inputwith outputcandidatesby show(via subscripts)between the segmentsof/pcu/ andthe candiing correspondences dates (Figure 2). The mappings shown in Figure 2 are representativeof the kinds of things that GEN can do. GEN can show correspondenceswith an output candidate that is but completely faithful, [plC2U3], it can also show correspondenceswith output candidatesthat are not completely faithful. Differences can occur in a variety of ways: by the featuresof the correspondingelements, [klr233];by the orderof the correspondingelements, [c2u3pl];and by a mismatchin the numberof elements, ['2u3] and [pli'2u3]. Additionally, it is formally possible to have output candirelationto the inputat all, regardlessof the segdates thatbearno correspondence mentalmakeupof the string, [c7u9].In principle,GEN provides correspondences to an infinite numberof candidates;2in practice,linguists try to select the candidates that are closest to the winner and to show how these are eliminated by EVAL.

EVAL
The device used in OT researchto prove thata given constraintrankingsucceeds in selecting the optimaloutputform is called a tableau.The tableauhas a specific
21nher computational model, Heiberg(1999) proposesan algorithmto avoid the problemof GEN producingan infinite set of candidates.

THEORY OPTIMALITY

539

PI23

kr233

Figure 2 The resultof feeding/pcu/ into GEN. form, with constraintslisted across the top and competing candidateslisted down the left side (Tableau 1, online only; also available through http://AnnualReviews.org,ElectronicMaterials).The constraintrankingis depictedby the left-toright orderof constraints,with the leftmost being the most dominant.Where the rankingof constraintsis critical,a solid vertical line separatesthe constraints(e.g. in and between FAITH NOCODA Tableau 1). Where the rankingof constraintsis not critical(i.e. neithercruciallyoutranksthe other),eithera dashedline or no line
at all separates the constraints (e.g. between ONSETand COMPLEX).

Fatal violations are violations of the highest relevantconstraint.In Tableau 1, and these two constraints,then, both candidatessucceed with ONSET COMPLEX; are not relevantfor selecting between the candidates.Candidateb violates FAITH, are a constraintthatcandidatea satisfies. Cells for NOCODA shaded,because success or failure with this constraint is irrelevant:The contest has already been decided.

A Return to Tibetan
It was shown above thatwhen /pcu/ compoundswith /ra/, "five,"the fully faithful Why does the [ijapcu] surfaces because of the subordinatestatus of NOCODA. consonantnot surface every time? Tableau2 (online only; also available through http://AnnualReviews.org,ElectronicMaterials) selects the unfaithful CVviolation. syllable candidate[cu], because of the absence of a COMPLEX The two candidates in Tableau 2 are equivalent with respect to ONSET, but the faithfulcandidatea violates COMPLEX, constraintthatthe unfaithfulcandidateb a satisfies. Here, then, the outputis not identical to the input. There is a further candidate to consider. The COMPLEX violation can be avoided throughepenthesis (the addition of a vowel), giving a form like [pi.cu] (Tableau 3, online only; also available throughhttp://AnnualReviews.org,Electronic Materials). The completely faithful candidatea is eliminatedby COMPLEX. Candidatesb How will the grammarselect and c, however, tie in their violation of FAITH: between them? The answer is that faithfulness is asymmetric. The unattested *[pi.cu] (with vowel epenthesis)violates DEPwhereasthe attested[cu] (with consonant deletion) violates MAX.As Tableau4 (online only; also available through

540

ARCHANGELI http://AnnualReviews.org,Electronic Materials)demonstrates,rankingDEPover MAXidentifies the correctcandidate.3 Were MAXto outrankDEP,the surface patternwould show epenthesis rather thandeletion. This is a desirableresult. Epenthesisoccurs in numerouslanguages (possibly includingEnglish: Epenthesismight be invoked to account for the contrast between past-tense [t/d] in [kist], "kissed," and [fizd], "fizzed," and pasttense [ad] in [fited], "fitted").

Consequences
Universal constraints,constraintviolability, and constraintrankingprovide the explanatorypower of OT, giving the means to explain strong cross-linguistic similarities (universals), cross-linguistic variation (violability), and typological differences (ranking).This section examines the impactof hypothesizinguniversal andviolable constraints,specifically markednessand faithfulnessconstraints. First, universal constraintsprovide a means of characterizingthe similarity among the phonologies of the world's languages. The phonologies are similar because they are expressed in terms of the same set of constraints. Second, the violability hypothesis allows languagesto differ from each other. The degree of violation tolerated for any given constraintis a propertyof each particularlanguage. In this way, OT characterizesthe two key factors of language: the high degree of cross-linguisticsimilaritycombinedwith the possibility of certain types of differences between languages, namely differences attributable to differentrankingsof CON. Finally, ranking differences provide a means of characterizingtypological variationamong the world's languages. As shown in the examples above, different rankingsof the syllable markednessconstraintswith faithfulnessproducethe attestedtypology of languagesbased on surfacesyllable structure. Differentrankings of the two classes of faithfulness constraints,MAX and DEP, divide lanconstraints guages in a differentway, those that avoid violation of higher-ranked
through insertion (MAX > DEP) and those that do so through deletion (DEP > MAX).

