Deriving the Feature-Filling/ Feature-Changing Contrast: An Application to Hungarian Harmony Charles Reiss

Thearticle explores an alternative to the interpretive procedure adopted in SPE andproposes a unifiedinterpretive procedure for all . Theproposal solves long-standing problems by makingit unnecessary torefer to athirdvalue of binaryfeatures [ ÀF],to introduce negation intolexical representations (e.g., [ NOT ` rd]),or tointroduce a feature filling/featurechanging diacriticon rules.The article provides a metric forcomparing extensionally equivalent rule systems and argues that themost concise formulation is notalways the correct one, by appeal- ingto crosslinguistic evidence. The proposal is illustratedby applica- tionto the target/ triggerrelations in Hungarian .

Keywords: underspecification,Hungarian vowel harmony, subsump- tion,structural descriptions, phonological rules, conciseness

Unfortunately,within linguistics it has notbeen generally recognized how important such formal, theoretical workis; instead there is a feelingthat too much concern for theoretical detail is a wasteof time.. ..[T]he attitudethat formal, theoretical work is boundto be bothad-hoc and sterile is, I amconvinced, fundamentally mistaken. .. MorrisHalle, ‘ ‘ConfessioGrammatici’ ’

1Distinctness andthe Interpretationof Structural Descriptions Inpractice, as inspectionof any introductory book will show, it has been implicitly assumedin generative phonology that a rulewill apply to any representation that contains a supersetof theinformation contained in the rule’ s structuraldescription (SD). Inother words, if theSD ofaruleR subsumesa representationQ, thenQ isaninput to R. Rulesapply to natural classesof segments, and a naturalclass is represented by a representationthat subsumes the representationof eachof itsmembers.

Thisarticle has benefitedfrom discussion with several audiences,including the Acme Balkanica Conferenceat ConcordiaUniversity in 2001 and the Ohio State UniversityLinguistics Department. Especially helpful individuals have beenAfton Lewis, MarkHale, SylviaBlaho, David Odden, Alain The ´riault,Peter Liem, MorrisHalle, Daniela Isac, and anonymousreviewers. Finally,Pe ´ter Sipta´rdeserves special thanksfor vastly improving the discussion of Hungarian throughcareful readingand helpful criticism.

Linguistic Inquiry, Volume 34, Number 2,Spring 2003 199–224 q 2003 bythe Massachusetts Institute ofTechnology 199 200 CHARLESREISS

Itturns out, however, that this is not theinterpretive procedure developed in Chomskyand Halle1968 (hereafter, SPE),thefoundational work in the field.

(1) Interpretiveprocedure from SPE (p. 337) Aruleof the form A N B / X —— Y appliesto any string Z 4 . . . X¢A¢Y¢ . . ., where X¢, A¢, Y¢ arenot distinct from X, A, Y, respectively;and it converts Z to Z¢ 4 . . . X¢B¢Y¢ . . ., where B¢ containsall specified features of B inaddition to all features of A¢ notspecified in B. Distinctnessis definedas follows: 1

(2) Distinctnessin SPE (p. 336)

Two units U1 and U2 aredistinct if andonly if there is atleast one feature F suchthat

U1 isspecified[ aF] and U2 isspecified [ bF] where a is plus and b is minus . . . The(typically implicit) appeal to subsumptionin generalphonological practice derives from the assumptionof a logicalequivalence between subsumption and nondistinctness, the idea that if x isnondistinctfrom y, then either x subsumes y or y subsumes x. Thisequivalence does not hold, however,except under the working assumption of the SPE erathat representations are fully specifiedfor allfeatures. It is truethat if either x subsumes y or y subsumes x, then x and y are nondistinct,but a simpleexample can illustrate that the converse is not valid if we allowfor partiallyspecified feature matrices in lexical entries. Let x 4 [` round, 1 back]and let y 4 [` round, ` high].The representations x and y do notdisagreewith respect to any features and are thus nondistinct, but oneclearly does not subsume theother. And we clearlydo not expect, say, that x wouldsatisfy an SD specifiedas y. As afurtherexample, consider that by strict application of the SPE interpretiveprocedure, anunderspecified vowel that had only the feature [ 1 round]would satisfy the SD ofarulelike (3),since the representation [ 1 round]is not distinct from therepresentation [ 1 nasal].

(3) [1 nasal] N [1 voiced]

Thisis surelyan undesirable result. Nondistinctrepresentations are, in the general case, what is called‘ ‘consistent’’ inunifica- tion-basedframeworks —thatis, they have no incompatible feature values. But nondistinctness, orconsistency,does not reduce to subsumption. 2 Thepreceding discussion should make it clear thatthe interpretation of SDs ingenerative phonology warrants reexamination.

1 Ihaveomitted reference tononbinary feature values. 2 Ihavefound the same pointmade byBayer andJohnson (1995:sec. 2) in a discussionof Lambek Categorial Grammar: ‘‘Interestingly,in cases where features are fullyspecified, these subsumptionand consistency requirements are equivalent.’’ However,I donot think that the relevance ofthisobservation to the application of phonological rules has beennoted. DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 201

2Feature-CountingEvaluation Metrics Perhapsall that is needed is to reject the SPE interpretiveprocedure in favor of one appealing tosubsumption, since this is whatthe practice has been for thelast several decades. Under this view,any representation subsumed by (containing a supersetof theinformation contained in) a rule’s SDistaken to be a licitinput to the rule. This interpretive procedure has the desirable effectof allowing rules to apply to more than just single representations— they can apply to a naturalclass of representationswhose description is subsumed by the rule’ s SD. Thisinterpretive procedure entails that a rulethat changed feature values —say,from ` F to 1 Ffor somefeature F— wouldapply vacuously to representationsthat are already 1 F before theapplication of the rule. 3 For example,a straightforwardstatement of Polish coda devoicing mightbe writtenas follows: (4) [` cons, 1 son] N [1 voiced]in coda Thisrule applies nonvacuously to [ ` voiced]inputs that are [ ` cons, 1 son];in other words, itmakes them [ 1 voiced].However, according to the subsumption-based convention of rule interpretation,the rulealso applies, albeit vacuously, to [ 1 voiced]inputs that are [ ` cons, 1 son]. Toreiterate, both [ ` voiced]and [ 1 voiced]can satisfy the SD tobe inputs to the rule. Thisinterpretation of SDs isrelated to the SPE feature-countingevaluation metric, the over- archinggoal of which is to minimize redundancy in the grammar, as seen in the Conciseness Conditionformulated by Kenstowicz and Kisseberth (1979). (5) TheConciseness Condition (one component of the SPE evaluationmetric, from Kens- towiczand Kisseberth 1979:336) If thereis more than one possible grammar thatcan be constructed for agivenbody ofdata, choose the grammar thatis most concise in terms of the number of feature specifications. Withhindsight, it is now apparent that the Conciseness Condition is flawed by virtue of its parochialness—inchoosing a modelof grammar,the analyst should take into account the models necessaryto generateother languages as wellas theone in question, and not just choose the most concisegrammar thatcan generate a givencorpus. Thus, we cansee that the Conciseness Condition asstated here is in direct conflict with the search for UniversalGrammar (UG), thegrammar of 4 S0 ,theinitial state of thelanguage faculty.

3 Iadoptwithout argument a binary-valuedfeature system. Thearticle is compatiblewith theories that allow various kindsof underspecification. 4 Animportant question is whetherthe correct formulationof a ruleis necessarily themost concise onethat is consistentwith the data andwith the crosslinguistic (universal) demands discussed here. I arguein Reiss 2002that learnabilityconsiderations provide yet another reason to favor less concise rulesthan are traditionallyposited. 202 CHARLESREISS

(6) Choosingamong extensionally equivalent grammars (Chomsky 1986:38) Becauseevidence from Japanesecan evidently bear on the correctness of atheoryof

S0 ,itcan have indirect —butvery powerful —bearingon the choice of the grammar thatattempts to characterize the I- attained by a speakerof English. Inother words, evidence from onelanguage should bear on thebest analysis of otherlanguages. If twohypotheses, A andB, concerningUG areempirically adequate to provide an explanatory accountof English, but only one of the two —say,A— isadequate to provide an explanatory accountof Japanese,then we shouldselect A asthebest available hypothesis for atheoryof S 0 thatcan lead to acquisition of bothlanguages. Thetraditional interpretation of SDs suchas (4) isnot the only logical possibility. It could havebeen argued that a rulelike Polish devoicing should be formulated so as not to apply vacuously,as in (7). (7) [` cons, 1 son, ` voiced] N [1 voiced]in coda Itseems that the decision to adopt the Conciseness Condition, and thus the rule format of (4), ratherthan (7), was motivatedby theinfluence that engineering approaches to informationtheory hadon thepioneers of generativephonology, an influencethat has been described as leadingto adeadend (Morris Halle, 1975:532 and personal communication). Formulation (4) was seenas themore efficient, and thus better, engineering solution since it was moreconcise than (7). 5 Inthis article, I exploreanother logical possibility for theinterpretation of SDs andshow thatit solveslong-standing problems in phonological theory: the question of howto allow rules totargetunmarked or unspecifiedfeature values and the intimately related issue of the distinction betweenfeature-filling and feature-changing rules. I thenapply the solution to the generation of thealternations seen in Hungarian vowel harmony.

