* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download The Lexical Basis of Sentence Processing
Old Norse morphology wikipedia , lookup
French grammar wikipedia , lookup
Modern Greek grammar wikipedia , lookup
Junction Grammar wikipedia , lookup
Lithuanian grammar wikipedia , lookup
Udmurt grammar wikipedia , lookup
Old Irish grammar wikipedia , lookup
Polish grammar wikipedia , lookup
Proto-Indo-European verbs wikipedia , lookup
Macedonian grammar wikipedia , lookup
Kannada grammar wikipedia , lookup
Chinese grammar wikipedia , lookup
Portuguese grammar wikipedia , lookup
Ukrainian grammar wikipedia , lookup
Japanese grammar wikipedia , lookup
Navajo grammar wikipedia , lookup
Germanic weak verb wikipedia , lookup
Germanic strong verb wikipedia , lookup
Swedish grammar wikipedia , lookup
English clause syntax wikipedia , lookup
Ancient Greek grammar wikipedia , lookup
Modern Hebrew grammar wikipedia , lookup
Italian grammar wikipedia , lookup
Yiddish grammar wikipedia , lookup
Old English grammar wikipedia , lookup
Ancient Greek verbs wikipedia , lookup
Russian grammar wikipedia , lookup
Spanish grammar wikipedia , lookup
Hungarian verbs wikipedia , lookup
Sotho verbs wikipedia , lookup
Latin syntax wikipedia , lookup
Icelandic grammar wikipedia , lookup
Serbo-Croatian grammar wikipedia , lookup
Georgian grammar wikipedia , lookup
Kagoshima verb conjugations wikipedia , lookup
Natu ral [.irnguage Processing liditor P r o f .l l r r s l a nM i t k o v .Scf rool of I{runanities, Languagesand SocialSciences Urrivcl sit y o{-Wolverhampton Srafli)KlSt. Wolverharnpton\MVl lSB, United Kingdom Ilrnail:l(.Mit [email protected] The LexicalBasis of SentenceProcessing Formal,computational and experimentalissues Advisory Board Christiarrlloitet (Universityof Grenoble) John Carroll (University of Sussex,Brighton) Eugerre Charni ak ( Brown University, Providence) Editedby Ilduard IIovy (lnformation Sciences Institute,USC) llichard Kittredge (University of Montreal) Geoffr'eyLeech ( LancasterUniversity) (larkrs Martin-Vicle(Rovira i Virgili Un., Thrragona) Arrdrei l\likheev(Universityof Edinburgh) University of Geneva (Universityof Groningen) John lJerlrorrrre Nicolas I.Jicoktr,(lBM, T. J.Watson ResearchCenter) Kcrrral( )flazer(SabanciUniversity) University of Toronto PaolaMerlo SuzanneStevenson Allarrlla rrrsey( UMISI Manchester) I\,lolrirlrrc Ikrlbert (Universit6de Marseille) l(ic:lrardSlrt'oat(AT&T Labs Research,Florham Park) Keh-YilrSu (RehaviourDesignCorp.) l s a b c l l e ' l i ' a r r c o s( loN E S C ,L i s b o n ) 'lsou (City Universityof Hong Kong) llenj;rrnirr 'llsujii(University of Tokyo) f un-iclri Iivclyrrc'lToukernrann(Bell Laboratories,Murray Hill) Yori<k lVilks (Universityof Shelfield) V<rltrrne 4 'l'he Lexical Ilasis of Sentence Processing:Formal, computational and e x D c l i l r r e r r t iasls r r e s L,tlitcdhy PaolaMerlo anclSuzanneStevenson Iohn BenjaminsPublishingCompany Amsterdam / Philadelphia 'l'he /o^,tt \--' rrsetlin tlris publicationmeetsthe minimum requirementsof American 1',r1'er Nari,rrrirl sr;rrrrl,rrd - permanenceof paperfor printed for InformationSciences L i b r a r yM a t e r i a l s^,N s r2 3 9 . 4 81- 9 g 4 . Thbleof contents Preface vtI words, numbers and all that: The lexicon in sentenceunderstanding SuzanneStevensonand Paola Merlo Part I: Fundamental issues The lexicon in Optimality Theory loan Bresnan Optimality-theoretic Library of Corrgrcss Cataloging-in-publication Data 59 Mark lohnson -I'he l.r:xicalbasisol st rrterrceprocessitlg: formal, computationaland experimentalissues / editerlby PaolaMcrlo alrtlSuz.anne Stevenson. processirrg, p. crn. (Naturall,anguage rssN1567_9202v.4) "'l'his volunrc <lclivesft'onrthe specialconferencesession...held in conjunction with the I I th Arrnual( l(.lNY( irnfcrenceon lluman Sentence processing, March l9-2I, 199g"_pref. Irrcludes bibliographical references and index. l- (-lrarnrrrar, oo'rparative and general--syntax--congresses.2. Lexicology-C)otrgresses. 3. Pltyclroliguistics--Congresses. 4. Computational linguistics--Congresses. I. Merftr,Paola,1949--ll. Stevenson,Suzanne.IIL Natural Languageprocessing(Amsterdam, Netherlands): v.4. P29s.t.4882002 415--<lc2l tssN90 272 49873 (Iirrr.)/ I -sBSlI 1563 (US) (Hb; alk. paper) @ 2(lO2* JohnRt'njanriir.s Il.V. Lexical Functional Grammar 39 The lexicon and the laundromat /> Jerry Fodor Semanticsin the spin rycle: Competenceand performance criteria for the creation oflexical entries 85 Amy Weinberg Connectionist and symbolist sentenceprocessing 95 Mark Steedman Part II: Division of labour between syntax and the lexicon 20010583 I3 No Pi111 of this brxrk rrraybc reproducedin any form, by print, photoprint, microfilm, or any otlrcl rrreans, rvitlrontrvrittenpermissionfrom the publisher. (lo. .P.O.Rox36224.1020ue Amsterdam.The Netherlands I'rrl'11511111g ItrhrrIfenjtrrrirrs r s. r r l r A n r e r i c a . P . ( B J .o x 2 T 5 l g . p h i l a c l e l p hpi ae l g l l g - 0 5 l 9 . u s a J o h r rl l c n j a r r r i rN A computational model of the grammatical aspectsof word recognition as supertagging tog Albert E. Km, Bangalore Srinivas and lohn C. Trueswell Incrementality and lexicalism: A treebank study VincenzoLombardo and Patrick Sturt r37 -lahle oI cotttcnts Modular archi I cctu resand statisticalmechanisms: 'l'he casellonr lexicalcategorydisambiguation It4atthew W. Crttckerond Steffan Corley r57 Ilncoding and storagein working memory during sentenceconrltlcltensi<ln Law'ie,zl.S/orye,RienkCl. Withaar, AlbertusA. Wijers, r8r Preface CeesA.l. IJrttereatrd AtrrrcM.J. Paans '['he time course of inftrrrtration integration in sentenceprocessing Miclmd J. Spivey,StonkaA. Fihrcva"WhitneyTabor n,td Sdilrc(r Ajrt:l-:-tti 207 I)art III: Details of lexical entries 'l'lrc participantsand their role lexicals()rrrceof ttnexpressed i u s e r r t c r r caer t r l ,l i s c o u st e u t t d e r s t a n d i n g Gnil h4nrtrrcr,Jtnrr-l'ierreKoenig, AlissaMelinger 233 antl lJretorr Bicrr retrrte ll.educeclt'elalivcsjudgetl hard require constraint-basedanalyses Honn ltilip, Nliclnel K. Thnenhaus,GregoryN. Carlson, Porrll). AllopennoortdloshuaBlatt 255 Prerlictingtlrerrraticrole assignmentsin context Gerr\,7'. h4.Altnnrtrt z8t [.cxicalsernanlicsas a basisflorargument structure frequencybiases VeredArcnrrtnnmul NeolJ. Pearlmutter 303 probabilities Vcllr senseatrd verb sulrcategorization Drtrtgloslloland ruul l)aniel lurafsky 325 A r r l l r o irr r d e x 347 I l r - r ni r r d e x 355 This volume derives from the special conference sessionentitled "The Lexical Basisof sentenceProcessing:Formal and computational Issues,"held in coniunction with the 1lth Annual CUNY Conference on Human Sentenceprocessing,March 19-21, 1998.The special sessionhigtrlighted talks and posters on current theories of the lexicon from the perspectiveof its use in sentence understanding. Lexical influences on processing are currently a major focus of attention in psycholinguistic studies of sentence comprehension; however, much of the work remains isolated from the vast amount of scientific activity on the topic of the lexicon in other subdisciplines. In organising the special session,we felt that the time was ripe to bring together researchersfrom these different perspectivesto exchangeideas and information that can help to inform each others'work. Participants included a multi-disciplinary slate from theoretical linguistics, computational linguistics, and psycholinguistics,representing various theoretical framework within each of these disciplines. The specialsessionwas quite successfulwith about 250 registrants,many of whom do not belong to the usual CUNY crowd. A primary motivation for the specialsessionwas that a focus of attention on the lexicon has the potential to bring the structural and probabilistic approachesto sentenceprocessingcloser together.To gain a deeper understanding of the lexicon and its impact on processing,we need to elaborate both the structure and the probabilistic content of lexical representations,which together influence sentenceinterpretation. These questions bring up general issuesin the architecture of the mind, and meet up with work in computer science,linguistics, and philosophy on the relation between conceptual knowledge, grammatical knowledge, and statistical knowledge. The contents of this volume reflect this range of issues from various perspectivesin the multidisciplinary study of the lexicon in languageprocessing.A selectionof the contributors to the specialsessionwere invited to preparewritten versions of their presentationsto be included in the volume. The preparation of the final version of the papers was assistedby very careful and detailed multiple peer reviewing. Very few of the reviewers were also contributors to 2 5 4 ( i i l l N { i r u n c re t a l . I'carlrrrrrrrer.N., & MacDonald, M. (1992). Plausibility effects in syntactic ambiguity of the cognitivesciencesociety of the I4th annual conference resoltrli,rrr.ht I'roccedings IN. (p. 498-'5t)i). Rloonrington, lleichle,ll.I).,I\rllatsek,A.,Fisher,D., &Rayner,K.