OTHER DOMAINS OF EXPLANATION


Extensions to other domains involving prosody, such as prosodic morphology [reduplicationand templaticphenomena(McCarthy& Prince 1993a, 1995; Hendricks 1999)] and metrical phenomena [stress and other foot-based alternations (Hung 1995, Hammond 1997, 1999)] are relatively straightforward, particularly 3Afurther The candidate set [Eu]: challenges [pu]hasanidentical of violations. tieis broken elethat which constraint, CONTINGUITY, demands contiguous by a further correspondence mentsin the inputbe contiguous the output. /pcu/,/p/ and/u/ arenot contiguous In in whereas and/u/ are.Thus,[cu]winsover[pu](see McCarthy Prince1995). & /c/

OPTIMALITYTHEORY

541

when an additional family of constraints is introduced, namely alignment (McCarthy& Prince 1993b). In an effort to provide a broaderunderstandingof how OT addressesdifferenttypes of phonologicalpatterns,vowel harmonyrather than other prosodic phenomenais used. Vowel Harmony The basic property of vowel harmony is that vowels in a sequence surface as similar to each other. Thus, within a single word, vowels might be either [+round] or [-round] [Hungarian(Ringen 1988)], [+back] or [-back] [Turkish (Clements & Sezer 1982)], [+high] or [-high] [Tibetan(Dawson 1980)], [+ATR] (advanced tongue root), or [-ATR] [Pulaar(Paradis 1986, Archangeli & Pulleyblank 1994)]. In Tibetan, nonhigh vowels rase to become high vowels when concatenated with a high-voweled morpheme. Table 4 shows compounds where the second membercontainsa high vowel, and Table 5 shows compoundswith an underlying high vowel in the first member. There are three importantgeneralizations about Tibetan [+high] harmony: [+high] regresses fromrightto left; [+high] progressesfrom left to right;andprogression does not affect long, low vowels. Vowel Harmony via Alignment In orderto accountfor the regressive andprogressivepatternsof height harmony, I use the alignment family of constraints from McCarthy & Prince (1993b). Alignment constraintspreferthatthe edges of two categories,eitherphonological or morphological,match. The patternseen here can be expressedby matchingthe edge of the feature[+high] with the rightand left edges of the word. As with faithfulness, alignment is asymmetric:Separateconstraintsenforce alignment to the TABLE4 Tibetan in The of is [+high]harmony compounds: secondmember thecompound
intrinsically [+high]a N1 Gloss N2 Gloss NJ + N2 Gloss

me JIEl to loo qhaa

Fire Sleep Stone Electricity Snow

siri

Wood Tent Pile Price Mountain

misii pIIlquu tupuu luurii qhaari

Firewood Mosquito net Pile of stones Price of electricity Snow-mountain

qhuu puu rii ri

aThe compound surfaces with high vowels throughout.An affected low vowel surfaces as [a]. Following Dawson 1980, this is a [+high] vowel.

542

ARCHANGELI of TABLE5 Tibetan in The harmony compounds: firstmember thecom[+high] [+high]a poundis intrinsically
N1 Gloss N2 Gloss N1 + N2 Gloss

ri
u

Mountain
Head

tse
Jpee

Tip,peak
Pillow

ritsi
upII

Mountain peak
Pillow

mi qu
u

Person Body
Head

tho loo
ta

List Lungs
Hair

mitu quluu
ute

Listof people Lungs


Hair

ripaa, *ripae

Wildboar

aThese compoundsalso show high vowels throughout;see footnote in Table 2. However, when the second vowel is a long /aa/, this vowel does not raise but surfaces intact as [aa].

right and to the left. Thus, under this hypothesis, Tibetan must have two constraints aligning [+high], one to the right and one to the left. ALHIR, aka right,word, right),where the rightedge of every [+high] aligns to ALIGN([+high], the right edge of some word; and ALHIL,aka ALIGN([+high], left, word, left), where the left edge of every [+high] aligns to the left edge of some word. Tableau 5 demonstratesthe role of ALHIL,and Tableau6 shows ALHIR(tableauxonline Materials). only; also available throughhttp://AnnualReviews.org,Electronic These constraints cover the first two generalizations about Tibetan [+high] harmony: [+high] shows both progressive and regressive harmony. But what about the thirdproperty,the variablebehavior of the long, low vowels?