3Subsumption andStructural Descriptions: A Problem TheSD in(4) subsumesvarious possible input representations. Crucially, all inputs must be specifiedfor atleast the features [ ` cons, 1 son].To further understand the nature of the set of representationsthat can serve as inputs to the rule, we needto focuson features that are absent from therule’ s SD. Thetraditional understanding of (4) dependson two distinct interpretations oftheabsence of aspecification: (8) a.Absence of a featurevalue implies that the feature is irrelevant tothe application of the rule.

5 Thebelief that the mind organizes language in a maximallyefficient manner may havealso motivated the Concise- ness Condition.However, it could also have been argued that avoiding vacuous application would have constituted a more efficient solution.Anderson’ s (1985:327)remarks onthe topic are alsotelling: ‘ ‘Earlyconcern for evaluation procedures. ..turnedout to be somethingof a deadend. . ..[T]heappeal of feature countingwent away ...notwith abang,but with a whimper.’’ DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 203

b.Absence of a featurevalue implies that the feature does not need to be mentioned inthe rule, because the rule neutralizes different values for thefeature. Theabsence of referenceto features for placeof articulationin the input of (4),for example, isinterpreted as in (8a) tomean that the rule applies regardless ofthe place of articulation of theinput consonant. In other words, features such as [coronal]and [labial] do not appear in the rulebecause they are irrelevant to its application: Polish devoicing applies to obstruents at all placesof articulation. Assumingthat Polish alternating stops are underlyingly [ ` voiced],6 therule applies vac- uouslyto underlyingly [ 1 voiced]stops and it changes underlyingly [ ` voiced]stops to [1 voiced].Thus, the rule’ s SDcontainsno reference to [voiced] since the rule neutralizes the distinctionbetween [ ` voiced]and [ 1 voiced].This is interpretation (8b). So,some features are absent from theSD becausethe rule does not affect them or depend onthem in any way (8a), and others are absent because the rule neutralizes their two possible values(8b). To reiterate, any representation that is subsumedby theSD ofthe rule satisfies that SD. Thus,(4) canapply to the following inputs: (9) a.representations in which the absent features are irrelevant to rule application and b.representations in which the absent features are neutralized by the rule. Incontrast to the standard view of Polish-typepatterns, consider a patternthat only became knownmuch later in the history of generative phonology, one that requires rules that fill in values onnecessarily underspecified segments. InTurkish, for example,Inkelas (1996; Inkelas and Orgun 1995) argues that there is necessar- ilya three-waycontrast in voicing. Some stem-final stops show a t/d alternation(10a), with [t] appearingin codas and [d] appearingin onsets. Inkelas convincingly argues for anunderlying segmentthat has all the features of acoronalstop, but is unspecified for [voiced].She denotes thisfeature bundle as /D/.Shestates that the segment is assignedthe value [ 1 voiced]in codas and [` voiced]elsewhere. Other stem-final stops consistently surface as [t] andthus are posited tobe / t/underlyingly (10b), and still others consistently surface as [d] andthus are posited to be/d/underlyingly(10c). (10) Turkishvoicing alternations 7 a.Alternating: [ Àvoiced](unmarked for [voiced])/ D/ kanat ‘wing’, kanatlar ‘wing-pl’, kanadgm ‘wing-1sg.poss’ b.Nonalternating voiceless: [ 1 voiced] /t/ sanat ‘art’, sanatlar ‘art-pl’, sanatgm ‘art-1sg.poss’

6 Iwilldo so without argument here. The reader willnotice that if Polishalternating stops are insteadunmarked forvoicing, the problem is to voice them inthe appropriate contexts without targeting the nonalternating voiceless ones. 7 These data havebeen challenged in discussion on the grounds that some Turkishspeakers donot pronounce the (10c)forms witha voicedobstruent in coda position. The irrelevance ofsuch an objection is apparent,as longas the data represent apossiblelanguage. Since Orgun is anativespeaker, I accept thedata as given.The presence ofinflectional 204 CHARLESREISS

c.Nonalternating voiced: [ ` voiced] /d/ etu¨d ‘e´tude’, etu¨dler ‘e´tude-pl’, etu¨du¨m ‘e´tude-1sg.poss’ Therule responsible for making/ D/surfaceas [t] incodas would be identical to (4), but wouldhave to beinterpreteddifferently, since it crucially cannot apply to underlying/ d/.Inother words,the representation of / D/ subsumesthat of / d/(and also that of /t/),butthe rule that affects /D/doesnot affect / d/.Itis necessaryto interpret the absence of [voiced]in theSD asin (11c), whichcompletes the list of interpretationsof absent features under discussion. (11)a. Absence of a featurevalue implies that the feature is irrelevant tothe application of the rule (4 (9a)). b.Absence of a featurevalue implies that the feature does not need to be mentioned inthe rule, because the rule neutralizes different values for thefeature ( 4 (9b)). c.Absence of a featurevalue implies that the feature must beabsentfrom apotential inputrepresentation for therule to apply. Withoutan intelligenthomunculus, a mentalgrammar needsa solutionto theproblem of correctly selectingthe relevant interpretation of anSD.

4EarlierApproaches

Oneway out of thisdilemma would be toallowthe grammar torefer to[ Àvoiced]as a possible specification.

(12) [` cons, 1 son, Àvoiced] N [1 voiced]in coda Below,I providea principledargument against allowing [ Àvoiced]as a possiblespecification. Traditionally,this move has been avoided by most researchers on the intuitive grounds that it representsan overly powerful enrichment of the representational apparatus of phonology. Instead,however, the notational apparatus of rules has been enriched. Typically, a rule label—featurefilling or,equivalently, structurefilling —isused, as in(13), to ensure that such aruleis not(over)applied to fully specified segments and can only apply to provide feature values tounderspecifiedsegments. In the absence of such a label,the correct interpretation is leftto the intelligenceof thereader.

morphologysuggests that the forms shouldbe treated as Turkishand not as French(the language they were borrowed from).I refer thereader tothecited works for further data showingthat this Turkish case is notisolated; Inkelas discusses several cases withthe same logicalstructure. She further shows that these data cannotbe handled by labeling certain morphemes as ‘‘exceptions.’’ Forinstance, she providesexamples ofa singlemorpheme with both an obstruent that alternates invoicing and one that is consistentlyvoiced (even in coda position): edZdat ‘ancestor’, edZdatlar ‘ancestor- pl’, edZdadí ‘ancestor-acc’. Thisexample showsthat failure to devoice obstruents in coda position cannot be a property ofindividualmorphemes, since thestem-medial /d Z/remains voicedalthough it is always incoda position, whereas the stem-final /d/alternates. Instead,the two segments mustbe representationallydistinct: the former is[ ` voiced]and the latter isunderspecified for voicing. DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 205

(13) Featurefilling: [` cons, 1 son] N [1 voiced]in coda Thisis the solution proposed by Inkelas and Orgun (1995:777). I reproducetheir rules in (14). (14) Feature-fillingrules from Inkelas and Orgun 1995:777 a. DEVOICING:Codaplosive N [1 voiced](structure-filling) b. VOICING:Onsetplosive N [` voiced](structure-filling) Theirrule (14a) is basically equivalent to (13), and they propose a secondfeature-filling rule (14b)to provide the alternating stops with [ ` voiced]in onsets. Anotherexample of the featurefilling labelis provided by McCarthy(1994:210), who formu- latesa rulespreading [pharyngeal] from aconsonantto a followingvowel. In addition to an autosegmentalrepresentation of spreading,McCarthy includes the following in the rule statement: ‘‘Condition:Feature-filling.’ ’ He notesthat the ‘ ‘intentof thecondition restricting [the spreading] tofeature-filling is to block the rule from loweringany vowel other than the featureless vowel .’’ Becauseit isfeatureless,consisting of justenough features to identify it asavowel,the representationof schwawill subsume that of every other vowel —lessspecification entails greater generality. Kiparsky’s (1985)discussion of coronalunderspecification briefly notes the problem treated here,stating that it isnecessary‘ ‘towork out someway of referringto unmarkedsegments’ ’ (p.98) whenrepresentations are not fully specified. However, his manner of distinguishing underspecified segmentsis notsatisfactory since he introducesa newdiacritic into representations just in places where,for example,[ ` coronal](or aCoronalclass node) would be specified.Kiparsky proposes that(15a) be the representation of a coronalfricative such as / s/,wherethe x onthe line to the missingnode means ‘ Thereis nospecification on thetier of place features’ . How dowe know(or, moreimportantly, how does the grammar ‘‘know’’) thatthe x doesn’t denoteunderspecification for someother feature?