(1998).Towardamodelofeyemov€ment Review,105,125-157ccrntrolirr reading.Psy'clnlogical 'f. ( arguments and the head-complementrelation. LitrguisticInquiry, 9ti7). I rnplicit I{<relrr:r; I ,l8,267-.1| O. l{umcllrart, l)., & ortony, A. (1977). The representationof knowledge in memory. ln of knowledge' l{. Anrlers6rr,l{. Spiro & W. Montague(Eds.),Schooling and theaccluisition Reducedrelativesjudged hard require constraint-basedanalyses Hana Filip*, Michael K. Thnenhaus+*,GregoryN. Carlson*, PaulD. Allopenna+and JoshuaBlatt+l 99-135. I lillsdale,NJ: LawrenceErlbaumAssociates. spccr, s.ll., & (jlifHrrr, c., Ir: (1998). Plausibility and argument structure in sentence and Cognition,26, 965'978. rcrrsiott.I'"lenrory c<rrrrlrref 'lhbossi, Sl,ivcy-l(tr,Nvlton,M., McRae,K.' &Tanenhaus, M. (1994)' Semanticeffects P., <r1sl,rrlar:ticarrrbiguityresolution:Evidencefcrra constraint-basedresolutionprocess. Irr M. lr.loscovitch(Ed.), Anention and performanceXV (p.5S9-615). Hillsdale' NJ: University of Rochester we takeasour point of departurestevensonandMerlo's(1997) observation that the differencesi' the processingdifficulty of sentences with reduced relativeclauses(RRs)arestronglydeterminedby the inherentlexical semanticclassof the verbsusedaspassiveparticiplesin RRs:namely, the unaccusative vs.unergativeclass.our main claim is that among the linguistic variablesresponsiblefor the relevantdifferencesa crucial role is playect by sema'tic variables,rather than just category-levelsyntacticcomplexity and/or complexityassociatedwith word-internal lexicalstructure of verbs (seeHale and Keyser,1993).First, we observea considerable overlapin the distributionsof acceptabilityjudgmentsbetweensentences with RRsbased on unaccusative verbsand thosebasedon unergativeverbs,and evennrore importantly, cleargradienteffectswith respectto acceptabilityjudgments for both typesof sentences that are influencedby the lexicalsemanticsof the main verb in the matrix clause.Second,suchdatacan be successfully motivated,if we characterizethe crucial unaccusative-unergative distinction in termsof thematicproto-Roleproperties(Dowty,lggg, lggl). Third, the linguistic analysisis consistentwith recentconstraint-basedgrammars, most 'otably HPSG,and our co'straint_based modelthat usesthe integration-competition architecture developed by spivey(1996)and appliecl to reducedrelarives by McRaeet al. (199g)and Spiveyand Tanenhaus ( 199g). []tlbattrl Associates, l,arvrertce '[iucswcll, M., & Garnsey,S. (1994).Semanticinfluenceson parsing:Useof 1.,'larrerrlraus, tfrerrraticr,rlc irrforrnationin syntacticambiguity resolution.]ournal of Memory and Ln rgtt,t yt',32, 285*318. 'li'ueswcll, J.,'lartenhaus,M., & Kello, C. (1993).Verb-specificconstraintsin sentence elfects of lexical preferencesfrom garden-paths.Journal of proccssirrg:Scpa111i11g Letrning Metnoryand Cognition,19' 528-553. ExltcrilrwrtnlPsychology: Van Valirr, lt., & I.apolla, R. (1997). Syntax:form, meaning,and function. Cambridge: CarnbritlgcUrriversityPless. Van Valin, It., & Wilkins, D. (1996).The casefor'effector': caseroles, agents,and agency (p.289& S. Thompson(Eds.),Granunaticalconstntctions Irr M. Shibatarri rcvisire<I. ()xford ( Press. )xford: University 322). Wif lianrs,ti. ( l9S7). Implicit arguments,the binding theory and control' Natural Language 5, I 5 1-180' nnd Lit ryrristic'['lrctn'1', 1. Introduction Beginning with Bever's (1970) classicarticle, sentenceswith reduced relative clauses,such as The horse racedpast the barn have served as an important fell, empirical testing ground for evaluating moders of sentenceprocessing. Bever observed that sentenceswith reduced relative clauses are difficult to under_ zi.6 | IanaFilip ct al stanrl, with p('()ple often judging the sentencesto be unacceptable,because 'NP tlrey initially assunlethat the V PP' sequenceis a main clause.In subse,.1trent dccatlcsone of the central controversiesrevolvedaround the question of rvlretherstlur:trrral cor-nplexityplays a prirnary causalrole in processingdifficrrlty clf sentcrrceswith reduced relative clauses,and other sentenceswith temIior exanrple,in recentconstraint-basedmodels,the diffil)()raryarrrhirirrities. r ulty o[ retlucctl relative clausesis argued to arise from an interaction of multiple corrstrrirrts,lnany of which are lexically-based(e.g.,MacDonald, Pearlrrrrrtterarr<lSt'itlerrberg, 1994;lbnenhaus and Trueswell,1995;Boland, 1997), I'rrt which d,' 111;1 irrclude any factors directly attributable to intrinsic easeor tlilficulty ol'p11r6qsci11g syntacticstructures.lmportant empirical evidencein approacheshas come from gradienteffectsin the srrpportof <orrstraint-lrased r lilJiculty reduced of relatives.For example,The eggscookedin butter I'rocessirrlt I'tsterldalit irtrr.sis clearly much easierto processIhan The horseracedpast the lnru .fell. lrr corrstrastto rnced,cookedis much more often used transitively,it is rrr<rrefr-ctlucntlyused as a passive,and eggsis a very poor Agent, but a very g,o<lcl'l'herne , in a cooking event.Due to such constraintsthe activeintransitive 'l'ltc reatfing r>f cggscookedwith butter... is lesslikely and the passiveparticiple lil<ely. rcaclirrgrnorc Gradient effectsin processingdifficulty for reducedrelative r'lauseslravebecn successfullymodeled using computational implementations ol-nrultiple constraint models (McRae,Spivey-Knowlton and Tanenhaus,1998; 'larrcnhaus, 1998). l i p i v e ya n c l Ilecerrtly,Stevensonand Merlo (1997) made the important observation tlrat the pnrcessing difliculty of sentenceswith reduced relative clauses is strongly rlclcrrninedby the inherent lexical classof the verbs used as passive with reducedrelativesheadedby pasI'irrticiplcsirr leduceclrelatives.Sentences sivc participl,:s tlet'ive<lfrorn unergative2verbs are "all mostly or completely urracce ptalrlc"(p. 355).In particular,manner of motion verbs"lead to a severe liirrclerrlrilth in tlre l{R construction" (p. 353), as is shown in Stevensonand lvlerlo's(p. l!' j) exarnples,hererepeatedin (1). In contrast,"unaccusativeRRs ;tr.eall cottrl'lclclyacceptable or only slightlydegraded"(P. 355).Stevensonand Merlo'scxarrrples are repeatedherein (2): 'l'lre ir. clippersailedto Portugalcarrieda crewof eight. 'l'lre b. troops rnarchedacrossthe fieldsall day resentedthe general. 'l'he nrodelplanet rotatedon the metal axis fell offthe stanc. c. 'l'lre dog walkedin the park washavinga good time. tl. 'l'[re (2) a. rvitchrneltedin the Wizard of Oz wasplayedby a famousactress. 'l'lre genesmutatetl in the experimentwere usedin a vaccine. b. ( l) Reducedrelativesjudged hard requireconstraint-based analyses c. d. The oil poured acrossthe road made driving treacherous. The picture rotated 90 degreeswas easyto print. stevensonand Merlo propose that the unergative/unaccusative clifferencecan be explained using Hale and Keyser's (r9g3) syntax-in-rhe-rexicon model, couched within Government and Binding Theory in which important aspects of lexical-conceptualstructure are mirrored by syntactic structures within the lexicon. unergative verbs are syntacticallycharacterized(among other things) by having an external argument, but no direct internul u.gu--"nt, while unaccusativeverbs have no external argument, and a direct inon-clausat, non_ PP) internal argument. Due to such lexical properties, transitive and passive structures, including those in reduced relative clauses, which are derived from inherently unergative verbs are significantly more complex than those derived from unaccusative verbs "in terms of number of ,rod., u.ra ,rrrr_ ber of binding relations, and in having the embedded comprement structure,, (stevenson and Merlo, 1997:364).when these linguistic assumptions are implemented in Stevenson's(r994a, b) competitive attachment parser, a kind of symbolic/connectionist hybrid, it turns out that the parser cannot activate the structure needed for a grammatical analysis of reduced relatives headed by passiveparticiples with unergative verbs, "becauseof its limited ability to project empty nodes and to bind them in the structure" (Stevenson and Merlo, 1997:397).Hence, the parser is viewed as confirming the earlierjudgment data, namely that there are "sharp distinctions between unergative RR llauses and RR clauseswith other verbs" (p. 396). In contrast to previous structural theories which attribute the difficulty of reduced relatives solely to category-level syntactic complexity differences, stevenson and Merlo propose that rexical constraints pray a central role in determining the processingdifficulty of reduced relative clauses.However. in contrast to constraint-basedmodels, they argue that differences among crasses of lexical items are due to differencesin structural complexity associatedwith their lexical structures.They argue that reduced relatives with participles based on unergative verbs are uniformly