VariableLow Vowels via Faithfulnessand Constraint Ranking


Low vowels show faithfulnessto theirinputvowel height in a way thatothervowels do not: Long, low vowels areexempt fromthe rightwardalignmentof [+high], althoughthey arenot exempt from leftwardalignment.The faithfulnessto [+low] suggests a constraintMAxLow (every input [+low] vowel has some [+low] output correspondent),forcing the preservationof input [+low]. MAxLow must be ranked between the two alignment constraints so that leftward alignment is oblivious to [+low] specifications and rightwardalignment is sensitive to them (Tableaux7 and 8, online only; also availablethroughhttp://AnnualReviews.org, ElectronicMaterials). Tableau 7 shows that leftward harmony is obtained in [qhoeri] because MAxLo is subordinate to ALHIL. By contrast, because MAxLo dominates ALHIR,rightwardharmonydoes not alter/aa/ in [ripaa](Tableau8). Statedmore

OPTIMALITYTHEORY TABLE 6 Ranking


ALIGN >> MARKEDNESS

543

Harmonytypology via constraintrankinga Effect Examples

Harmony proceeds regardless

Maasai, Turkana, Wolof, Nez

of the type of sound created


MARKEDNESS>> ALIGN Harmony is prevented from

Perce, Chukchee
Pulaar, Yoruba, Wolof,

creatingmarkedfeature combinations

Menomini, Maasai, Turkana, Tibetan

aLanguagesthat fall into both categories are like Tibetan:Harmonyproceeds in two directions and is blocked in one direction but not in the other. Examples other than Tibetan are from Archangeli & Pulleyblank 1994.

generally, harmony is obtained when ALIGN> FAITH;harmony is blocked when FAITH> ALIGN.4

Markedness and Feature Alignment


There are other ways in which harmony might be prohibited, for example through feature markedness. Feature markedness refers to the likelihood (or the unlikelihood) of certain features co-occurring. For instance, vowel height features and tongue root features have a close connection (see Archangeli & Pulleyblank 1994): Tongue root advancement, [+ATR], and [+high] are compatible, as are [+low] and tongue root retraction (or [-ATR]). The opposite combinations are not compatible. These constraints interact with alignment of [ATR] predicting a typology of tongue root retraction harmony (Table 6).

EMPIRICAL EXTENSIONS
Alignment can readily be extended to other empirical domains, such as prosodic phenomena. Alignment of the word edge to a foot edge forces initial or final stress (Hung 1995; Hammond 1997, 1999). Alignment of a reduplicative morpheme to a particular prosodic category forces the reduplicant to take on that prosodic shape (McCarthy & Prince 1993b, 1995). Interactions between alignment and other constraints can force infixation rather than prefixation or suffixation (Hendricks 1999). The list goes on.

4Thedifferencebetween long, low vowels andshort,low vowels is thatthe short,low vowel always harmonizes,even where the long, low vowel fails to. This introducesa furthercomplexity to the constraintfamilies: the need to limit constraintsto a subset. Here, MAXLO must be restrictedto long vowels (MAxLO0,). This constraintoutranksALIGNHIR, which in tur outranksthe generic MAXLO: MAxLot > ALHIR> MAXLO.>

544

ARCHANGELI Cases involving the interaction of faithfulness, alignment, and markedness illustrate the complex patterns that result from the interaction of relatively simple constraints. A case in point is syllabification in Tonkawa. The analysis here assumes the input forms motivated in the work of Kenstowicz & Kisseberth (1979), shown in the leftmost column in Table 7. Of interest is the observation that although Tonkawa roots have three vowels in the input form, there are no surface forms in which all three vowels surface simultaneously. The challenge is to characterize not only the syllable structure but also the pattern of vowel loss in Tonkawa verbs. The first step is to determine which of the syllabification and faithfulness constraints are inviolable and which are violable. To do so, it helps to make further observations about the Tonkawa data. First, surface syllables are either CV or CVC (pi.cen). No syllables start with a vowel, no syllables have a complex margin, and no syllables have a consonantal peak. However, codas are permitted. Second, the only alternation seen is between the presence and absence of vowels. As observed in Kenstowicz & Kisseberth (1979), the quality of the vowels is idiosyncratic. This fact suggests that they are present in the input and that in

TABLE 7 Tankawaverb paradigmsshowing vowel/0 alterationsa Rootb Gloss Cut Active picn-o? we-pcen-o? Castratedone, steer notoxo Hoe ke-pcen-o? picen notx-o? we-ntox-o? ke-ntox-o7 Hoe (noun) netale Lick notox netl-o? we-ntal-o? ke-ntal-o? naxace Make into a fire naxc-o? we-nxac-o? ke-nxac-o? netle-n-o? we-ntale-n-o? ke-ntale-n-o? naxce-n-o? we-nxace-n-o? ke-nxace-n-o? He...it He...them He...me He...it he...them he...me Progressive picna-n-o? we-pcena-n-o? ke-pcena-n-o? Gloss He...it He...them He...me

picena

notxo-n-o? we-ntoxo-n-o? ke-ntoxo-n-o?