(15) Representations of /s/ and /f/ (Kiparsky 1985) a. /s/ b. /f/ [1labial] x 1continuant 1continuant 1obstruent 1obstruent

C C

Kiparskyrepresents a noncoronalvoiceless such as / f/as in(15b),where the root node isassociated to a labialplace node. The natural class including both these voiceless /s/and/ f/would presumably be represented by Kiparsky as in (16), unspecified for featureson theplace tier. 206 CHARLESREISS

(16) Representation of all fricatives in Kiparsky’s (1985) system 1continuant 1obstruent

C

Obviously,Kiparsky’ s systemof representing/ s/in (15a) is the equivalent of specifying [ ` coro- nal],since the representation of / s/contains information not present in (16) and can hardly be calledunderspecified. 8 Similaruse of a diacriticdenoting the absence of an association to a givennode can be foundin Archangeli 1988, where a ‘‘melodyunit or anchor’ ’ Zcanbe linked to a featureF via normalassociation lines between Z andF, obligatorilyunlinked to FifZisenclosed in a circle, orambiguouslylinked or unlinkedto F inthe absence of anassociation line or circle.

(17) Linkage notation (adapted from Archangeli 1988); Z is a “melody unit or anchor” a. Z unlinked to F b. Z linked to F

F c. Z ambiguously linked or unlinked to F

Thisnotation presents a problemsimilar to the one facing Kiparsky’ s (1985)system, since the meaningof thecircle around the Z inagivenrule is ‘ unlinkedto thefeature F thatthe rule will provide’. Thus,this notation of ‘ ‘underspecification’’ requiresreference to the very feature whose mentionit is meant to avoid. Inthe next section, I developa unifiedinterpretive procedure for SDs thatvitiates the need for explicit featurefilling or featurechanging diacritics,as well as the need to refer tofeatures thatare absent from arepresentation—Idonot use [ ÀF]asapossiblespecification.

5TheUnified InterpretiveProcedure

For agivenrule R a ,we canrefer toitsstructural description as SDa ,itsstructural change as SCa , andits environment as Enva ,givingus thesimple rule schema in (18).

8 Kiparsky’s suggestionleads toother problems as well: doesthe line in (15a) block spreading? DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 207

(18) Rule schema9

Ra : SDa N SCa in Enva We willemploy Greek letter variables in the usual way: a [ $` , 1 %. Ialsoassume that something like Chomsky’ s (1967)Principle 6 isvalid:‘ ‘Twosuccessive linesof aderivationcan differ by at mostone featurespecification.’ ’1 0 Thisprinciple helps restrict thenotion of possiblephonological rule, and thus it is desirable to conformto it.For ourpurposes, thisprinciple means that SC a willalways contain a singlefeature specification, ` F, 1 F, or aF.We cannow formulate the unified interpretive procedure (UIP) for structuraldescriptions, substituting aF for SCa . (19) Unifiedinterpretive procedure for structural descriptions

ArepresentationQ isaninput to a ruleR a

SDa N aF in Enva

ifand only if SD a subsumesQ andone of thefollowing holds:

a. 1 aF [ SDa (SDa andthus each Q thatsatisfies SD a isspecified 1 aF) or

b. 1 aF [/Q(noQ thatsatisfies SD a isspecified 1 aF,andthus SD a isnot specified 1 aF, either).

Firstconsider (19a). Since (19a) requires that SD a bespecified 1 aF,itfollows that every representationQ subsumedby SD a bethusspecified. Since SD a mustsubsume every input to the rule,each input must also be specified 1 aF.ArepresentationQ thatsatisfies this condition will undergofeature changing to aF. Now consider(19b). The requirement of (19b)is that Q notbe specified 1 aF,soit can be eitherspecified aF or notspecified at all for featureF. SinceQ isrequired not to be specified

1 aF,anyQ thatsatisfies this condition will not be subsumed by an SD a that is 1 aF. In other words,if Qisnot 1 aF,andQ isan input to R a , then SDa is also not 1 aF.(We thussee that (19a)and (19b) are mutually exclusive: they cannot be satisfied simultaneously.) If condition (19b)is fulfilled,the rule will either fill in the value aForvacuously‘ ‘change’’ aF to aF. Thetwo conditions thus require either (a) Qis 1 aFor(b) Qisnot 1 aF.(Furtherconditions areimposed by Env a ,ofcourse.) The existence of underspecificationmeans that not 1 aF does not mean ‘aF’,butinstead means ‘ either aForunspecifiedfor F’.

Thus, if SCa is 1 F, and SDa isspecified ` F,thenan input to R a mustcontain ` F in order tosatisfy SD a bycondition (a). However,if SC a is 1 F but SDa isnot specified ` F, then an

9 Morediscussion is requiredfor deletion, insertion, and metathesis rules.These problemsare addressedin work inprogress. Also, the environment Env is, strictly speaking, part of theSD, butI willtreat them separately forthe sake of clarity. 10 Thisnotion can beadapted for more recent theoriesof representation.I donot make use of,for example, feature geometryin thisarticle, andI haveargued elsewhere (Reiss, toappear a,b)that feature geometryis nota necessary or desirablepart of phonological theory. 208 CHARLESREISS input to Ra maynot contain ` F.Itmay be specified 1 F (Ra willapply vacuously in thiscase) oritmay be unspecified for F(R a will fill in 1 Finthis case) and thus satisfy SD a bycondition (b).

Similarly,we canswitch all the signs. If SC a isspecified ` F and SDa isspecified 1 F, then an input to Ra mustcontain 1 Finorder to satisfy SD a bycondition (a). However,if SC a is specified ` F, but SDa isnot specified 1 F,thenan input to R a maynot contain 1 F, but it may bespecified ` F (Ra willapply vacuously in this case) or it may be unspecified for F(R a will fill in ` Finthis case) and thus satisfy SD a bycondition (b). So,(19a) corresponds to traditional feature-changing rules and (19b) to traditional feature- fillingrules. However, the UIP precludesthe necessity of referringto unmarked values such as [Àvoiced].The crucial advance we havemade is this:instead of having‘ ‘towork out some way ofreferring to unmarked segments,’ ’ we havea wayto ensure that they are treated as a class withrepresentations that are vacuously affected by rules. Ina languagelike Polish, which has a two-wayvoiced/ voicelesscontrast in obstruents, we mightbe tempted to retain the traditional formulation of the devoicing rule and the traditional interpretationof SDs. We wouldstill generate Polish-type output. However, a trulyexplanatory approachto phonology allows us to see that Turkish can tell us something about Polish: the correctformulation of therule must be somethingcloser to (7) thanto (4). Since we areinterested inUG, we areinterested in a singleinterpretive procedure for allgrammars. This was thepoint ofthequotation from Chomsky1986 in (6). Inthe particular case of the representation of Polish devoicing, UG shouldbe assumed to usethe same interpretive procedure as isusedin Turkish.More concise rules can be written for justthe Polish data, but they would not be rules of human phonology, if the UIP iscorrect.Since theinterpretive procedure for alllanguages is assumed to be identical, but the patterns to be accountedfor aredifferent, the rules themselves must differ as well. Polish uses (7), whereas Turkishuses (4). Inother words, the traditional account of Polish devoicing using the subsumption-based interpretiveprocedure would be extensionally equivalent to theaccount proposed here (using (7) andthe UIP), butwe nowcan choose between them in a principledfashion. Again, this is the typeof argumentationsuggested by (6). Afurtherimplication of theUIP isthat we nowderive the intuitively valid result that rules donottreat representations that are ` Fandrepresentations that are 1 Fasanaturalclass to the exclusionof representations that are unmarked for F.Withrespect to theTurkish data discussed above,this means that, for example,/ t/and / D/constitutea naturalclass (they are not [` voiced]), and/ d/and / D/doso as well (they are not [1 voiced]),but / t/and / d/do not, to the exclusion of /D/. For thesake of explicitness,let me reiterate the difference between the simple subsumption- basedinterpretation of SDs andthat given by theUIP. Arulelike (4) shouldapply to [ ` voiced] stopsaccording to the traditional interpretive procedure. The SD of(4) isgiven as [ ` cons, 1 son], andsince the representations of, say, / t/,/D/,and/ d/are subsumed by [ ` cons, 1 son],the rule DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 209 shouldapply to all three. But this would not let us distinguish / d/from /D/,whichwe needto dofor Turkish. However,using the UIP, arulelike (4), where the SC is[ 1 voiced],cannot apply to [` voiced]representations like / d/.