difficult to process,regardless of factors such as frequency and plausibility, that is, "structural compiexity alone can cause failure to interpret a sentence,evenwhen all other factors would help its correct interpretation" (Stevensonand Merlo, 1997:392). If correct' Stevensonand Merro's claims would have a number of important implications for theories of sentenceprocessing.First, they would provide the clearestevidenceto date for structural complexity effectsin sentenceprocessing,due to the internal syntactic structure of words, thus helping to resolve 2\ j8 Reducedrelativesjudged hard requireconstraint-basedanalyses 259 Ilarra Filip et al. cotrtroversyinthe freld.Second,since there are both syntaca lolrg-starrdilrg distinction,Stevenson asPects of the unergativelunaccusative lic and serrrarrlit,l alr,l Melkr's rcstrllsrvottlclstrongly support an approach in which syntacticcorrcl;rlcsof scrrlrrrticdistinctionsplay the primary causalrole in accountingfor vali;rtion in processirtgldifficulty. ln this r'lrirplql rve evaluate Stevensonand Merlo's claims in light of ad<liti,rnalcrnpirical data and modeling within a constraint-basedframework. Sc,.tion 2 prcserrtsthe resultsof a questionnairestudy which replicatesStevenson irncl Merlris fincling that reduced relativeswith passiveparticiples derived frorn uncrgative verbs are, as a class,more difficult than reduced relativeswith passiveparticiples basedon uuaccusativeverbs.However,the resultsalso show that there is a consitlerableoverlap in the distributions of acceptabilityjudgdifliculty, as would be expectedon a constraint-basedacnrents atrtl 1'rrrrsing count. Sectiotr3 shorvsthat the processingdifficulty that is due to the unergadifferencefalls out of a computationalimplementation of a tivc/unaccusativc corrstraint-lrasctlmodel, using only those constraints that recent constraintb;rsccltlteoristsltirveclailned account for processingdifferencesamong reduced rclativcs (Macl)onalcl et al., 1994; Thnenhausand Trueswell, 1995). Thus the differencedoes not require appeal to structural comutrcrgative/rrtact:ttsative plcxity differelrces.In section 4, we argue that a semantic approach based on tltclnatic roles presents a promising alternative to the syntax-in-the-lexicon 'l'lrc tlrentatic properties, which characterizethe two fuzzy cluster apgrroach. c()n('eplsProto-AgerrtanclProto-Patient(Dowty, 1988,1991),can accountfor a y,rcatdeal ol'pr'ocessingdifferencesbetween sentenceswith reduced relative verbs, [irserlotr tutergativeverbs,on the one hand, and On unaccusative cl:rrrses qrr thc ollrer lrarrd.One advantageof this novel way of looking at the gardenp;rtlr phcrrorrcrrorris that it allows us to understandthe similaritiesbetween [lrr.st.two lypcs of sentencesin exhibiting clear gradient effects with respect kr acceptabilityjtrcignrentsand parsing difficulty that are influenced by the lcxit.alsernanlicsof the mair.rverb in the matrix clause.The influenceof the rrr;rirrpretlicirlr:itr a sentenceon the magnitudeof the garden-patheffecthas so firl y.oncrrrrrroliceclanelit is problematic for structure-basedaccountsthat assrrn'e cilhcr <irtcgory-levelsyntactic complexity and/or complexity associated willr word-irrlerttallexicalstructure of verbs.We also show that a constraintthesesemanticnotions can be naturally embedbirs"dapl rcluchirrcorprorating dcrl withiu rccenl collstraint-basedapproachesto grammatical representation' Wc conclucleby tlescribing a rating study that showsthe effect of the main verb on the processirrgdifficulty of whole sentenceswith reduced relative clauses,as is predictcdby rttrr linguisticanalysis. 2. Gradient effects we observed that sentenceswith reduced relatives based on unergative verbs, including manner of motion of verbs, manifest a considerabledegreeof variability in acceptability,and, in fact, perfectly acceptablesentencesof this type are easy to find. Examples are given in (3). At the same time, some reduced relatives with unaccusative verbs are relatively hard, such as those in (4). (3) a. b. c. d. The victims rushedto the emergencyroom died shortly after arrival. The pig rolled in the mud was very huppy. The Great Dane walked in the park was wearing a choke collar. The prisoners paraded past the mob were later executed.a (4) a. b. c. d. The theatredarkenedfor the movie frightenedsomepreschoolers. The Klingon disintegrated during the battle had launched a rocket. The solution crystallizedin the oven burned a hole into the petri dish. The plasterhardenedin the oven crackedwith loud popping sounds. In a questionnaire study we had twenty-four University of Rochester undergraduates recruited in introductory courses use a five point scale (l = very easy,5 = very difficult) to rate the difficulty of a mix of sentencesthat included reduced relative clauseswith inherently unaccusative and unergative verbs, as well as transitive and passivemain clausesentencesusing the same verbs. The full set of materials used in the rating studies are availableby requestfrom either of the first two authors. Table I presentsthe mean ratings. There was a significanteffectof constructiontype, Fl(1,23)=62.00,p<.01; F2(2,32)=82.O2, p<.01. Reducedrelativeswere significantly harder than passivesor transitives, regardlessof verb type (all planned comparisonswere significant at p<.01). We replicated Stevenson and Merlo's finding that sentenceswith reduced relatives headed by passiveparticiples based on unergative verbs are harder to process than sentenceswith reduced relativesheaded by participles derived from unaccusative verbs. For reduced relatives with passive participles derived from unaccusative verbs, the mean was 2.95; and for those with unergative verbs, the mean was 3.45. This difference was reliable in the analysis by subjects, F(1,23)=5.51, p<.05. However, there was substantial overlap in the distributions, and in fact the difference between the unaccusativesand unergatives was only marginally reliable in an item analysis,F(1,32;=3.15, p=.085. Four of the eighteen unergative verbs used as passiveparticiples in reduced relatives were rated as yielding sentenceswith reduced relatives judged easier than the mean rating for sentenceswith reduced relativesbased on unaccusativeverbs. The sentenceswith these verbs are in (3) above. ln addition, some sentenceswith Reducedrelativesjudged hard require constraint-basedanalyses 26 16o I larral;ilip et al 'lrrlrle l . fr r r l l i c t<l l i l l i c u l t yo f r e d u c e drelatives Transitive AA .----f | tl o ooo Trans Pass 2.95 1.63 1.60 3.45 ',t.52 Class O Urracr:rrsatives A [JnFrqatives o oo A A o ooooo Interpretive Nodes \lHl t=t 'l: 1.81 re(luce(lr-elativesheadedlry passiveparticiplesderived from unaccusativeverbs \vcrerate(las nl()r'eclifficult than the mean rating for sentenceswith unergativebaseclrcclucc<llelatives(3.45). Six of the sixteenunaccusativeverbs fell into this the sentences in (4). cntegory,inc:lnclirrg fb surnrrrarize,the ratings showedthat sentenceswithunergative-basedretlrrceclrelativcs\vereon the whole more difficult to processthan sentenceswith rrrraccusrlivr.'heserl reclucedrelatives,but also that there was a considerable rlt'grccof'ovcrlap betweenthesetwo types of sentenceswith respectto the pro- Intransitive .l ^l I lt ^: ll lol u N t(ol t-l @) l@) E o F cessingdilli<'rrlty.'l'heoverlapin the distributionsand the continuum of difficulty is pr oltlt'nrrrlicl'or an account in which the inherent structural complexity vo bs preclicts"sharp distinctions between unergativeRR clauses oI nner11:rtivt' autl llll cfarrscsrvith other verbs" (Stevensonand Merlo, 1997:396).They do definitive evidenceagainstsuch a proposal, however,berrot, holvcvcr',1rr.,rvi<le cirusenleasurt'rncrrtel'ror or other differencesamong materials could lead to ovcrlaf in tlrc tlirtaevenif the underlyingdistributionsdid not overlap. j. model A conslrairrt-trasecl We irnplernctrtetla constraint-basedmodel using the integration-competition arcfiitccture tlevelopedby Michael Spivey and applied to reduced relativesby McRae et al. ( 1998) ancl Spivey and Thnenhaus( 1998). In this model alternative syntacticslrLrctrlrescornpetewithin a probability spacewith multiple conslraints prrovidirrgprobabilistic evidencefor the alternatives.This model is not a firlly ilrrplerncntctl parser;rather, it is an architecturefor predicting the diffiapculty of arnbiguityresolutionusingprinciplescommon to constraint-based 'l'he qrrestionwe addressedwas whether an unergative/unaccusative prrraches. Figure l. The Integrationand competition model usedin the current simulations. Eachvertical rectanglerepresentsthe input ofa particular constraint.The horizontal rectangalrepresentsthe output of the model at any point in time for the three integration nodes for the embeddedclause:the active transitive,active intransitive and passivein a reducedrelative.Constraintsfor tense,voice,transitivity and thematic fit were introduced at the embeddedverb. The PP constraint was