He.. it He...them He...me

aThe roots in the active take either a monovocalic CVCC or CCVC shape (e.g. picn- or -pcen-), whereas those in the progressive show up as the bivocalic CVCCV or CCVCV (e.g. picna- or -pcena-). Whetherthe surface form of the root itself startswith CV or with CC depends on whetherit is word-initial(picno?) or follows a prefix (wepceno?), respectively. bInputforms motivated in the work of Kenstowicz & Kisseberth.

OPTIMALITY THEORY

545

Tonkawathere are restrictionson where vowels may surface. To account for the ability of vowels, but not consonants, to delete, it is necessary to do something hinted at above-to separateMAXinto faithfulness constraintstargetingspecific DEPC,and DEPVare types of segments. Here, MAXVis violated whereas MAXC, inviolable (Table 8). From Table 8, we can extract a working hypothesis that the inviolable constraints outrank the violable ones: ONSET,PEAK,MAXC, DEPC, DEPV > MAXV,

Note how faithfulness (MAxC,DEPC,DEPV,MAXV)and markedness NOCODA. interweave in this hierarchy.This intricateinteractionis expected under the OT hypothesis, which allows each language its own individualconstraintranking. This constrainthierarchyfails to select the correct candidatein certaincases, such as [pi.cen] < /picena/, rather than the well-syllabified by unattested [pi.ce.na] (Tableau9, online only; also available throughhttp://AnnualReviews. org,ElectronicMaterials). The tableau also demonstratesthat given an an input like /picena/, there is no way to rankmarkednessand faithfulness constraintsto prefer [pi.cen] over [pi.ce.na]. A faithful string of CV syllables is exactly what these two constraintfamilies prefer. Something is necessary to force vowel loss, violations. incurringMAxV and NOCODA Further examination of the data reveals two critical generalizations. First, all words are consonant-final. Effects of this sort, where two edges match (here the word edge and the consonantedge), are the hallmarkof alignmentconaka straints,alreadyfamiliarin phonology: ALWDCR, ALIGN(word, right, consowhere the right edge of every word aligns to the right edge of some nant, right), consonant. This constraintis violated by words that end with vowels and is not violated by words that end with consonants. As shown in Tableau 10 (online only; also available through http://AnnualReviews.org,Electronic Materials), where a outranksboth MAXVandNOCODA, word-finalconsonantresults, ALIGNWDCR throughvowel loss, violating MAxV. violating NOCODA,

TABLE8 Tonkawa constraint violability


Violability Inviolable Constraint ONSET
PEAK MAXC DEPC DEPV

Example of violation Not applicable

Violable

MAXV NOCODA

pic.no? "He is cutting it" *pi.ce.na.no? pic.no? "He cuts it"

546

ARCHANGELI ALWDCRaccounts for the loss of final vowels, but what accounts for the loss of vowels word-medially?A general tendency in Tonkawais relevant here. The initial syllable tends to be closed (underlined in the following examples): [picno?], "he cuts it"; [wepceno?], "he cuts them";[picnano?],"he is cutting it"; [wepcenano?],"he is cutting them";etc. This suggests a constraintaligning the left edge of a word with the left edge of a heavy syllable: ALWD(G,tL,aka left, ALIGN(word, (Jim,left), where the left edge of every word is aligned with the left edge of some heavy (bimoraic)syllable. The effect of adding ALWD(6,tL to the constrainthierarchyis illustratedin Tableau 11 (online only; also available Materials). throughhttp://AnnualReviews.org,Electronic is In cases where ALWDG(g,L violated, no affixes have been added, e.g. thanthe initial heavy syl[pi.cen], indicatingthatthe final coda is more important lable (not [pic.na]) [shown in Tableau 12 (online only; also available through http://AnnualReviews.org,ElectronicMaterials) by the ranking ALWDCR > ALWD(gtL]. An alternativeone might consider is simply that the language prefers closed right, C, right), where the syllables wherever possible: ALYCR,aka ALIGN(G, right edge of every syllable is aligned with the right edge of some consonant. However, forms like /we-notoxo-n-o?/, "they are hoeing it," show that this ALoCR is not the critical constraint.The candidatethatbest satisfies the putative ALaCRis [wen.tox.no?]. Under the analysis offered here, this latter candidate fails because of excessive violations of MAXVandNOCODA (Tableau 13, online Materials). only; also available throughhttp://AnnualReviews.org,Electronic