· (19a)is notsatisfied since [ ` voiced]is notin the SD oftherule as stated. · By(19b), any Q thatis aninput to the rule cannot be specified as [ ` voiced].

Sinceneither condition is satisfied, the rule cannot apply to [ ` voiced]representations. However, both/ t/and / D/satisfycondition (19a), since / t/is [ 1 voiced],and the rule applies vacuously; and/ D/satisfiescondition (19b), since / D/is not [1 voiced],and thus the rule fills in thisvalue. Thiswould work perfectly for ourmodel of Turkish. Considernow the rule in (7), repeated here:

(20) [` cons, 1 son, ` voiced] N [1 voiced]in coda

Sincethe SD ofthe rule contains [ ` voiced],the rule can obviously apply to [ ` voiced]representa- tions.It cannot apply to representations that are either specified [ 1 voiced]or unspecified for [voiced]—the SD doesnot subsume such representations. This would work perfectly for our modelof Polish.

6Discussion

Areviewerof anabstractof thisarticle objected to theUIP onthe grounds that it isanextremely powerfuldevice, since it is meant to be relevant to all rules in all languages. This objection reflectsa misunderstandingof the notion of powerin theory construction. In fact, a singleinterpre- tiveprocedure that holds for allrules in all languages provides a less powerful(and thus better) modelthan one that allows various devices on anadhoc rule-by-rule and language-by-language basis.The objection can thus be dismissed. Notethat the result of the UIP in(19) can also be derived by allowing the use of logical negationin phonologicalrepresentations. For example,if Turkish/ t/and/ D/were bothspecified [NOT` voiced],then the rule that fills in [ 1 voiced]on /D/incodas could refer tothis specification.

(21) [` cons, 1 son, NOT` voiced] N [1 voiced]in coda

Allowingnegation in representations to havescope over a single-valuedfeature such as [ ` voiced] willnot obviously create problems. However, allowing negation to havescope over sets of valued featureswould wreak havoc with the notion of natural class. It would allow us to treat the complementset of eachnatural class as a naturalclass. For example,the segments described by the set NOT[` voiced, ` labial]would include both [d] and[p], but not [b]. Whatthe UIP doesis introduce logical negation into the interpretation of rules without enrichingthe set of primitives that can appear in lexical representations. The UIP doesredefine thenotion of natural class, in fact, but in a veryrestricted fashion: ‘ ‘[ aF]’’ and‘ ‘unspecifiedfor 210 CHARLESREISS

F’’constitutea naturalclass to theexclusion of ‘‘[ 1 aF]’’ inthe sense that the differencebetween otherwiseidentical members of a naturalclass ‘ ‘NOT[1 aF]’’ isneutralized by the rule. Under thisview, natural classes are defined, not by a featurematrix that subsumes a setof phonological representations,but by a setof phonologicalrepresentations that are accepted as inputs to a rule, giventhe UIP. Inother words, natural classes are derived from thenature of rule application, ratherthan constituting a primitivenotion of phonologicaltheory. I thinkthis is adesirableresult. We cannow return to the rejection of the use of À asa coefficientvalue for features. Obviously,we couldintroduce this value, allowing the set of values to range over $` , 1 , À%. However,this move would have implications for thebehavior of natural classes that appear to beundesirable. If phonologicalrules could refer to[ ÀF]inSDs, thenit would be possible to applyrules to segments so specified without affecting other segments. For example,it wouldbe possibleto affect Turkish / D/tothe exclusion of both / d/and / t/—say,by rounding it before roundvowels.

(22) Ahypotheticalrule [Àvoiced] N [` labial]before [ ` labial]

Myintuition is thatwe willnot find such processes; but introducing À asafeaturevalue allows suchpossibilities since [ Àvoiced]describes a naturalclass to the exclusion of / d/and/ t/.Incontrast, sucha processcannot be modeled using the UIP approach.Underspecified segments can never bereferred towithout referring to the segments with which they neutralize on the surface. Eitherthe introduction of negation into the set of lexical representational primitives (e.g., [NOT` voiced])or the use of À asapossiblefeature coefficient ([ Àvoiced])can be used to correctly modeldata with the logical structure of theTurkish stop alternations. However, I haveprovided argumentsthat introducing logical negation into the interpretiveprocedure for rules isempirically preferableto both of these alternatives. Let us turn now to discussion of a morecomplex rule type.

7Rules ContainingVariables

Inthe previous discussion, a was usedas a metalanguagevariable to refer toeither ‘ ‘` ’’ or ‘‘1 ’’withinrules. In this section, I showthat the UIP applieswithout modification also in cases where a denotesa variablewithin the rules themselves.

If SCa containsthe variable aF,thena representationQ willserve as input to R a if SDa subsumesQ andeither (a) QandSD a arespecified 1 aF(tosatisfy (19a)) or (b) QandSD a are notspecified 1 aF(tosatisfy (19b)). Since a ranges over ` and 1 , 1 a alsoranges over ` and 1 .Thismeans that if SC a isspecified aF, and SDa is not specified 1 aF,theninputs to R a cannotbe specified either ` F or 1 F(19b).Thus, R a mustbe featurefilling: its inputs must be unspecifiedfor F. Immediatelywe seethat this forces us to revise a typeof rule that is commonly invoked: thetype that involves feature-changing to either 1 F or ` F.For example,the voicing assimilationseen in Russian prepositions in (23a) is represented in standard generative phonology DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 211 asin (23b)(Halle and Clements 1983). We seethat / t/voices to [d] andthat / z/devoices to [s], whenpreceding a voicedor voicelessobstruent, respectively. (23) Traditionalstatement of Russian voicing assimilation (to be rejected in accordance with (19)) a. ‘from’ ‘without’ ‘next to’ ‘rose’ at ro´zé by iz ro´zé u ro´zé ‘Ala’(name) at a ´lé by iz a´lé u a´lé ‘cow’ at karo´vé by is karo´vé u karo´vé ‘beard’ ad barad´-õ by iz barad´-õ u barad´-õ ‘sister’ at sy istr´-õ by is sy istr´-õ u sy istr´-õ b. [` cons, 1 son] N [avoiced] / —— [1 son, avoiced] Accordingto the UIP (19),the presence of [ avoiced]in the SC andthe lack of reference to [voiced]in theSD forcethe interpretation that inputs lack [ 1 avoiced].Since a [ $` , 1 %, this rulesout any value for [voiced]in the input. Thus, in accordance with the UIP, thisrule would onlybe licitas a feature-fillingrule, applying to consonants that lacked a valuefor [voiced]. Assuming,as is traditional,that Russian obstruents are all specified as either [ ` voiced] or [1 voiced]underlyingly, the analysis could, in accordancewith the UIP, besplitinto two rules:

(24) Two-valuedassimilation a. [` voiced, ` cons, 1 son] N [1 voiced] / —— [1 voiced] b. [1 voiced, ` cons, 1 son] N [` voiced] / —— [` voiced] Rule(24a) would apply only to [ ` voiced]stops. Rule (24b) would apply only to [ 1 voiced] stops. However,it now becomes apparent that we cancollapse the two processes and still honor the UIP.

(25) [` cons, 1 son, 1 avoiced] N [avoiced] / —— [1 son, avoiced] Therule, like Polish (and Russian) coda devoicing, is thus purely feature changing: whenever thereis featurechanging, there is no vacuousapplication. If we compare(25) with (23b), we see thatonce again the UIP drivesus to accept a lessconcise representation of the rule. The new formulationcontains three mentions of a,ascompared with two in the original. 1 1

11 Of course,it is commonto represent sucha process as aruleautosegmentally spreading the [voiced] specification froman obstruent to the one preceding it. Elsewhere (Reiss, toappear a,b),I arguethat much of autosegmentalphonology mustbe replaced byanalgebraic notationusing variables and indices. The more powerfulalgebraic notationis necessary, since phonologicalquantification appears torequire the use ofquantifiersin order to express nonidentityconditions. In otherwords, some versionof theold-fashioned alpha notation may bethebest way toexpress suchprocesses. 212 CHARLESREISS

Iwillnow apply the UIP tothe generation of a morecomplex body of data, the trigger- targetrelations found in Hungarian vowel harmony.