introduced after the model completedrycles(i.e., until the dynamic criterion was reached)for eachinput at the embeddedphrase.Similarly,the main verb constraintwas introducedafter processingwascompletedfor the PP,The weightsshown are thosethat wereusedwhen a constraintwasfirst introduced,i.e.,beforenormalizationat the next input. difference would fall out of such a model using just those constraints that have been previously identified in the constraint-based literature. Figure I presentsa schematic representation of the model. In the model, three constructions competed, beginning with the first verb in a sentencewith a reduced relative clause:NP V(-ed) PP V. The constructions were: active tran- iz Reducedrelativesjudged hard require constraint-basedanalyses z( I I;rrralrilip cl al. McRae and colleagues(cf. McRae et al., 1998). Ratings were collected using a five point scale.Questions we used are here exemplified using the verb melt as 'How an example: common is it for ice to melt someone or something?' (Active Transitive),'How common is it for ice to melt?' (Active intransitive),'How common is it for ice to be melted by someone or something?' (Passivein RR). We tested the model on six unergative verbs, danced, raced, paraded, rushed, mqrched, hurried, and on four unaccusative verbs, dissolved,cracked,hardened and melted. This small subset of verbs represents those for which we had corpus counts, difficulty ratings and ratings for thematic fit. Table 2 presents the biasesused in the model for each of the four constraints that applied at tbe -ed and passivein a reducedrelative.The full passivewas sitive,activeirrlrirttsitive, -rrl verb form becauseof the absenceof a preceding copula, ruletl out- at the a n , l t h u s\ \ ' a sl r o t i n c l t t c l e d . J'|e conslrairrtsused were those identified by MacDonald and colleagues (e.g., MircDr)nirl(let al. (199a)) and by Tanenhausand his colleagues(e.g., 'l-t-tteswell, '[hncnhaus 1995).The followingfour constraints came into play anrl at tlre ,erl verb firrtn: ( | ) The frequencywith which a verb was used transitively or intransitivcly; (2) the frequenry with which it was used in tensedvs. tenseless constnrctions; (3) the frequency with which the -ed verb form was used in rhe passiveand aclive voice, and (4) the plausibility with which the first NP could fugcti1rtras sttt)jectof an active transitive, subject of an intransitive, and subject of a prrssive("thernatic fit"). An additional frequency constraint came verb form. Table2. into play at lhe I'l), atrclanother at the main verb. In the intcgration-cotnpetition model, eachconstraint provides probabilistic:strpport for the sytrtacticalternatives.The normalized bias on the constraint is rrrrrltiplieclhy the weight assignedto the constraint. The weights of all the corrstraiutsapplyl"* at a given input are normalized so that they sum to 1.0. 'l'lrr. rnotlclrvorlls iu tlrreesteps.First the biasesare multiplied by the weightsto Word delt.rnrinellrc cvirleltce(activation)eachprovidesin support of the competing irrtr.rplelatiorr(irrteglation)nodes.Activations are summed at each integratiorr rrotlc. Scr,rrrtl,feedtrackto the constraints is provided by multiplying the prolrx[rililyol'crrlr integrationnode byits weight and adding that value to its 'l prr.viousllias. lrinl, the biasesfor eachconstraintare then renormalized.The rrr,,,lclcontilrncscycling until a designatedcriterion; the criterion is lowered alter each cycle. (lr0r cletails,see McRae et al., 1998; spivey and Tanenhaus, l9,r[].) Wlrerr lhe criteriolr is reached,the model moves onto the next region q[ tlrc text, in tlris casethe PP.The new constraintprovided at the PR namely, strorrg evi<lerrcef<rr either an intransitive or a passiYe'was assigneda weight of 1.0,follorvirrgthe procedureused in McRae et al. (1998).All of the weights lvert- tfiep relorrrraliz.ed,resulting in a weight of .5 for the PP and '125 for equal weiglrt ol'.25. Biasesfor transitiviry tense, and voice were determined flrrtrr c<tr'pusirrrall'sesusing the ACLiDCI corPus' comprising the Brown corpus rrrcl 64 rrrilliorr rvortls of the Wall Street Journal that were kindly provided to us lry Paola N4erloancl SuzanneStevenson.The biasesfor thematic fit were {etr:r.urinetllry typicality ratings collected using the procedure developed by Bias lntransitive RR Tense Thematic Fit Transitivity Voice 0.31 0.12 0.37 0.41 0.31 0.38 0.45 0.41 0.38 0.50 0.18 0.19 Danced Tense Thematic Fit Transitivity Voice 0.40 0.2r 0.15 0.43 040 0.s6 0.77 0.43 0.21 o.24 0.08 0 .l 4 Dissolved Tense Thematic Fit Transitivity Voice 0.l6 0.17 0.50 o.2l 0.16 0.43 0.25 o.2l 0.68 0.41 0.25 0.58 Hardened Tense Thematic Fit Tiansitivity Voice Tense Thematic Fit Transitivity Voice 0.05 0.21 0.43 o.23 0.32 0.3t 0.39 0.34 0.45 0.22 0.06 0.49 0.15 0.05 0.49 0.36 o.23 0.32 0.35 0.42 o.34 0.45 0.43 0.91 0.49 0 .l 5 0.91 0.30 0.21 0.55 0.37 0.34 0 .l 9 0.31 0.09 0.35 0.03 0.01 0.71 Marched Tense Thematic Fit Transitivity Voice Melted Tense I i I Transitive Cracked Hurried tense,voice, tlretnatic fit antl transitivity. The same procedure for normalizing weights was firll<xvedwhetr the model moved on to the main verb' IJecauserve tlicl 6qt have an independently motivated way of setting the weiglrts glt the lirur constraints at the -ed verb form, we assignedeach an Constraint Reducedrelativesjudged hard requirc constraint-basedanalyses 2,65 t 4 I h r r ; rI ; i l i l ' c ;l r l 'lhlrlc (corrtitrtttrl) 2. W<rlrl ( onstt'airrt Transitive few,'lry Mclrcd Wilth 'l l r c r r r a tFi ci t 'li rrrisitivity Voice 'l i'rrse 'flrenratic Fit ltansitivity Parirrled Vrice 'l t'rrse 'l'hernatic l { a <r ' d Itrslred 0.16 0.34 0.31 0.28 0.t5 0.35 0.34 0.28 0.28 0.15 o.32 0.49 0.28 $.25 0.28 0.55 0.31 (r'r< Fit 0.26 liarrsitivity 0.30 0.3I \i rice 'l Bias lntransitive i'rrsc I l r r r r r a l ilcr i t li;rrrsitivity 0.50 0.10 0.05 \ / rr i c e li'rrse 0.50 0.40 l l r c r r r a t iFc i t 'li ansitivity \/oice 0.26 0 .l 4 0.44 Q.49 0.50 0.45 0.93 0.50 0.40 0.33 0.80 0.44 RR 0.53 o.t7 0.44 0.71 0.33 0.17 0.44 0.s0 0.46 0.15 0.39 0.01 0.45 0.02 0.01 0.20 0.41 0.07 0.t2 I I , i 'Jhble 2, unergativeverbs tend to be used more often than As crrn be seerrfrrlrrr vetlrsitt intransitiveconstructionsand lessoften as passives'For runrccusative of an turergativeverlrs these factors mean that the active intransitive reading 'Nl, V(-erJ) I)I)' li.agtlent will be more strongly biased relative to the reduced rel:rtiveclattsclertlitr g. In or{cr to cv;rlrratet|e output of the model, we considered three measures.'l'[e {irst was the total number of cyclesuntil the criterion was reached at f at tlre ruaiu vellr (cyclesat the -ed verb form, + cyclesat the PR cycles the rnain verb).'l'lreseconclwas the probabilityassignedto the reducedrelative take structure at thc rrrail vertr. The third was the number of rycles it would verb' main the the rno{el to assigrrthe reclucedrelative a probability of .9 aI lVc asstrrnedllrlt each of these lneasuresshould correlate with the difficulty Ill three measurespredictedthat as a classreducedrelatives ol'tlre scntcrrt.rl. deriveclfrom unergativeverbs would be more diffiwitlr passivePat.ticiPles cglt tfiarrrc.lrr,,',1relativeswith passiveparticiplesderivedfrom unacccusative vetbs:firr total rttttrtberof cycles,t(9)=3'16' P<'01; for probability at the main all verlr t(9)=1.95, p- .()2;anclfor cyclesto a criterion of '9, t(9;=Z '99' p<'02' teststwo-tailed. The model also correctly predicted some gradient effects.For example, the reduced relative beginning with The witch melted.. . was correctly predicted to be harder than the reduced relative beginning with The jewelry melted... . In addition, the reduced relative with paraded was predicted to be easier than the reduced relatives vnth danced, raced or marched. However. The victims rushed to the hospital died was incorrectly predicted to be quite difficult even though it was rated as fairly easyby subjects. It is important to note that the model we presentedis incomplete in important ways. There are constraints that are not included and as a result the model generally overestimates the availability of the reduced relative analysis. Moreover, we were working with only a few verbs for which we had data. Nonetheless,it is clearthat the processingdistinction betweenreduced relatives headed by passiveparticiples derived from unergatives and unaccusativesfalls out of a small set of constraints, primarily verb-based frequencies,that have been independently argued for by proponents of constraint-basedmodels. In the light of the results we reachedso far, a proponent of the syntax-inthe-lexicon approach might appealto two types of counterarguments.The first might be that frequencies reflect the unergative/unaccusative distinction; however, the structural complexity associatedwith the lexical structures of these two classesof verbs results in those frequencies and actually plays the causal role (but cf. MacDonald, 