Conclusions
The Tonkawa example illustratesthree importantpoints. First, it illustratesthe role thatalignmentmay play in determiningthe phonotacticsof a language.Here, alignmentforces words to end with consonantsand,wherepossible, to begin with heavy syllables. Second, it illustratesthat general tendencies play a significant role in determiningthe structureof a language, even when those tendencies are that In not surface-true. this case, the requirement wordsbegin with a heavy syllable is not always true, despite it being a general tendency. Finally, the Tonkawa example illustratesthe complex patterns(here, patternsof vowel loss) that may arise from the interactionof simple constraints.

Consequences
As shown above, the threemajorclasses of constraintsintroducedhere-faithfulness, markedness, and alignment-interact to characterizethe cross-linguistic typological patternsfound in the world's languages, an exciting result. Furthermore, markednessconstraints,like faithfulness and alignment constraints,may refer to different aspects of the phonological representation.Although study of any one languagewill quickly reveal phenomenathat do not fall readily into one to of these threeclasses of constraints,it is important recognize the significantdegree of coverage possible from this group. It is also importantto realize that

OPTIMALITY THEORY

547

althoughfaithfulnessand alignmenthave generalforms, the formulationofmarkedness constraints is not as rigorously defined. Categorizing the constraints below as markednessconstraintsmakes sense from the point of view of strong cross-linguistic tendencies and/orphonetic properties,but the formal statements are not obviously cohesive.

Some Markedness Constraints


The following are examples of markedness constraints:ONSET,syllables have if onsets; HI/ATR, [+high] then [+ATR]; and ICC[voice], a sequence of consonants must be identical in voicing (see Pulleyblank 1997). Whether analyzing dataor studyingsomeone else's analysis, it is not always obvious when one posits a markednessconstraintthatit is indeed a constraintof that class. This is different from faithfulness and alignmentconstraints.

PHONOLOGICAL ISSUES AND HOW OT ADDRESSES THEM


This section sketches briefly how OT might handle the linguistic organizationof sounds.

Phonotactics and Alternations


Phonotacticsand alternationsgo hand-in-hand underOT. The analysis of the C/0 alternationin Tibetanrelies on the overall syllabificationphonotacticsof the language; the analysis of the V/0 alternationin Tonkawa relies on the interaction between syllable phonotacticsand on two alignmentconstraintsdefining phonotactics of Tonkawa word edges. The [+high]/[-high] alternation in Tibetan resulted from alignment outrankingfaithfulness. In short, the overall patternof sounds in words will follow largely from the highest-rankedconstraintsin that language, and alternationsresult when markednessor alignment outranksfaithfulness. A second result of OT is called the emergence of the unmarked(McCarthy& Prince 1994). The point is that subordinatemarkednessconstraintsnonetheless can make decisions aboutthe optimalcandidate,providedthathigher-ranked constraints fail to decide. In these cases, an unmarkedconfiguration arises, even when that configurationis not centralto the language. Tableau 13 illustratesthis phenomenon.

Inventories
Under OT, all constraintsare in the constrainthierarchy.The inventory of segments possible in the outputof a language are thus the set of segments permitted by the constrainthierarchy,which evaluates outputs. In general, the constraints most responsible are the highest-rankedsegmental markednessconstraints.

548

ARCHANGELI

In constrastto the straightforward way in which OT handles output inventories, OT has no direct means for expressing the inventory of input segments. Because all constraintsin OT are in the constrainthierarchy,no constraintshold over the set of inputs. There are, however, devices proposed in the literatureto provide covert limits on the input. The most popular of these is called lexicon optimization(Prince& Smolensky 1993; see also Ito et al 1995). Lexicon optimization comes into play when each of a varietyof inputsproducesthe same output; lexicon optimizationthen selects among these inputs for the one that incurs the fewest violations of the highest-ranked constraints(see in particular et al 1995 Ito for more on lexicon optimization;see Archangeli & Langendoen 1997b for discussion of alternatives).

Morpheme Structure Constraints


The most naturalway to express restrictionsthat hold morphemeinternallyis to include constraintslimited to particularlexical domains. This approachhas the interestingconsequence of claiming that morphemestructureconstraintsreflect universalsthatmay not hold elsewhere in the language.(Inkelaset al 1997 argues against this approachand in favor of not representingthis type of generalization in the grammarat all).