8RegularPatterns inHungarian Vowel Harmony Inthe following discussion, I treatonly the fully productive patterns of Hungarianvowel harmony inregularstems. For example,I leaveaside the representationalissues associated with stemsthat take suffixes, rule iterativity, and the treatment of transparent . Thereare four main patterns of harmonicalternations, which I firstpresent with the orthographic vowels,since the orthography has typically guided (excessively, I believe)other treatments of thesefacts. I turnin section 9 tothephonetic values of thevowels, showing some of the deficiencies ofrelying too heavily on the orthography, and in section 10 to the lexical representation of alternatingand nonalternating vowels. In section 11, I developa setof rules for generatingthe surfaceforms from theproposed lexical representations, and in section 12, I provideschematic derivations. Nothingcrucial relies on deriving all of these alternations from rulesthat correspond to ‘‘structure-filling’’ rules;however, the complex patterns and the relative familiarity of thedata makeHungarian vowel harmony an attractive testing ground for anewproposal.

8.1Harmony involving a/e Suffixessuch as inessive -ban/-ben surface as -ban afterback vowels and -ben afterfront vowels: dobban ‘in a drum’, szemben ‘inaneye’ . Thesesuffixes are never rounded and thus surface with theunrounded front variant also after a frontround vowel: to¨kben ‘ina pumpkin’.

8.2Harmony involving a´/e´ Thereare also alternations involving the so-called long versions of a/e, asin the translative suffix -va´/-ve´ (the v assimilatesto a precedingconsonant): dobba´ ‘(turn)into a drum’, szemme´ ‘(turn) intoan eye’ , to¨kke´ ‘(turn)into a pumpkin’.

8.3Harmony involving u/u¨, u´´/u´, and o´´/o´ Theother alternations with two forms alwayshave round vowels. They agree with the backness ofthepreceding vowel. Some consistently have a shorthigh round vowel: angolul ‘inEnglish’ , to¨ro¨ku¨l ‘inTurkish’, lengyelu¨l ‘inPolish’ . Othershave the long counterparts of these high vowels: la´bu´ ‘-legged’, fejuÍ ‘-headed’. Finally,one set consistently has a longmid round vowel: dobto´l ‘from adrum’, szemtoÍl ‘from aneye’ , to¨ktoÍl ‘from apumpkin’. Thus,the height of the preceding vowelnever affects the harmonic vowel: tuÍztoÍl ‘from afire’.

8.4Harmony involving e/o¨/o Finally,we mustconsider harmony patterns with three alternants. These are restricted to asingle patterninvolving mid vowels: e/o¨/o. Thispattern is seen in the superessive suffix -en/-o¨n/-on: DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 213 szemen ‘on an eye’, to¨ko¨n ‘ona pumpkin’, dobon ‘on a drum’.1 2 Thevowel of this suffix is presentunderlyingly, and it can be accounted for straightforwardly. However,there are other vowels that participate in a vowel-zeroalternation and may be analyzedas epenthetic. These vowels generally show the same three-way e/o¨/o contrast.Such is thecase of the accusative suffix: szemet ‘eye-acc’, to¨ko¨t ‘pumpkin-acc’, dobot ‘drum-acc’. These ‘‘unstable’’ vowelsalso surface as a/e afterso-called lowering stems. I leavea discussionof such stemsaside for thepurposes of thisarticle.

9VowelPhonetics Inthe preceding section, I referred tothe vowels using standard orthography and occasional referenceto their phonetic features. In this section, I providethe IPA symbolsthat represent the vowelsin standard pronunciation, along with a featuralrepresentation. The umlaut marks front roundvowels. The lengthened umlaut marks long front round vowels, and an acute accent marks allthe other long vowels. Thus, one of the two orthographic marks of length is always present ona phoneticallylong vowel, and never present (aside from afew frozenspellings) on a short vowel.However, the phonetically long vowels do notalways differ from their‘ ‘short’’ counter- partsin lengthalone; they may also differ featurally, with respect to height, rounding, or tenseness. TheIPA symbolsand featural descriptions I haveassigned are based on examination of a varietyof sources, most importantly consultation with native-speaker linguists Pe ´ter Sipta´r and SylviaBlaho. Manydiscussions of Hungarian vowels abstract away from predictableand nondistinctive features.For example,Sipta ´r(1994:175)says of the a´ vowelthat the height(extra low, in compari- son with a/e, whichare often treated as low) and backness (fairly central) are ‘ ‘amatterof phonetic implementationsince in the (morpho)phonological pattern of Hungarian [this vowel] behaves as alowback vowel ( e.g. withrespect to vowel harmony, long/ shortalternations, etc.).’’ I will assumea minimalamount of underspecification(following Inkelas 1996) and thus include such features.Assuming an inventory of binary vowel features that consists of $hi,lo, bk, rd, ATR %, Iprefer toremain faithful to the phonetic facts wherever possible and to show how doing so helpsus toaccount for acomplexset of interactions. Note that all the high vowels are [ ` ATR]. For themidfront unrounded vowels, the short one is lax ([ 1 ATR]), thelong one tense ([ ` ATR]). However,the other four mid vowels, all the round ones, are all tense. Finally, note that orthographic a represents[ O],amidback round lax vowel, whereas a´ islowand unrounded and tense.

12 Thesuperessive does also surface as just -n whenthe stem endsin a vowel,as in kapu-n ‘ona gate’. Iassume thisresults from a straightforwardvowel deletion rule. 214 CHARLESREISS

(26) Surfacevowels of Hungarian OrthographyIPA Features Length

i [i] [` hi, 1 lo, 1 bk, 1 rd, ` ATR] short ´õ [i:] [` hi, 1 lo, 1 bk, 1 rd, ` ATR] long u¨ [y] [` hi, 1 lo, 1 bk, ` rd, ` ATR] short uÍ [y:] [` hi, 1 lo, 1 bk, ` rd, ` ATR] long e [E] [1 hi, 1 lo, 1 bk, 1 rd, 1 ATR] short e´ [e:] [1 hi, 1 lo, 1 bk, 1 rd, ` ATR] long o¨ [ï] [1 hi, 1 lo, 1 bk, ` rd, ` ATR] short oÍ [ï:] [1 hi, 1 lo, 1 bk, ` rd, ` ATR] long u [u] [` hi, 1 lo, ` bk, ` rd, ` ATR] short u´ [u:] [` hi, 1 lo, ` bk, ` rd, ` ATR] long o [o] [1 hi, 1 lo, ` bk, ` rd, ` ATR] short o´ [o:] [1 hi, 1 lo, ` bk, ` rd, ` ATR] long a [O] [1 hi, 1 lo, ` bk, ` rd, 1 ATR] short a´ [a:] [1 hi, ` lo, ` bk, 1 rd, ` ATR] long

Mostanalyses of Hungarian abstract away from allof thesesurface distinctions— for exam- ple,treating a asan unrounded vowel for thepurposes of harmony.In the following discussion, Itakea differentcourse.

10Vowels in UnderlyingRepresentation

Followingwork by Inkelas (1996) and others, I assumethat learners posit underspecification only whenalternations in the data force them to do so. In other words, it is logically possible that Polishhas the alternation / t/versus/ D/,ratherthan / t/versus / d/,butwe assumethat / D/isposited onlyif / t/and / d/are ‘ ‘alreadyassigned.’ ’ Thechild’ s defaultparse is thatunderlying segments areidentical to surface segments, and this default is rejected only if alternations force a new analysis,1 3 asin Turkish.Polish does not present the child with data that will force such a change. Sincethe transparentparse works, the learner sticks with it anddoes not consider other extension- allyequivalent grammars. 1 4

(27) Inkelas(1996:1) on underspecification [U]nderlyingrepresentation is determined solely by optimization with respect to the grammar,not by imposing any type of constraints directly on underlying representa-

13 See Hale andReiss 1998for arguments concerning the transparency of thechild’ s initialparse. 14 Anadditional source of underspecificationmay arise inorder to capture distributional patterns, but the point here isthat there isno obviousneed to minimize theamount of informationstored in lexical representations,since thisrepresents alongerlearning path; see Hale andReiss, toappear, for discussion. DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 215

tion.. ..[This]results in the use of underspecification only when there are alternant surfaceforms ... Inkelasspecifically rejects philosophical arguments for underspecification. (28) ‘‘Grammar-blind’’ approachesto underspecification rejected by Inkelas (1996, and referencescited there) a.Markedness (universal, language-specific, or contextual):Unmarked material is un- derspecified. b.Redundancy: Redundant feature values (determined on the basis of the segment inventory)are underspecified. c.Predictability: Predictable material is underspecified.