1997). The second argument is that the syntax-inthe-lexicon approach implemented in stevenson'sparser is superior becauseit presupposesa full-fledged linguistic theory, namely, Government and Binding Theory, whereasthe constraint-basedapproach is not supported by independent Iinguistic assumptions in a similar way. In the next two sections we address these issuesin turn. First, we explore and motivate the claim that among the linguistic variablesresponsiblefor the processingdistinction a crucial role is played by semantic variables, rather than just syntactic variables. Second, we show that the ideas implemented within our simple model are broadly consistentwith recent constraint-basedgrammars, most notably HPSG. 4. { t: The linguistic basis of unaccusative/unergative distinction in processing Our primary observation,and one that has so far gone unnoticed, is that both types of sentenceswith reduced relatives exhibit similar gradient effects in acceptability judgments that are crucially influenced by the lexical semanticsof the main verb in a matrix clause.To put it in the simplest terms, the fewer Reducedrelativesjudged hard require constraint-basedanalyses 267 6 6 F l a r r aF i l i P e t a l argument position is unfilled, and it can be filled by an 'external cause' argument, when they are used transitively. This does not hold for unergative verbs, because they already have one external subject argument. By this test,yellow in (lOa) and solidifu in (lOb) are unergative,while hardenin (lOc) and.yellow in (10d) are unaccusative.(Examplesin (9) and (10) are taken from stevenson and Merlo, 1997:365.) asagelt-lil<e pr()l)clties ancl the more patient-like properties the main verb clause relative reduced a with sentence the easierthe whole sigls to its srrfijet:1, isludgctl.'l'his itlea rvill be discussedin detail in Section4.2, but let us illustrate it here witlr a fcrv exarlples. In (5a) the su[ject of complained,the patients,is is a volitional agclrl in the denoted event, and we seethat the whole sentence undergoes lessaccepta$lellras (5b) with tlied as the main verb, whose subject a clrangeo[ statc.A sinrilarcontrastcan be found in (6): (s) 'f 'lre to the nurse' a. lratientsrylhgd to the emergencytoom#com,lained 'l'f died. room emergency to the rc patientsrushed lr' (6) a. b' (e) a. #The candy caramelisedin an hour burned. b. #The wax solidified into abstract shapesmelted. c , #The paper yellowed in the sun shrank. ( r 0 ) a . #The chain-smoker yellowed the papers. b. #The sculptor solidified the wax. l'frc ( ireat l)ane walkedin the park#tug:rcdat the leash' 'lrc ( ilert I)anewalked inthe parkworea chokecollar' f c. The sculptor hardened the wax. d. The sun yellowed the paper. Sirrrilarlyirt rr',lrrceclrelativeswith passiveparticiplesderived from unaccus a t i v c v t ' r ' l r s , s r r r l rtal nsr k e n e r i i n ( 7 ) , w e s e e t h a t t h e u s e o f f r i g h t e n e d a s o p p o s e d irr tlre rnatlix clauseis correlatedwith a differencein the acceptability t, srrtelletl 'I'l're reasonis that frightened,but not smelled,presents tr{'tlrc rvlrrrlcsqrrlettce. in tlrc strlrjcct tlrc thentreas the causeof the change of the psychologicalstate are examples similar Other preschoolers. tlrc refcrent ol'the direct obiectsome g i v e ni r t ( t t ) : (7) a. b. (S) a. b. 'l'lrc #frightenedsomepreschoolers' theatre darkened for the movie thcatredarkenedfor the movie smelleQlikepopcorn. 'l'[rc 'l'hc getresn.rutatedin the experiment#attackedtheirhost' 'l'lre genesmutated in the experimentwereusedin a new vaccine. (5)-(8) resist N{ost irnportailtly, different degreesofacceptability observed in in the couched arr explanatirrrrin structure-basedterms aSwell as explanations that the syrrtax-irr-thc-lexiconapproach of Stevensonand Merlo (1997). Recall inherently by headed relatives reduced with lirtter pretlict tlrat n/l sentences 'rrersrrtivcv<'rl'sare predictedto pose'sharp difficulty'(p. 392) for an intergrammatical analysisbythe parser.In l)rclcr, a'{ tlrci,car'rot [e assigneda reduced relativesthat are not easyto .r.tfer lrr ifcroulrl firr r.rnaccusative-based as thosein (9), Stevensonand Merlo resortto the semanticdisirrterprct,srr<.lr 'external 'itrternal causation' (see I-evin and causation' and tirrctiolt lrt.lrvr.crr are unergative.According they that I lov:rv, 1995:210-211)to argue ILaplratrrnrt 'internal causation' in trr tlrenr, vcrlrs like (orotnelise,sotidify and yellow entail from tlrcir serlatrtic <lescription,a feature that distinguishesunergative verbs unac(see Since ibid.). caused' Irlaccusativeorres,the latter being'externally cusativeverlrs ltave one internal direct object argument' the external subject i I i ii ,!l The problem with this test is that unergative verbs, including agentive manner of motion verbs, when used transitively require their subject argument to be an Agent: cp. *The explosionjumped the horse vs. The jockey jumped the horse. (This observation was made by cruse, 1972;lackendoff, L972;Levin ancl Rappaport Hovav, 1995; seealso stevenson and Merlo, 1997:357and footnote 4 below.) This inconsistency clearly indicates that a test based on the possibility of the overt expression of an Agent argument cannot be the right diagnostic for deciding the membership of verbs in the unaccusative and unergative class. The main source of confusion stems here from correlating'external causation, 'possibility and of an overt expression of an external agent', on the one hand, 'internal 'prohibition and causation' and against an overt expression of an external agent',on the other hand. what is lacking is a precise characterization of 'internal the notions causation' and 'external causation', introduced by Levin and Rappaport Hovav (1995), and the motivation for the correlation of these semantic notions with the syntactic structures associatedwith unergative and unaccusativeverbs.Moreover, ( l0a) is claimed to be lessacceptablethan ( lOd), becauseits subject referent may be intentionally involved in the denoted event, while in (l0d) the denoted changeof stateis "indirectlybrought about by some natural force" (p. 365). However, it is not shown how such a fine-grained distinction between '(volitional) Agent'and'natural force',and the suggesteddifference in acceptability judgments, can be viewed as being correlated with the external subject argument in the caseof unergative verbs, and with the internal object argument in the caseof unaccusative verbs. Reducedrelativesjudged hard requireconstraint_based analyses z 2 6 8 I l a n a l r i l i pt t r r l 'l'he faci lh;rl Slevettsonand Merlo do resort to rather subtle semantic crileria irr orcler.to accoLlntfor difficult casesis instructive, becauseit shows that explarrationsirr tenns of categoricaldifferencesbetween syntactic configurations in tlre lexicqrr are insufficient. Indeed, one may ask to what extent syntactic lhc:torsal e necessaryin addition to semanticones in order to account for the garden-1)atlrl)henomenon. If we focus on the differential semanticsof the vertrsirt tlrc rrr:ttetialtliscussedhere' we can begin to acount for the overlapol'setrtences with reducedrelativesaswell as the greatdeal of Iring rlistrilrrrtirrtt vrrrialrilityrvillr r espectto how good or bad they arejudged to be, leavingopen llrc tprcs(iotr trl.rvltatrole, if any,a word-internal syntacticdifferencesare left to pl:ry.\Vc nr)w lurll tcl characterizingthose semanticconstraints more precisely' 4 . r . ' l ' l r e n r a t i ,l.' r o t o - l l o l e s 'l'he idea tlrat ar.gurnentpositions ofverbs are associatedwith certain "thematic 1rles" (()asclltrles,CaseRelations)such asAgent,Patient,Instrument, and so forth, has rcccived varying characterizationsin the linguistic literature' Here, however,wc lirllorv the analysisof David Dowty (1988, 1991)' who proposes tlrat the only tlrcrnaticrolesare two clusterconcePts,Proto-Agentand ProtoPaticr-rt,each rharacterizeclby a set of verbal entailments, given in (11) (see Dowty, 199| : i,t72)." IAl n argument of a verb may bear either of the two proto(or |otlr) t11valying {egrees,according to the number of entailments of r<.rles cach kind lhc verb givesit" (Dowty, l99l 547). : (11) a(tlrrilptllC"roperties for the A state or the event in a. volitional iuvolvement r erception) h . ( ( ' n t i c n c(ea r r d / o p (:. c;rrrsittgalt evelltor changeof statein anotherparticipant (1. rtoveltlettt (relativeto the position of anotherparticipant) ( c. t eferetttexistsindependentof action of verb) (irlrtr ilttrtirlSpropertiesfor the PatientProto-Role: charrgeo[ state trrrrlergoes h le m e lr. i r r c r c r r r e n t a .;rrrsally allectedby anotherparticipant C. (l. stirliorraryrelativeto movementof anotherparticipant (c. 