Contexts (or Environments)


In many cases, contexts for alternationsare irrelevantunder OT: The putative context is handledindependently.For example, Tibetan [+high] spreadis conditioned by the existence of a [+high] vowel in the input.This fact can be characterized by a high-rankedDEPHI,a constraintentirely independentof the [+high] alignment constraints.DEPHIprohibits insertion of [+high], so that only input [+high] induces harmony. Similarly, it is the ranking of MAxLo that prohibits rightwardspreadof [+high], ratherthan including this as a propertyof rightward [+high] alignment. In this way, OT is a theory of conspiracies in the sense of Kisseberth(1970): Surfaceeffects are the resultof the interactionof several independent constraints,not the result of a single, highly complex constraint. empirical Despite these successes, contexts remainone of the least-understood domains within the OT paradigm. Exceptions Despite the broad generalizationsthat we can make about languages, there are nonetheless exceptions. Threegeneralproposalshave been made for dealingwith lexical items that are exempt from partof the constrainthierarchy.First, alternative constraint hierarchies have been set up for exceptional forms, typically treatedas a rerankingof constraintsthat are critical for the general cases (Ito & Mester 1993, Ito et al 1995). Second, exceptional forms are simply lexically specified as such (Inkelas et al 1997). Finally, exceptional behavior is encoded directlyin termsof OT, lexically enteredas a necessary constraintviolation (Golston 1996, Archangeli 1999).

OPTIMALITY THEORY

549

ExtralinguisticPhenomena
A small but growing body of OT work is examining first- and second-language acquisition (Demuth 1995, Smolensky 1996), language change (Jacobs 1995, Zubritskaya1995), etc. At the center of much of this work is the OT hypothesis: Differences will be reflected in different constraint hierarchies. This holds whetherthe differences are between two stages in acquiringa language,between the hierarchies of an adult's first and second languages, between diachronic stages of a single language, and so on. OT provides a clear paradigm-constraint reranking-so that results from extralinguisticdomains influence the development of the linguistic model. Relatedto this are studies of the learnabilityof OT grammars (Tesar 1995, Pul& Turkel 1995) as well as the interpretation leyblank ofpsycholinguistic phonological studies in terms of OT (Frisch 1996; Hammond 1997, 1999).

OT Beyond Phonology
Work in OT syntax can be found as well (Samek-Lodovici 1996, Aissen 1997, Grimshaw 1997, Pesetsky 1997, Speas 1997, Barbosa et al 1998). There is little consensus at this point in OT syntax. OT, being built for phonological analysis, makes critical use of familiarphonological devices-the UR/input, the SR/output. With syntax, the natureof the inputis ill-defined, an issue exploredby Archangeli & Langendoen (1997b): Is it a concept or a set of words? Does it have structure? Depending on the answersto these fundamental questions, otherissues arise. For example, if the inputis simply a stringof words (such as Mico, Marina, and loves), multiple outputs might correspond to a single input: Mico loves Marina;Marinaloves Mico; Mico, Marinaloves; etc). This contrastswith phonology, which traditionallymaps each inputto an unambiguousoutput.

NEW QUESTIONS,NEW ANSWERS


This article focuses on three main goals: to enable the reader to understand mechanically how OT works, to understandhow OT handles different types of phonological data, and to understand the richness of the domains that OT addresses. This concluding section addresses areas where OT challenges us to rethink argumentsand conclusions once thoughtto be irrefutable.

Representations
For the past two decades, much phonological researchhas focused on the nature of the representation. The representation developed from the linear,unstructured advocatedin the late 1960s (Chomsky& Halle 1968) to the highly representations articulatedrepresentations,including word, foot, syllable and moraic structure linked to independent hierarchically arrangedfeatures (see Goldsmith 1994). issues Interestingly,with OT, the focus has largely shifted from representational

550

ARCHANGELI to issues of the natureof constraintsand constraintranking.Phonologists arenow are askingwhetherthe highly articulatedstructures necessaryor whethersimpler, more-linearrepresentations indeed adequate(see Padgett1999, Cole & Kisseare berth 1994).

Nature of Constraints
The original OT proposal is that all constraintsare universaland that each represents a truthaboutlanguagethatis not an absolutetruthabouteach language.This raises new areasof exploration:identifyingpotentialconstraints,motivatingconstraints,andmotivatinguniversality.One attractiveline of researchboth formally and empiricallyexplores whether it is possible to combine simple constraintsto make more-complex constraints(Smolensky 1995). Doing so allows the basic constraintsto remain simple, yet it also allows for language complexity in two ways: constraintinteraction,as discussed in this article, and constraintcompilation, mentionedbriefly here.