Accordingto Inkelas, ‘ ‘Theonly motivation for underspecificationis to capture alternations in theoptimal way’ ’ (1996:2).The underspecification posited for theharmonic vowels of Hungarian willbe guided by this principle.

10.1Nonalternating Vowels

Inthe case of nonalternatingvowels, I assumethat they enter and exitthe phonology with the same featuralrepresentations. There are no redundancyrules, for example.Thus, both nonalternating and alternatinghigh vowels are [ ` ATR] inthe lexicon, and they surface as [ ` ATR]. Thefact that [` ATR] ispredictablefrom [ ` hi]is not encoded in the grammar.

10.1.1Stem Vowels Asidefrom afew lengthalternations and vowel-zero alternations that will notconcern us, all stem (and prefix) vowels in Hungarian have a constantvalue. 1 5 All of the vowelsin (26)appear in stems, and many in prefixes. I assumethat they all have the underlying andsurface featural representations given above. 10.1.2Suffix Vowels Vowelharmony in Hungarian only affects suffixes. However, not all suf- fixesare subject to harmony. The following suffixes are among those that do not alternate: -ig ‘until’, -ke´nt ‘as’, -kor ‘at a time’: hatkor ‘atsix o’ clock’ , he´tkor ‘atseveno’ clock’ .

10.2Alternating Vowels

Inowpropose underlying representations for thealternating harmonic vowels. I willassume that eachof the alternations mentioned in section 8 canbe captured via a partiallyunderspecified representationand a setof phonological rules that fill in somefeature values; no feature-changing

15 Of course,a fullevaluation of myproposal will ultimately require accounting for the length alternations as well. Ileave thisfor future research. 216 CHARLESREISS rulesare needed. Length is mentionedin theprose, but not represented featurally, since I assume itis represented through multiple linking to a timingtier. 1 6 10.2.1The Lexical Form of the a/e Vowel Suffixessuch as the inessive -ban/-ben, illative -ba/ -be, andsublative -ra/-re areall stored with the following feature values for theiralternating vowel, a or e: [1 hi, 1 lo, 1 ATR]. Thisshould not be controversial, since such suffixes do in factalternate with respect to [bk] and [rd] ( a is [` rd]). Thespecification [ 1 lo]is maintained, sinceboth realizations of thevowel, [ O, E],aremid vowels. 10.2.2The Lexical Form of the a´/e´ Vowel The vowels a´ and e´ arenever high or round, but theyalternate between mid front and low back. I assumethat both forms are[ ` ATR], although thismay be debatable.Accordingly, the underlying form is[ 1 hi, 1 rd, ` ATR]. 10.2.3The Lexical Form of the Consistently Round Vowels Thealternations among the consis- tentlyround vowels can all be generated by the same rule— the only alternating feature is [bk]. The u/u¨ alternationis between front and back vowels that are consistently high, round, and tense.These vowels are represented as [ ` hi, 1 lo, ` rd, ` ATR], withno specificationfor [bk]. Thisfeature is filled in by context. The u´/uÍ alternationis between long vowels that are consistently high, round, and tense. These vowelsare represented as [ ` hi, 1 lo, ` rd, ` ATR]. The o´/oÍ alternationis between long vowels that are consistently mid, round, and tense. These vowelsare represented as [ 1 hi, 1 lo, ` rd, ` ATR]. 10.2.4The Lexical Form of the e/o¨/o Vowel Themost complex alternation to deal with is the onethat shows three surface forms, e/o¨/o. As mentionedabove, the presence of avocalicfeature bundleis underlyingin somecases of three-wayalternations and epenthetic in others. Of course, inthe latter case, it is misleading to refer tothe vowel’ s lexicalrepresentation. However, I will assumethat both underlying and epenthetic three-form suffixes enter the phonology as [ 1 hi, 1 lo].‘ ‘Feature-filling’’ rulesprovide the values for [bk],[rd], and[ATR].

10.3Personal Pronoun Forms Correspondingto the case endings discussed above are a setof stems that inflect for person.For example,the inessive -ban/-ben correspondsto bennem,benned, benne, bennu ¨nk,bennetek, ben- nu¨k ‘inme,in you,. ..’.Thestem appears with the front vowel [ E],withwhich the person endings harmonize.Similarly, the delative -ro´l/-roÍl correspondsto ro´lam,ro ´lad,ro ´la,ro ´lunk,ro ´latok, ro´luk ‘from off ofme,from off ofyou,. ..’,allwith back vowel suffixes. It is thus tempting to seethese case stems as providingevidence for theunderlying vowel quality of the case suffixes.

16 AlthoughI rejected some uses ofautosegmental phonology in footnote 11, here Iam appealingto anautosegmental representationof length. This will have to be dealtwith. DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 217

However,in several categories, the case stem and case suffix do not matchphonologically. Thesuperessive ending -en/-o¨n/-on correspondsto thecase stem in rajtam,rajtad, rajta, rajtunk, rajtatok,rajtuk ‘onme,on you,. ..’.Theelative -bo´l/-boÍl correspondsto the case stem in beloÍlem, beloÍled, beloÍle, beloÍlu¨nk, beloÍletek, beloÍluk ‘from inme, from inyou, . ..’.Thus,I willassume thatthe case endings and the case stems are listed independently in the lexicon. In other words, contraRingen and Vago (1998), I concludethat there is a setof suffixes such as / -bVn/with underspecified,alternating vowels, and a setof stemssuch as /b En-/with fully specified, nonalter- natingvowels. In some instances (e.g., inessive / -bVn/and / b En-/),theforms areclearly related etymologically,whereas in others (e.g., elative / -bV1/and / beloÍ l-/or superessive/ -Vn/and / rajt-/), therelationship is opaque. The case stems contain fully specified vowels that trigger harmony in thepersonal suffixes; the case endings harmonize with the stems they attach to. There is no need toposit feature-changing processes in those instances where the case stem and case suffix just happento look alike.

11 The Rules Inthis article, I amconcerned only with capturing the interactions between triggers and targets ofharmony.Therefore, I willnot address the issues of transparent vowels or ofrootswith surface frontvowels that trigger back harmony. I willalso not propose an explicitmechanism for harmony interms of syllable structure or higher-level structures. I willadopt a simplifiedview in which harmonyoccurs between a voweland the vowel that precedes it. The following seven rules are positedwith crucial ordering noted. All seven are what would be called feature-filling or structure- fillingrules in other accounts. 1 7 Thisdoes not need to be stated for eachrule, since it follows from theUIP.

R1 : V N [abk] / [abk] —— Thisrule applies first. According to the UIP, itcan only apply to vowels that have no specificationfor [bk].Thus, it applies to allof the alternating vowels and none of thenonalter- natingones.

R2 : [1 lo, ` bk] N [` rd] Inaccordance with the UIP, thisrule applies to [ 1 lo, ` bk]vowels that are not [1 rd]. The ruleapplies vacuously to already round vowels that are also [ 1 lo, ` bk],and in feature-

fillingfashion to [ 1 lo, ` bk]vowelsthat are unspecified for [rd]. SinceR 2 appliesto vowels

thatbecame [ ` bk] via R1 , R2 mustfollow R 1 .For example,a vowelthat enters the phonology

as [1 hi, 1 lo]could become [ ` bk] by R1 ifit follows a backvowel, and then [ ` rd] by

R2 .Itwould then receive an [ATR] valuefrom R 6 ,andat this point it would be fullyspecified for allthe vowel features.

17 Somealso happen to be what are traditionallycalled fill-inrules— rules that supply default values only with reference tocontext within the segment itself. 218 CHARLESREISS

R3 : [` bk, 1 rd] N [` lo] Thisrule ensures that a´ isrealized as alowvowel, the only one in the language. Crucially, thevowel of the a´/e´ alternationis underlyinglyspecified as [ 1 rd].The rule refers to[ ` bk],

andsince all the alternating vowels are underlyingly unspecified for [bk],R 3 must follow

R1 .

R4 : [1 hi, 1 bk, 1 ATR] N [1 rd] Thisrule ensures that the e of the e/a alternationsurfaces as unrounded. It cannot effect this featureinsertion for the e of the e/o¨/o alternation,since the latter is not specified as [ 1 ATR]

until R6 applies.

R5 : V N [ard] / [ard] —— Bythe UIP, thisrule cannot apply to any vowel that is already specified for rounding.It turnsout that it only affects the front vowels of the e/o¨/o alternation.All other vowels that

receiverounding by rule get [ ` rd] from R2 or R4 . Thus, R5 followsthese two.