'l()esrtot existindependentlyof the event'or not at all) il. The Argument SelectionPrinciple determines the direct associationof clusters of Proto-Agent and Proto-Patient properties with grammatical relations in a many-to-one fashion: (LZ) fugument SelectionPrinciple (Dowty I99l:576) In predicates with grammatical subject and object, the argument for which the predicateentails the greatestnumber of proto-Agent properties will be lexical_ ized as the subject of the predicate; the argument having the greatest number of Proto-Patient properties will be lexicalizedas the direct obiect. 4.2 Compatibility between subjects in sentences with reduced relative clauses In reviewing the contrasts found in examples,such as (5)-(g), it appearsthat the following is a reasonable description of one effect of the main verb on a reduced relative clause: (13) Hypothesis The acceptabilityof sentenceswith reduced relative clauses,headedby passive participles derived from unergative and unaccusative verbs, increaseswhen the passiveparticiple and the main verb of a matrix clauseassigntheir subiect-Nps more Proto-Patient,and fewer Proto-Agent, properties. The intuition behind the hypothesis ( l3) is that sentencesare easierto in_ terpret when there is an internal coherenceamong the interpretations of their constituents.one way this coherencecan be achievedis in terms of compatible assignmentsof thematic properties to different Np arguments that are associated with one and the sameparticipant in the domain of discourse.In sentences with a reduced relative clause the internal coherence depends in part on how well the thematic make up of the subject Np in the matrix clause matches the thematic make up of the PRo-subject of the reduced relative clause:namely, the passive participle in the reduced relative requires that its pRo subject be a"very good" Patient. Let us take (la) #The horse raced pastthe barn felt. At the point wlten raced is processed,the preferred syntactic-semanticpattern is that of the main clause with an agentive subject-Np. However, when /e// is processed,raced must be understood instead as a passiveparticiple. passive participles typically presupposethe existenceof corresponding active transitive verbs whose subjects correspond to active direct objects (seesag and wasow, 1997:164, for example; however, passivesubjects do not always correspond to active direct objects,seeZwicky, l9g7; postal, 19g6, and others). Let 70 Reducedrelativesjudged hard require constraint-basedanalyses 2., lIan.r Fil(r et al properties by the verb racedin its rrs frow lcrok at the assignmentof thematic (lexicalcausative)use. ('PA standsfor irrtr.nsitive(rrrr.gativei and transitive 'PP' f o r P r o t o - P a t i e not n e s ' ) lie | ) r o t o - A p , t ' nI II1 1 1 r e t sa n d ( | 4) other participant', and given that thehorse in (15) is a sentient being with a (potentially) certain volitional involvement in the racing event, 'movement' can be here taken as the Proto-Agent property. (This is not uncontroversial. However, feII does not assign clear Proto-Patient properties to its subject ei'undergoes ther. A candidate might be a changeof state',but here it would not mean a permanent change,rather just a change in bodily posture, and hence ultimately'movement'.) Hence, the thematic make up of the subject Np in the matrix clausedoes not match the thematic constraint of the reduced relative clause which requires that its PRO subject be a "very good" patient. past thebarn' 'l'lrt' lrorscI,,\C):D pastthe barn' The rider MCED the horte I !r\ ( I vt'litiotr) Iscnti{lllce -l ilt()v('lllelrt I PP PA and PA causallY + (+ volition) + volition + sentience affected + sentience + causing change + movement "usual" transitive in that it semantiA t.ausativclirlrrr of au ttnergativeis not a prototypical transitives cally cleparts[Krrrr prototypical transitives.Intuitively, 'billiard ball model" as Langacker(1986) calls can be utrclcrsto.d in term.sof a an asymmetric and unidiit, which involvt's tlvo participants that interact in (possibly ,r'. .rfih.n', is directly affectedby some action rectiorralrviry,rvlrereLry instigated or causedby the involving tttovcttteltt, cotltact, effect, and the like) the direct object has many other participunt. h'r l)owty's terms, tl-rismeans that pr'tolPatie't (a.cl a few Proto-Agent) proPerties, and the subject has many A typical unergative verb Proto-Agertt (arrd a ferv Proto-Patient) properties' becauseits usecltrarrsirivelycloesDot lrt the semanticsof a transitive PrototyPe, (14) example our "good" Agent: in dilcct object hrrsa therl-raticmake up of a of the to the object tlre strbject tlrehorse<;f the intran sitiveraced corresponds proto-Agent properties. At the same, tfte trrrrsiti'e rtrrcrl a'rd tlrcy share three affected'by the transitive hr|..seis assigrretlorre lrroto-Patient property'causally use of unergativeverbs rncrd.,l'ltealvkrvardnessoften related to the transitive these two different roles or tDaybe sccil ils slcrrrnrirrgfrorn having to reconcile (an Agent-like and a Patient-like) on one and the trvo cli{lererrtl''-'r-sirectivJs This carries over to passarne|articiParrt in the denoteclcomplex eventuality. verbs' The reason is that a siuc particif ies <letivetlfrom inherently unergative to have a high number of prototypicrrl |assive cQnstructionrequiresits subject an unergative.verbsupplies Pflrto-Patierrt prrrperties,yet a passiveparticiple of properties,given that a subject argunrctrt that carriesa number of Proto-Agent verb (The rider raced it correspondsto the direct object of an active transitive of the active intransitive tlre lrorse),rvhiclrin ttlrn corresPondsto the subject in (15) we seethat the verb (The horse rnt:ed).Toreturn to our lead example' as the plto subject of the passiveparticiple has the same thematic properties a "good" Patient' The main correspouditlg actrveobject in (14), hence it is not verbfellassignstheproperty.movement'toitssubjectthehorse.Insofaras 'movement relative to the position of anthis can bc irrtcrprctetl itr terms of (15) Thehorsei I [< PROi > RACED past the barn] fell. I PA PA and + movement (+ volition) PP + causally affected a sentience + movement lf, on the other hand, the main verb of a matrix clauseassignsProto-patient, rather than Proto-Agent, property (or properties) to its subject, the magnitude of the garden path effect is diminished, as (16) shows: the subject of died is clearly a "better" Patient then the subject of fell, as it is entailed to undergo a permanent changeof state.Hence, (16) is somewhat easierto interpret than (15). (16) Thehorsei I I<PROi > RACED pastthe barn] died. I I PP PA and PP + undergoeschange (+ volition) + causally affected of state * sentience + movement of course,not all transitive and passiveusesof inherently unergativeverbs are odd. Other factors, such as expectations related to the occurrence of highly conventionalized combinations of words and general world knowledge, may come into play and oyerride the semantic mismatch described above. For example, John walked hk dog and Fido was walked by lohn tonight sound highly natural. Let us now look at sentences with reduced relatives headed by passive participles derived from unaccusativeverbs. In (l7a) the subject ofthe un- $ accusativemelted, the butter, corresponds to the object of the active transitive melted in ( l7b), they are both entailed to have at leasttwo Proto-patient prop- 7, FfarraFilip et al '(lrrlrl',c 'Incremental of slale' and Theme'.Hence,they are "very good" cl tics: [)at ic rrts,arr<I rvr'(.;rl I expect that both the transitive and passivelusesof n'telted i l l r ' P c 'l1r ' .1. 1 r; 't rr r ' P I a l 1 l e . (.17) :r. 'l hc lruttgl Mlltl'ED in the pan. 'l'lrc r r'ol<IUEUI'EDthe butterin the pan. b. ( l8) a. .l'lrcllqltg1MELI'EDin the pan wasfresh. b. # !'!Lebu{ql MEUIED on the stovedripped onto the kitchen floor. As is prcclictcrll,y the hypothesisin (13), (18a) is judged easierto processthan (l8b). (lSb) corrtainsthe matrix verb drippedthat entailsthat the referentof its subjectargunrentntovesrelativeto the position of another participant,and berrcecan be vioved as entailing one Proto-Agent property in its subject argunrcnt.This, lrolvcvcr,is inconsistentwith the requirementstatedin our hypothesis(13) that tlrc subjectNP in the matrix clausematchesin its Proto-Patient properties the lhernatic rnake up of the PRO-subject of the reduced relative clarrse.(l [ta) contains tlre stativepredicate befresh in the matrix clause,which entrrilsrro Prolo-Agentpropertiesin its subjectargument, and hence(18a) is n l ( ) r ea c c e l ) t u l rtl h calr( I8b). 4.1 An tll'S(; it1'proaclr Irr the pasl tcn ycars or so tlrere has been a growing convergenceof results and rrrclhoclokrgicalirssrrnrptionscoming from psycholinguistics and theoretical irr llrl rlornain <lf constraint-basedapproachesto natural language lirrrlrristics tk'sr.ription(r'.ri.,l\rllarcl and Sag, 1987;Pollard and Sag,1994;Sag,1998,for forareview e r ' r r r r p f e ; s e ea 'rlr c n h i r r r s a n d l i u e s w eIl9l ,9 5 a n d M a c D o n a l d , 1 9 9 7 approacl.res in psycholinguistics). Theysharetwo main asol'r orrstlairrl"lr;rscd interpretation requiressatisfactionof multiple srrrrrptions:l;irsl, a serrtence's (1'r'ssilrly,lif1i'rerrt ialll' weighted)constraintsfrom various domains of linguisli< ;rnclnorr lirrlirristicknowledge.Second,the integrationof such diverseconstrlints is lacilitatedby the information containedin lexicalentries.Verb-based syrrtacticarrd sernanticpatternsprovide a guide for interpretingkey aspectsof sl nrcture and rneaning,whereby semantic constraints often have tlrc senterrce's a l ' r i v i l e g c tsl t a l l r s . 