Nature of Languages
Each constraintembodies a falsifiable generalizationaboutlanguage. Some such constraints are readily acceptable as generalizations: "sonorants are voiced," "syllables have onsets," "sentenceshave verbs."Othersare harderto understand as linguistically significant generalizations:"Align the right edge of X with the left edge of Y" (alignment);"the left edge of X and the left edge of Y are identical" (anchoring).Constraintsof this sort embody a new type of generalization,a type that was not expressedpriorto the developmentof OT.

Evaluationof Analyses
Prior to OT, linguists had relatively clear evaluation metrics. When comparing two analyses with equal coverage, Occam's razor came into play: The simpler analysis was preferred.This methodis no longer relevant.OT claims that all constraintsare universal. OT claims that the grammarof a particularlanguage is a particularrankingof that set of constraints.There is no means to select between two grammarswith identical empirical coverage. A corollary is the question of whether all grammarsare both equally complex and equally simple, or whether certainconstraintrankingsare preferred.

CONCLUSION
My effort here has been first to provide, without getting too tangledup in technical aspects of the model, sufficient detail abouthow OT works. This was done by focusing on threemajorconstraintfamilies-faithfulness, markedness,andalignment-and workingthroughtheirvaryingroles in threelanguagefragments,ending with a brief discussion of issues facing OT researchers,both in terms of the empiricaldomainsthatOT might addressand in termsof understanding larger implications of the model. For interested readers, Archangeli & Langendoen

OPTIMALITYTHEORY

551

(1997a) provide a deeper introductionto OT, covering phonology, morphology, and syntax. Also, the RutgersOptimalityArchive offers a searchablecollection of cutting-edge works in OT (http://ruccs.rutgers.edu/roa.html).

Visit the Annual Reviews home page at www.AnnualReviews.org.

LITERATURE CITED
Aissen J. 1997. On the syntax of obviation. Language 73:705-50 Archangeli D. 1999. Lexical irregularity in OT. DOT vs. variable constraint ranking. Presented at Annu. Meet. Linguist. Soc. Am., Los Angeles. Archangeli D, Langendoen DT. 1997a. Optimality Theory:An Overview.Maiden, MA: Blackwell Archangeli D, LangendoenDT. 1997b. Afterword. See Archangeli & Langendoen 1997a, pp. 200-15 ArchangeliD, PulleyblankD. 1994. Grounded Phonology. Cambridge,MA: MIT Press Barbosa P, Fox D, HagstromP, McGinnis M, Pesetsky D, eds. 1998. Is the Best Good Enough. Cambridge,MA. MIT Press Blevins J. 1994. The syllable in phonological theory. In The Handbook of Phonological Theory,ed. J Goldsmith,pp. 206-44. Cambridge, MA: Blackwell Chomsky N. 1957. Syntactic Structures. The Hague: Mouton Chomsky N. 1975. Aspects of the Theory of Syntax. Cambridge,MA: MIT Press Chomsky N. 1986. Knowledge of Language: Its Nature, Origins, and Use. New York: Praeger Chomsky N, Halle M. 1968. The Sound PatternofEnglish. New York:Harper& Row Clements GN, Keyser SJ. 1983. CV Phonology: A Generative Theory of the Syllable. Cambridge,MA: MIT Press Clements GN, Sezer E. 1982. Vowel and consonant disharmony in Turkish. In The Structure of Phonological Representations, PartII, ed. H van der Hulst, N Smith, pp. 213-55. Dordrecht:Foris Cole J, Kisseberth C. 1994. An Optimal Domains Theory of Harmony. Cogn. Sci. Tech. Rep. UIUC-BI-CS-94-02. (Lang. Ser.) Univ. Ill., Champaign-Urbana Dawson W. 1980. Tibetan phonology. PhD thesis. Univ. Washington, Seattle. 143 pp. Demuth K. 1995. Markedness and the development of prosodic structure. Proc. NE Lingist. Soc.2:13-25 Frisch S. 1996. Similarity and frequency in phonology. PhD thesis. Northwestern Univ., Chicago, IL. 190 pp. GoldsmithJ, ed. 1994. TheHandbookof Phonological Theory.Cambridge,MA: Blackwell Golston C. 1996. Direct optimality theory: representation as pure markedness. Language 72:713-48 GrimshawJ. 1997. Heads, projectionand optimality. Lingist. Inq. 28:373-22 Halle M, Clements GN. 1982. Problem Book in Phonology. Cambridge,MA: MITPress Hammond M. 1997. Optimality theory and prosody. See Archangeli & Langendoen 1997a, pp. 33-58 Hammond M. 1999. English Phonology. Oxford, UK: Oxford Univ. Press Heiberg A. 1999. Featural optimality theory: a computationalmodel. PhD thesis. Univ. Ariz., Tucson. 318 pp. Hendricks S. 1999. Reduplication without template constraints: a study of bareconsonant reduplication.PhD thesis. Univ. Ariz., Tucson. 283 pp. Hung H. 1995. The rhythmicandprosodic organization of edge constituents. PhD thesis. BrandeisUniv., Waltham,MA. 221 pp. Inkelas S, OrgunO, Zoll C. 1997. The implica-