R6 : [1 hi, 1 lo, ard] N [aATR] Thisrule supplies mid vowels that are unspecified for [ATR] withan [ATR] valuethat agreeswith their rounding value. The UIP requiresthat inputs to this rule not have an [ATR] value;thus, there is no chance that long e´ willbe affected by this rule. This rule applies

after R2 andalso after R 5 ,sincethese fill in thevalues of [rd] requiredfor therule to apply.

R7 : [1 hi, 1 bk] N [1 lo] Thisprovides the [ 1 lo]value to thefront member of the a´/e´ alternation.Since the SD refers

to [bk], R7 mustfollow R 1 .Therule will apply vacuously to a derived[ 1 hi, 1 bk, 1 lo] e, as well.

12Derivations

Thefollowing tables provide derivations of each of thealternations. Input and output are given withorthographic vowels. The only features of thetrigger of harmony that are relevant are [bk] and[rd], soit isnotnecessary to provideall trigger-target pairs. For example,the vowel u triggers thesame harmony patterns as o. Ihaveprovided schematic derivations for fourtrigger vowels representingeach of thefour possible combinations of thesefeatures: [ ` bk, 1 rd], [` bk, ` rd], [1 bk, 1 rd], and [1 bk, ` rd]. Inthe first table, (29), I haveannotated the cases where the rule does not affect the input, whereasin the subsequent tables I merelynote NA for notaffected. If aboxis marked NS (no subsumption ),itmeans that the SD ofthe rule does not subsume the potential input representation.

For example,R 3 doesnot apply to anyforms intable (29), since the structural description of R 3 doesnot subsume any of the input representations in this table. Thus, such cases conform to standardpractice: the rule does not apply because it does not match the SD. However,some cells are marked UIP toshowthat the current rule’ s SDwouldbe metunder traditionalassumptions, but it is notmet under the UIP. For example,R 5 doesnot apply to any DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 219 forms in(29) since all the inputs are specified for [rd] bythis point in the derivation, and R 5 mustbe featurefilling by the UIP. Finally,some cells are marked VAC; here,the SD issatisfied under the UIP, butthe effect oftherule is notvisible (it applies vacuously). This is the case for R 7 inthe second and fourth datacolumns in (29), where it applies vacuously to vowels that are already [ 1 hi, 1 lo, 1 bk].

Theonly rules that affect the output are R 1 (for allforms inthis table), R 2 (forms precededby abackvowel trigger), and R 4 (forms precededby a frontvowel trigger). (29) Generating the a/e alternation

á__ e__ o__ ö__ 2hi 2hi 2hi 2hi 1lo 2hi 2lo 2hi 2lo 2hi 2lo 2hi In. 1bk 2lo 2bk 2lo 1bk 2lo 2bk 2lo 2rd 2ATR 2rd 2ATR 1rd 2ATR 1rd 2ATR 1ATR 2ATR 1ATR 1ATR 2hi 2hi 2hi 2hi 2lo 2lo 2lo 2lo R 1 1bk 2bk 1bk 2bk 2ATR 2ATR 2ATR 2ATR 2hi 2hi 2lo 2lo R2 1bk NS 1bk NS 1rd 1rd 2ATR 2ATR

R3 NS NS NS NS 2hi 2hi 2lo 2lo R4 NS 2bk NS 2bk 2rd 2rd 2ATR 2ATR

R5 UIP UIP UIP UIP R6 UIP UIP UIP UIP R7 NS VAC NS VAC Out. á-a e-e o-a ö-e

Table(30) shows the derivation of the a´/e´ alternation.We seethat rules R 1 and R3 affect theforms triggeredby precedingback vowels, whereas rules R 1 and R7 affectthe forms triggered bypreceding front vowels. Thenext two tables, (31) and (32), show the derivation of front/backalternations of consis- tentlyround vowels. In both cases, the only relevant rule is R 1 .Sincethe long high round vowels u´, uÍ behavein completely parallel fashion to the short ones u, u¨, Ishowonly the latter.

Intable (33), rules R 1 and R6 affectall forms. R 2 onlyaffects forms triggeredby back vowels,and R 5 onlyaffects forms triggeredby front vowels. The input vowel is merely specified [1 hi, 1 lo],but a three-way e/o¨/o alternationis generated. 220 CHARLESREISS

(30) Generating the á/é alternation

á__ e__ o__ ö__ 2hi 2hi 2hi 2hi 1lo 2hi 2lo 2hi 2lo 2hi 2lo 2hi In. 1bk 2rd 2bk 2rd 1bk 2rd 2bk 2rd 2rd 1ATR 2rd 1ATR 1rd 1ATR 1rd 1ATR 1ATR 2ATR 1ATR 1ATR 2hi 2hi 2hi 2hi 2rd 2rd 2rd 2rd R 1 1bk 2bk 1bk 2bk 1ATR 1ATR 1ATR 1ATR

R2 NA NA NA NA 2hi 2hi 1lo 1lo R3 2rd NA 2rd NA 1bk 1bk 1ATR 1ATR

R4 NA NA NA NA

R5 NA NA NA NA R6 NA NA NA NA 2hi 2hi 2lo 2lo R7 NA 2rd NA 2rd 2bk 2bk 1ATR 1ATR Out. á-á e-é o-á ö-é

(31) Generating the u/ü alternation

á__ e__ o__ ö__ 2hi 2hi 2hi 2hi 1hi 1hi 1hi 1hi 1lo 2lo 2lo 2lo 2lo 2lo 2lo 2lo In. 1bk 2bk 1bk 2bk 1rd 1rd 1rd 1rd 2rd 2rd 1rd 1rd 1ATR 1ATR 2ATR 1ATR 1ATR 1ATR 1ATR 1ATR 1hi 1hi 1hi 1hi 2lo 2lo 2lo 2lo R1 1bk 2bk 1bk 2bk 1rd 1rd 1rd 1rd 1ATR 1ATR 1ATR 1ATR

R2 NA NA NA NA

R3 NA NA NA NA R4 NA NA NA NA R5 NA NA NA NA

R6 NA NA NA NA R7 NA NA NA NA Out. á-u e-ü o-u ö-ü DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 221

(32) Generating the ó/oÍ alternation

á__ e__ o__ ö__ 2hi 2hi 2hi 2hi 2hi 2hi 2hi 2hi 1lo 2lo 2lo 2lo 2lo 2lo 2lo 2lo In. 1bk 2bk 1bk 2bk 1rd 1rd 1rd 1rd 2rd 2rd 1rd 1rd 1ATR 1ATR 1ATR 1ATR 1ATR 2ATR 1ATR 1ATR 2hi 2hi 2hi 2hi 2lo 2lo 2lo 2lo R1 1bk 2bk 1bk 2bk 1rd 1rd 1rd 1rd 1ATR 1ATR 1ATR 1ATR

R2 NA NA NA NA

R3 NA NA NA NA R4 NA NA NA NA R5 NA NA NA NA

R6 NA NA NA NA R7 NA NA NA NA Out. á-ó e-oÍ o-ó ö-oÍ

(33) Generating the e/ö/o alternation

á__ e__ o__ ö__ 2hi 2hi 2hi 2hi 1lo 2lo 2lo 2lo 2hi 2hi 2hi 2hi In. 1bk 2bk 1bk 2bk 2lo 2lo 2lo 2lo 2rd 2rd 1rd 1rd 1ATR 2ATR 1ATR 1ATR 2hi 2hi 2hi 2hi

R1 2lo 2lo 2lo 2lo 1bk 2bk 1bk 2bk 2hi 2hi 2lo 2lo R NA NA 2 1bk 1bk 1rd 1rd

R3 NA NA NA NA R4 NA NA NA NA 2hi 2hi 2lo 2lo R NA NA 5 2bk 2bk 2rd 1rd 2hi 2hi 2hi 2hi 2lo 2lo 2lo 2lo R6 1bk 2bk 1bk 2bk 1rd 2rd 1rd 1rd 1ATR 2ATR 1ATR 1ATR

R7 NA NA NA NA Out. á-o e-e o-o ö-ö 222 CHARLESREISS

We thussee that a complexset of vowel-vowel interactions can be derived using a setof simplerules, appropriately applied in accordance with the UIP. As mentionedin passing, other problemsclearly remain to be solvedin the domain of Hungarianvowels, and it will be interesting tosee if the present proposal can shed new light on these classic problems. While there is no logicalnecessity to derive these patterns exclusively with the equivalent of feature-fillingrules, doingso allows us tosimplifythe account of vowelharmony in a straightforwardfashion, a topic we nowturn to.