'I'hc lcxicalcorrstrairrt-based approachproposedherehas all the main hallrr:rrks of recerrtvcrsions of HPSG (see Sag, 1998, for example). Assumptiorrs aborrt lexit:alsernanticsof verbs and linguistic information directly associatedwitlr cxtla-lirrguistic context and generalworld knowledge are influ- Reducedrelativesjudged hard require constraint_based analyses enced by Fillmore's work and construction Grammar (seeFillmore and Kay, in press).The grammar assumedhere is monostratal, non-derivational and nonmodular. It is characterized declaratively by specifying types of well_formed linguistic expressions(e.g., words, phrases,part of speech classes,argumenr structure classes,and traditional morphological classes,for example) and con_ straints on those types. Alr properties oflinguistic expressions are represented as feature structures- Language-particurar rules and universal principres are characterized as systems of constraints on feature structures. The main explanatory mechanism is unification in the narrow sense of structure sharing oftoken-identical feature structures (cf. pollard and sag, 1994). Since lexical entries constitute the key ingredient for interpreting the main aspectsof the sentencet structure and meaning, and facilitate integra-tionof diversetypesof knowledge,let us introduce their main features using a simplified lexical entryfor the transitive activeracedin (19): ( ie) .l -itll (19) contains phonological,syntactic,semantic and pragmatic information, encoded as valuesof the feature attributes pHoN, syN, sEM and CONTEXT, respectively.The value of SyN encodessyntactic information required for con_ structing syntactic projections headed by raced.The linking bei*e.n the syntactic (SYN) and semantic (sEM) structure in the lexicon is mediated via coindexation of syntacticarguments and thematic argument slots, and motivated by Dowty's Argument selection principle (here given in (12)). Each argurnenr slot in the thematic structure of a verb correspondsto a cluster proto-Age't of and/or Proto-Patient properties (cf. Dowty, 1991).Thematic argument slots in turn are co-indexedwith individuars in the predication feature structure PRED, which together with 'psoa' (param etrized state of affairs) constitutes the value of GoNTENT. The feature structure pRED capturesthe assumption that verbs semantically expressrelations between individuals. The attributes .racer, and 'racee', which correspond to'frame-specific participants'in Fillmore (19g6) or 'individual thematic roles' in Dowty (19g9), include properties that we asso_ ciate with the individuals 'i' and'j' on the basisof knowing that the statement 27, Reducedrelativesjudged hard require constraint-basedanalyses 275 4 I l a r r aF i l i pc t a l . 'i r:r,r:<lj' is Il.rre.Irr a given single-clausepredication, further semantic restrictiorrson particilxrrrtsare imposed by the interpretationof noun phrases'For exrrrrirlg,'lr.r.,'r'il' rvill be constrainedby the content of the NP filling the 'IllNP'pl:r<'c. l'lllil) does not provide an exhaustiveaccount of all that we l<rrolvalrotrt llre rrreattittgof a given verb. What role an individual plays in a g,ivcnsituirtiotr rlelrettdson a number of other factors,includingworld knowledge,which is errc,rtleclunder'psoa'.(For a related,though not identical' use of 'psrra' 5.. f \ rllald an<lSag,I994; Sagand Wasow, 1997.)I'extcalentries of verbs itrfcrrmationabout the occurrenceof a given verb form also inclrrrle[r'cqtrr,'ncy ;rlrotttits arguntentstructures,and the like. in tlre langrrage, Apart liorrr tlre lexicon, the grammar will minimally include the level of verb forurs arrrl tlre syntactic level with phrasal templates.This is illustrated in a h i g h l ys i r r r p l i l i e tI l; i g u r e2 . VFoRMS ,aced tacecl passive participle raced pas! active participle LEXICON PHRASAL TEMPI"ATES model. outlitreof a constraint-based Figrrre2. A sirnplilietl tylres each level of representationare cross-classifiedin multiple In g,cr-reral, "l inhcritancehierarchiesaccordingto their sharedinformation. (Due to the limitation of space,this is not representedin Figure 2.) The information shared by a given class0f'objects is associatedwith a general type and is automatically passeddorvn frotn the general type to specific members of the class.For exar)rple,I{n(ll.,l)Z antl IIACEDS inherit infonnation from the generic lexical cntry ftrr trnlr.sitiveverbs, here representedby the node Vt. Types direcdy sul)sunted urrtler the salne supertype represent mutually inhibitory alternativcs, which ofterr representmultiple interpretation alternatives and differ in frcrltrencyof occurrettceitr the language.For example, RACED2 (active past terrse) an<lI{ACIil)3 (passiveparticiple) are mutually exclusive,here indicated by the thick slartctl line lretweenRACED2 and RACED3.The activeintransilivr,'rrse<tI'roccrlis ntclte ft'equentthan the activetransitive one. We assumethat srrrlrllctlrrcnt.yirrftrrrttaliouis encodedin the lexicalentriesof verbs. Unification allows us to represent dependenciesand connections within one particular level of representation and also among different levels. Feature structures representing compatible types are unified in a new coherent structure by linking them to a single feature structure, which is shown with straight lines (not all such possible connections are here indicated): e.g., IVFORM PAST.ACTIVEI U [SYN Vi]. Featurestructures representingincompatible types cannot be unified: for example, active verbs cannot be projected into a passiveclause. One advantage of this system is that it allows us to capture the observation that different types of information that characterizethe use of a given word are dependent on each other so that accessingone type ofinformation during sentenceprocessingresultsin accessingothers compatible with it. For example, if the sequenceThe horseraced... is understood as the main clause, the information associatedwith the verb racedwilTbe a complex feature structure comprising the information that this verb shareswith all active past tense verbs. If the same sequenceis understood as the head noun modified by a reduced relative clause, racedwilT be associatedwith the information shared with all passiveparticiples, and due to its passiveargument structure it will also activatethe information associatedwith the active transitive useof race. S. Empirical study of effects at the main verb We conducted a rating study in which we had six subjectscomplete questionnairesin which they made judgments about four of the dimensions that Dowty identified as being part of the Proto-Agent cluster: 'volition', 'sentience','causing an event or changeof state',and'movement'.s The questions concernedthe subject argument of the main verb in the matrix clause.Thus to obtain ratings for The horse raced past the barn died, the subject would rate The horse died. Each simple sentencein the latter set of data was associatedwith four questions designedto illicit judgments about the four main Proto-Agent properties entailed by the verb for its subject argument. Each question was answeredby our subjects using a scale from I to 5. For example, in the caseof 'volition', I would indicate a completely non-volitional participation of the individual denoted by the subject argument (e.g.,The horsedied) and 5 would a fully volitional participation (e.g.,Thepatients complained).Wethen averagedtheseratings to come up with a composite Proto-Agent rating, with 1.0 being the lowest and 5.0, the highest. Subsequently,we selectedmatched pairs of reduced relativeswith different main verbs, e.g., The victims rushed to the hospital complained/died, in which participants assigneddifferent Proto-Agent ratings for the two main Reducedrelativesjudged hard require constraint-basedanalyses ,.76 I [ ; r r r ' lr i l i l ' , ' t : r l . 