552

ARCHANGELI

tions of lexical exceptions for the natureof Ariz., pp. 145-83 grammar.In Derivations and Constraints Paradis C. 1986. Phonologie et morphologie in Phonology, ed. I Roca, pp. 393-418. Oxlexicales. les classes nominales en pulaar ford: Clarendon (Fula). PhD thesis. Univ. Montreal,MontIto J. 1986. Syllable theory in prosodic phoreal, Canada.425 pp. nology. PhD thesis. Univ. Mass., Amherst. Pesetsky D. 1997. Optimalitytheory and syn228 pp. tax: movement and pronunciation. See Ito J, Mester A. 1993. Japanese Phonology. Archangeli & Landendoen 1997a, pp. ConstraintDomains and StructurePreser134-70 vation. Tech.Rep.LRC-93-06, Univ. Calif. Pinker S. 1994. The Language Instinct. New York: Morrow Santa Cruz Ito J, Mester A. 1994. Japanese phonology. PrinceA, SmolenskyP. 1991. Optimality.PreSee Goldsmith 1994, pp. 817-38 sented at Ariz. Phonol. Conf. 3, Univ. Ito J, Mester A, PadgettJ. 1995. Licensing and Ariz., Tucson underspecification in optimality theory. PrinceA, Smolensky P. 1993. OptimalityTheLingist. Inq. 26:571-613 ory. ConstraintInteraction in Generative Jacobs H. 1995. Optimalitytheory and sound Grammar.RuCCs Tech. Rep. No. 2, Rutchange. Proc. NE Lingist. Soc. 2:219-32 gers Univ. Center for Cogn. Sci., PiscataKenstowicz M, Kisseberth C. 1979. Generaway, NJ tive Phonology. New York: Academic Pulleyblank D. 1997. Optimality theory and KisseberthC. 1970. On the functionalunity of features. See Archangeli & Langendoen rules.Lingist.Inq. 1:291-306 1997a, pp. 59-101 phonological KisseberthC. 1994. An OptimalDomains The- Pulleyblank D, Turkel W. 1995. Optimality theory and learning algorithms. the repreory of Harmony. Cogn. Sci. Tech. Rep. sentation of recurrent featural asymmeUIUC-BI-CS-94-02, Univ. Ill., Chamtries. Presented at Curr. Trends Phonol., paign-Urbana Abbaye de Royaumont,France McCarthyJ. 1995. Extensions offaithfulness. Rotuman revisited. Tech. Rep. ROA-110, Ringen C. 1988. Transparencyin Hungarian vowel harmony.Phonology 5:327-42 Univ. Mass., Amherst J, Prince A. 1993a. Prosodic mor- Samek-LodoviciV. 1996. Constraintson subMcCarthy jects. an optimalitytheoreticanalysis. PhD phology I: constraintinteractionand satisthesis. RutgersUniv., New Brunswick,NJ faction. Ms., Univ. Mass. Amherst and Smolensky P. 1995. On the internal structure RutgersUniv. of the constraint component CON of UG. McCarthy J, Prince A. 1993b. Generalized Presentedat Linguistics Dep. Colloquium, alignment. In Yearbook of Morphology Univ. Ariz., Tucson 1993, ed. G Booij, J van Marle,pp. 79-153. Dordrecht:Kluwer Smolensky P. 1996. On the comprehension/production dilemma in child lanMcCarthy J, Prince A. 1994. The emergence of the unmarked: optimality in prosodic guage. Lingist. Inq. 27:720-31 morphology. Proc. NE Linguist. Soc. Speas M. 1997. Optimalitytheory and syntax: null pronounsand control. See Archangeli 24:333-79 & Langendoen 1997a, pp. 171-99 McCarthyJ, Prince A. 1995. Faithfulness and reduplicative identity. Univ. Mass. Occ. Tesar B. 1995. Computationaloptimalitytheory. PhD thesis. Univ. Colo., Boulder. 121 Pap. 18:249-384 PP. PadgettJ. 1999. Partialclass behavior and nasal place assimilation. In Proc. SW Optim. ZubritskayaK. 1995. Markednessand sound change in OT. Proc. NE Lingist. Soc. Theory Workshop,Ariz. Phonol. Conf 5. 2:249-64 WorkshopFeatures Optim. Theory, Univ.