13Mixed Stems WhenHungarian vowel harmony is generated in this way, it is unnecessary to explain why harmonydoes not occur between stems in a compound(e.g., balta` nye´l ‘hatchethandle’ ) or betweena prefixand a stem(e.g., meg` la´t ‘catchsight of’ ), despitethe standard view on this matteras expressed by authors like Spencer (1996:178) and Sipta ´r and To¨rkenczy(2000:chap. 6).These authors attribute the failure of harmony to apply in such cases to the assignment of prefixesand first compound members to separatephonological words from thestem that follows them.While such a divisionmay be reasonable on other grounds, vowel harmony, or rather its failureto apply, provides no evidencefor aphonologicalword break between such constituents. Underthe model presented here, all prefix and stem vowels are fully specified 1 8 andthe proposed vowelharmony rules affect only underspecified vowels, when interpreted using the UIP.

14Conclusions We haveseen that the goal of conciseness,or maximal generality, in ruleformulation was inconsis- tentwith the search for UG. Ruleswere formulatedin order to generate a givencorpus of data, notwith the aim of modelinga singlehuman phonological component. It is desirableto have a singleinterpretive procedure for determiningwhether a representationsatisfies the SD ofarule andthus serves as an input to the rule. This single procedure should work for allrules in all languages.In this article, I haveexplored a logicalalternative to theinterpretive procedure adopted in SPE andargued that this unified interpretive procedure solves long-standing problems by makingit unnecessary to refer toa thirdvalue of binary features [ ÀF], introducenegation into lexicalrepresentations (e.g., [ NOT` rd]), orintroducea featurefilling /featurechanging diacritic on rules. Ihaveargued that a smallset of rules,provided with the UIP, cangenerate a setof dataas complexas thetarget-trigger relations of Hungarian vowel harmony. The UIP providesa principled distinctionbetween feature-filling and feature-changing rules, one that is read mechanically off therule’ s representation.We donot need labels (or acleverhomunculus) to tell us if a ruleis featurefilling or feature changing, and we donot need diacritics in representations to refer to

18 Aside fromsome epenthetic,‘ ‘unstable’’ vowels,which receive theirfull specification from within their stem. DERIVINGTHEFEATURE-FILLING/-CHANGINGCONTRAST 223 featuresthat are not present. If thepurely feature-filling analysis of Hungarian presented here 1 9 isin fact valid, then we haveshown that this conclusion leads to an overall simplification in our understandingof Hungarianphonology: there is less evidence for positingphonological domains withinwords to accountfor thelack of harmonyin mixed stems and compounds. Not surprisingly, ratherthan being ‘ ‘ad-hocand sterile,’ ’ attentionto theoretical detail can lead to improvedanalyses. Onemight wonder whether the UIP isjust a notationalvariant of the rule labels feature filling and featurechanging, sinceit appearsto replacethese two labels with a disjunctivecondition onrule application. There is a goodargument that it is not. Labels could be used with great latitude,in principle. For example,using the mechanism of rulelabeling, a rulecould be stated thatfilled in valuesfor [rd] onlyon segments lacking specificationfor [hi].The notion of a rule labeldoes not preclude this possibility, though I believeit is not attested, and the UIP tellsus whyit is not:feature-filling and feature-changing differences derive from theUIP. TheUIP uses theSC ofaruleto determinewhich features must be absentfrom arepresentationfor ittoserve asan inputto thatsame rule. In other words, the UIP explainswhy just the feature that appears inthe SC istheone for whicha rulecan require obligatory absence. TheUIP allowsrepresentations in thelexicon to be nondistinct,and it evenmakes it possible for onerepresentation to subsume another, since it allows the two to betreateddifferently by the grammar ina principledfashion. Specifically, a rulecan affect a representationA, yetfail to affectanother representation B, evenwhen A ismore general than (subsumes) B. As thereferences toearlier work by Inkelasand Orgun, Kiparsky, Archangeli, and McCarthy demonstrate, this has beena long-standingproblem in phonological theory. Bydevelopinga workable(albeit incomplete) underspecification analysis for Hungarian,we areworking toward a simplermodel of Hungarianphonology, in general. If thecontrast between suffixalternations and prefix constancy, for example,does not call for astratalphonology, the questionarises how general this result is —bothfor Hungarianand for universalphonology. Can theUIP asformulatedhere be usedto simplify other analyses that in thepast drew onthearsenal oftraditionalgenerative phonology, such as theDerived Environment Condition, the Strict Cycle Condition,Structure Preservation, and theElsewhere Condition? This important issue will provide fertileterrain for futureresearch. At thispoint, we cancompare the introductoryquotation from MorrisHalle with a competing viewof theproper goals of phonology:

Weurge a re-assessmentof this essentially formalist position. If phonology is separated from the principlesof well-formedness(the ‘ laws’) thatdrive it, the resulting loss of constraint and theoretical depthwill mark a majordefeat for the enterprise. (Prince and Smolensky 1993:198; see also p. 3)

Thisarticle is an attempt to supportHalle’ s view.We shoulddevelop the necessary formal tools toexpress phonological processes, and we shouldunderstand how our notation works, whether itis constraintor rulebased. I findthis to be a moreinteresting and pressing problem, one that

19 Or anotherfeature-filling analysis. 224 CHARLESREISS canlead to greater ‘ ‘theoreticaldepth,’ ’ thana taxonomicencoding of dubious ‘ ‘principlesof well-formedness.’’

References Anderson,Stephen. 1985. Phonologyin the twentieth century. Chicago:University of Chicago Press. Archangeli,Diana. 1988. Underspecificationin Yawelmani phonology and morphology. New York:Garland. Bayer,Sam, and Mark Johnson. 1995. Features and agreement. In Proceedingsof the 33rd Annual Meeting oftheAssociation for Computational Linguistics, 70–76.San Francisco: Morgan Kaufmann. Chomsky,Noam. 1967. Some general properties of phonologicalrules. Language 43:102–128. Chomsky,Noam. 1986. Knowledgeof language. New York:Praeger. Chomsky,Noam, and Morris Halle. 1968. Thesound pattern of English. New York:Harper and Row. Hale,Mark, and Charles Reiss. 1998. Formal and empirical arguments concerning phonological acquisition. LinguisticInquiry 29:656–683. Hale,Mark, and Charles Reiss. To appear. The Subset Principle in phonology: Why the tabula can’t be rasa.Journal of Linguistics 39. Halle,Morris. 1975. Confessio grammatici. Language 51:525–535. Halle,Morris, and George N. Clements.1983. Problembook in phonology. Cambridge,Mass.: MIT Press. Inkelas,Sharon. 1996. Archiphonemic underspecification. Ms., University of California,Berkeley. Inkelas,Sharon, and Orhan Orgun. 1995. Level ordering and economy in thelexical phonology of Turkish. Language 71:763–793. Kenstowicz,Michael, and Charles Kisseberth. 1979. Generativephonology. New York:Academic Press. Kiparsky,Paul. 1985. Some consequences of lexicalphonology. PhonologyYearbook 2:85–138. McCarthy,John. 1994. The phonetics and phonology of Semiticpharyngeals. In Phonologicalstructure and phoneticform: Papers in laboratory phonology III, ed.by Patricia Keating, 191– 233. Cambridge: CambridgeUniversity Press. Prince,Alan, and Paul Smolensky. 1993. Optimality Theory: Constraint interaction in generativegrammar. Technicalreport, Center for Cognitive Science, Rutgers University, New Brunswick,N.J. Reiss,Charles. 2002. Acquisition and post-OT phonology. Ms., Concordia University, Montre ´al, Que´bec. Paperpresented at ‘‘Phonology2000,’ ’ HarvardUniversity and MIT, 1999. Reiss,Charles. To appear a. Quantification in phonological rules. TheLinguistic Review 20(2/3)(special issueon typology in phonology). Reiss,Charles. To appear b. Towardsa theoryof fundamental phonological relations. In Paperson linguistic asymmetry, ed.by Anna Maria Di Sciullo.Amsterdam: John Benjamins. Ringen,Catherine, and Robert Vago. 1998. Hungarian vowel harmony in Optimality Theory. Phonology 15:393–416. Sipta´r, Pe´ter.1994. The vowel inventory of Hungarian:Its size and structure. In Theeven yearbook 1994, ed. by La´szlo´ Varga,175– 184. Budapest: Eo ¨tvo¨s Lora´nd Tudoma´nyEgyetem, School of English andAmerican Studies Working Papers in Linguistics. Sipta´r, Pe´ter,and Miklo ´s To¨rkenczy.2000. Thephonology of Hungarian. Oxford:Oxford University Press. Spencer,Andrew. 1996. Phonology. Oxford:Blackwell.

ConcordiaUniversity LinguisticsProgram 1455de MaisonneuveW. Montre´alH3G 1M8 Que´bec, Canada [email protected]