'l'lrcyiclifits We were ableto idendied vs.The victimsconxplained' vcrl)s(c.tI., We then had tify 2l rrr;rltlretlpairs of reclucedrelativesthat met this criterion. verbs, e'g', main these ,tnotlrcr g,t.t,rrpof subjectsrate reduced relativesusing 'I']rc vit I i'rrs,r,rl,r',1la tlrc lnspital complained/dietlshortly after arrival' 'lhblc 3. llarc,l difficultyfor reducedrelativeswith main verbsdifferingin the ProtorePresent to their subjectargument.Numbersin parentheses assigned AgentPrOperti,:s Merlo'sclaimsthough, this resultis completelyconsistentwith currant constraint-based lexicalistmodels.We alsopresentedan analysisof the unergative/unaccusative distinctionusingthematicrole propertiesalongwith some preliminarysupportingevidence.In future researchit will be important to combinemore sophisticated thematicrole representations into a constraintmodel. basedprocessing l h e r r t c a t tI ' t o l o - A g c l t l r a t i n g Acknowledgments passiveprrticiple derivedfrom u n a c c u s a t i v ev e r b s u l r e r g a t i v ev e r l ) s 2.32 2.81 (t.37) (2.04) 2.50 3.3r (2.35) (3.83) 'l'he the mean tlata are presentedin Table 3. The numbers in brackets indicate argument, ratings ftrr verbs with low Proto-Agent entailments in their subject in their sub.nd th. meirn ratings for verbswith high Proto-Agent entailments a main revealed ratings Arr ANOVA conducted on the difficulty icct rrrg-trntt'ttt. of ,'ffe,t of vetl',.:lass,F(l'zO)=g'so, p<'01,a main effectof the Proto-Agency Overt l r c r r r : r i rvr r : r l r ,l r ( 1 , 2 0=) 5 . 0 2 ,p < . 0 5 a n d n o i n t e r a c t i o n ,F ( 1 , 2 0 ) = l ' 1 0 ' properties Proto-Agent higher ,rll,tlrcrr,r ctlrrceclrelativeswith main verbswith properties' \vcl('ilrr(r (' rliflictrlt than reclucedrelativeswith lower Proto-Agent 'l'rr this section, we showed that the unaccusative-unergative srrrrrrrrariz-e r l i s t i r r t : t i otrlrr a t s t e v e n s o n a n d M e r l o c h a r a c t e r i z e a s a s y n t a c t i c d i s t i n c t i o n c o r clausescan be r.clal.tl rvitlr tlil'liculty ot easeof processingin reduced relative roles' One adI.c-c;tstas a tlistinclioD that concernsthe assignmentof thematic is that phenomenon vaullrgc of tlris llovel way of looking at the garden-path comit all.lvs us t. turrlerstandsomething that has never been systematically a sentence in predicate main the rrrcutetl,rrr |cftlre: nanely, the influence of of the garden-patheffect.The analysisin terms of Dowty's ()n tlte rrragrritu<le predictions here. therrraticr.oles,lbrrnulated in (13), also makes the correct (1988), ,l'hcse Thnenhaus and carlson resrrltsalsg sqpport the claim made by Thnenhaus ,lhrrcnlraus by ancl(larlson ( 1989),and in a number of later studies in languagecomancl his collirlxrrators,that thematic roles play a central role with an is consistent analysis prc|c,si.1. We also showed that our thematic t l y nlotivatedlil guisticmodel' i ncleperttlcn ,lhkerr tt,eether, tlte current work confirms stevenson and Merlo's finddeing thirt scnlel)ceswith reduced relativesheaded by passiveparticiples sentences liie,l {l ,,rrr rtrtergativeverbs pose more processing difficulty than verbs.Contrary to Stevensonand relativeslrasedon unaccusative rvillr rr.,lrr,..,',1 The researchpresentedin this paperwassupportedNIH grant HD 27206ancj by NSFgrantSBR-9511340. Many thanksto MichaelSpiveyfor his comments on a previousdraft of this paper and to SarahScheidelfor her help with the questionnaires. gratefulto PaolaMerlo and SuzanneStevenWe areespecially sonfor helpfuldiscussions and for providinguswith the corpuscountswe used in our model. Notes r. '+' '*' Departmentof Linguistics, Department of Brain and CognitiveSciences. z. The unaccusative/unergative distinction (e.g.,melt vs.mce)wasintroducedby Perlmutter ( 1978),and alsonoticed by (Halt, 1965). 3. According to semanticcharacterizationsgiven by Var.rValin (1990) and Dowry (1991), for example,unergativeverbs tend to entail agentivity in their single argument and to be aspectuallyatelic. Unaccusativeverbs take a patient-like argument and are mostly telic. 4. It might be objected that our examples in (3) are easyto process,becausethey involve complex unaccusativepredicates,rather than unergativeverbs.However, for English at least, there seemto be no convincinggrammaticaltestsfor the unaccusativestatusof the com'unergativeverb + bination directional PP'.(SeeLevin and Rappapport-Hovav,1995:t88 and elsewhere,for a discussionofpossible candidatetests,such as the occurrenceofunaccusativesin the causativealternation.) 'referent exists inde5. Onc of Dowty's Proto-Agent properties was not included: namely, pendent of action ofverb'. It doesnot matter for our analysis,given that the constructions under considerationhavethe samevaluefor this feature. )-73 I [lrra l]ilil, r'l al Itefcrcnct:s Iicvcr', 1'.(i. (l()/0). "'l'lre (irgnitive Basis for Linguistic Structure." In J.R. Hayes (ed.) (.)tgrriti,trr rtrrrlthedevelopnrcnt of language.New York:Wiley. ( larlsrrr,tl.l':1.anrl 'lhrrcnhaus, M.K. (1938)."ThematicRolesand LanguageComprehension.".!, r/(r,rm rd Sunrt tttics,2 I, 263*300. ( lrrrsc,I ).A. ( 197Z). "A N()teon EnglishCausatives." LinguisticInquiry, 3,520-528f )orvty, I).1(.(1979). l\rord Meaning and MontagueGrarnmar.The Semanticsof Verbsand 'l'itrtcs (lerurotit'e irr Senmnticsand in Montague'sPTQ.Dordrecht:Reidel. 'Thematic Role'l'ln (lgtlc)). "Orr the SemanticContent of the Notion of I)orvly, I).11. (llricrclria, (1., Partee,Il. and R. Turner (eds.), Property Theory, Type Theory and Kluwer AcademicPublishers. VolunreII: SemanticIssrres. Senmntics. Mrtrrrrr/Larryunge l)ortJlecht,Ibston, London. I )rrrvly,I ).Ii. ( l99l ). "'Ihematic Proto-Rolesand Argument Selection."Language,67, 547(r19. Department of Linguistics,The lrillnrrrre,tllr.J. (t986). On GrantnraticalConstructior,ts. ms. I lrrivcrsi| ]', rf ( laliforniaat Berkeley. l;ilfnrrrrc, ( llr.l. an<lI'. Kay.In press.ConstructionGranmar. Stanford:CSLI Publications. of Ilalr.,K. ,rrr,ll. Kcyser.(1993)."On argumentstructureand the lexicalrepresentation Theviewfrombuilding,20,53-110. s1,rrt;r<'lir r,'l;rliorrs." In K. tlaleandf. Keyser(eds.), ( ) r r r l ' ri r l u c N . I A :M I ' l ' P r e s s . turdObjectin Modern English.Cambridge,MA: MIT Ph.D.Thesis. I Iall, It. ( l()(,5). ,Srr/rict1 ( Gratnrnar.Cambridge,MA: The rll, It. | 972). SeumnticInterprctatiortin Generative lack,'rr,l, N l ll l ' r c s s . lt. ( I 990). SurrortticStruchres.Cambridge,Mass.:The MIT Press. Iackr.'rrtlo{1. [,arrgircker', l{. ( | gft(;)."'liansitivity, Case,and GrammaticalRelations:A CognitiveGrammar N'ls. Prosl'cr'trts." Semarrtics At the Syntax-Lexical Levirr,It. antl lr'1.I{ap|aport }lovav. (1995). Unaccusativity: IrtttrJrtct.( )arnbridge,Mass.:The MIT Press. and Sentence Macl)rrrralrl, N,l.O. (1997). "Introduction." ln Lexical Represent:alions Britain: Great Processes. Hove, Language and Cognitive A special issue of Prrrrcssirrg. Psy'clrologl' PlcssLtd. M.S.(199a)."The lexicalnatureof N.J. and SeidenberS, Macl}rrralrl,M.C., l'earlrnutter, gicalReview,10 1, 676-7 03. ity resolution."Psycholo syrrl :rtt ic arnbigr.r M.K. ( 1998)."Modelingthe Influenceof Mcllae,K., Spivey-Knowlton,M.). and Thnenhaus, 'l'lrcrrralir: (and trit Other Constraints)in On-Line SentenceComprehension!'lournal tnge,38,283-312. t t o r y nr r tl Lrtt tgt ol Mt t Merlo, P. arttl S. Stevenson.(1998). "What grammars tell us about corpora: the caseof retlut'ctlrt'lativeclauses."Proceedingsof the Workshopon VeryLargeCorpora,6, I34142. Berkeley Hypothesis." and the Unaccusative l)erlrrrrrlto,l) Nl. (1978)."hnpersonalPassives Society, 4, 15-189. l.irrlrristir's Itrlf arrl, ( i. irrrrl I. Sag. ( 1994). Head-Driven Phrase Structure Grarnntar. Stanford: C)l.SI /(,lrit ago:'l'heUnivelsityof ChicagoPress. Reducedrelativesjudged hard require constraint-basedanalyses Pollard, c. and I. sag. (1987). An Information-BasedSyntax and senntrtics. vorume I. Fundamentals.stanford: GLSI/chicago: The university of chicago press. sag' I. ( 1998)-"English Relativeclause c-onstr action{' Journarof Liiguistics. Sag, I. and T. Wasow. (1997). Syntactic Theory: A Formal Introiuction, ms. Stanfbrd University. Spivey-Knowlton, M. f. (1996).Integration of Visualanil LinguisticInforrnatiotl: Hsl1an Data and Model simulations. ph.D. Thesis.university of Rochester,Rochester, Ny. spivey' M.f. and M.K. Thnenhaus.(r99g). "syntactic ambiguity resolution in discourse: Modeling the effects of referential context and lexical fiequency'' rournal of ExperirnentalPsychology:r.earning Memory, and cognition, z+, tsit-tsaz. stevenson,s. (1994a)."competition and Recencyin a Hybrid Nehvork Model of syntactic Disambiguatioil' lournal of psycholinguisticResearch,23, 2gS_322. stevenson,s. (1994b). A CompetitiveAttachmenr Model for Resolvingsyntactic Anfuiguities in Natural Language processing.ph.D. Thesis, Department of {lomputer Science, University of Maryland. stevenson, s. and P. Merlo. (1997). "Lexical structure and parsing Complexity." I' M'C' MacDonald (ed.) Lexical Representations and Sentenceprocessing.Aspecial issue of.Languageand Cognitive processes. Thnenhaus'M.K. and G.N. carrson. (19g9). "Lexical Structure and Languagecomprehension."In w. Marslen-wilson (ed.),LexicalRepresentation and processs,pp. szr-561 . Cambridge:MA: MIT Press. Thnenhaus,M.K. and f.c. Tiueswell.(1995). "sentencecomprehension.,'In J. Miller a'cl P. Eimas (eds.), SpeechLanguageand Comrnunication: Handbook of perception and Cognition.New York: Academic press. van valin, R.D.Jr.(1990)."semantic parametersof split Intransitivity.,' Language,22,22r_ 260. z