Toward an account of accented pronoun interpretation in discourse context

48 Pages·2001·0.58 MB·English

by  Jennifer Venditti

Toward an account of accented pronoun interpretation in
discourse context: Evidence from eye-tracking
JenniferJ.Venditti , MatthewStone , Preetham Nanda ,Paul Tepper
? y y y
UniversityofPennsylvaniaInstituteforResearch in CognitiveScience
?
Rutgers Center forCognitiveScienceand Dept. ofComputerScience
y
December 21,2001
Abstract
Pronouns uttered withintonational prominence (i.e.‘pitch accent’) are interpreted differently from
thoseutteredwithoutsuchprominence. Studieshaveshownthatanaccentedpronounwillshiftattention
fromthemostsaliententityinthediscoursecontexttosomeotherlesssaliententity.Forexample,inthe
spokenutteranceJohnhitBillandthenHEhitGeorge,listenersagreethattheaccentedHEreferstothe
lesssalientBill. Whilethisjudgmenthasbeendiscussednumeroustimesintheliterature,themajority
ofpreviousstudieshavereliedonintrospectiveoroff-linejudgments,andhavefocusedoninterpretation
instrictlyparallelclausalsequences. Thispaperreportstheresultsfromaneye-trackingstudyofon-line
interpretation of nuclear-accented (subject) pronouns in differing discourse contexts. We present data
suggestingthat(i)thetypeofinferreddiscoursecoherencerelation,and(ii)theabilitytolocallyresolve
the presupposition of contrast evoked by the accent influence the interpretation of accented pronouns.
Inaddition,ourdatatellussomethingaboutthetime-courseofincrementalinterpretationofutterances
withaccentedsubject pronouns. Wefindthat bothpotentialantecedents areevoked immediatelyupon
hearingtheaccentedpronoun.Apreferenceforonereferentovertheotheronlyemergesoncesubsequent
propositionalinformationisencounteredwhichlendssupportfortheinferreddiscourserelation.
1
Contents
1 Introduction 3
1.1 Constraintsonaccentedpronouninterpretation . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Theparallelfunctionstrategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Determiningwhichreferentis‘expected’or‘salient’ . . . . . . . . . . . . . . . . 4
1.1.3 Kameyama’saccountofaccentedpronouninterpretation . . . . . . . . . . . . . . 5
1.1.4 Whenaccentedpronounsdonotshiftattention . . . . . . . . . . . . . . . . . . . 6
1.2 Buildingadiscoursecontext . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Experiment1: Determiningpotentialbiases 8
2.1 Motivation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Discourses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.2 Auditorystimuli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.3 Visualstimuli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Experimentdesignandprocedure . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.3 Datacodingandanalysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Experiment2: Trackingeyefixationson-lineindiscourse 12
3.1 Motivation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.1 Discourses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.2 Auditorystimuli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.3 Visualstimuli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1 Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.2 Experimentdesignandprocedure . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.3 Datacodingandanalysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4 Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4.1 Canoureyes‘followalong’withaspokendiscourse?. . . . . . . . . . . . . . . . 17
3.4.2 Fixationonreferentsofnounsvs.pronouns . . . . . . . . . . . . . . . . . . . . . 19
3.4.3 Effectofaccentonfixationbehavior . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4.4 Theinfamous“parallelstructures” . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.5 Whenaccentedpronounsappearnottoswitchreference . . . . . . . . . . . . . . 31
4 Discussionandpreliminaryproposal 33
4.1 Theroleof‘parallelism’ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1.1 Smyth’ssyntacticaccount . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1.2 Kehler’sdiscoursecoherenceaccount . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Accentedpronouninterpretationbasedoncoherencerelations . . . . . . . . . . . . . . . 36
4.3 Thetime-courseofinterpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2
1 Introduction
Anumberofstudiesinboththeoreticalandcomputationallinguisticshaveobservedthatpronounsuttered
withintonationalprominence(i.e.‘pitchaccent’)areinterpreteddifferentlyfromthoseutteredwithoutsuch
prominence. Forexample, considerthe now famousexampleaboutJohn andBill’s aggressions, givenin
(1).
(1) a. JohnhitBillandthenhehitGeorge.(he=John)
b.JohnhitBillandthenHEhitGeorge.(HE=Bill)12
In these examples, the gender information conveyed by the pronoun is consistent with either of the
two malereferentsinthepreviousclause. However,despitethispotentialambiguity,nativeintuitionsare
unambiguous: the pronoun uttered without an accent (1a) is interpreted as referring to John, while the
accentedpronoun(1b)istakentorefertoBill.
Thecurrentstudytakesanexperimentalapproachtodocumentingthisintuition.Specifically,wepresent
datafromapsycholinguisticexperimentusingeye-trackingwhichexaminestheon-lineandoff-lineinter-
pretationofaccentedandunaccentedpronounsindiscoursecontext.Themainresearchquestionsaddressed
bythisstudyare: Istheinformationprovidedbytheaccentinterpretedon-line,aslistenersparseaspoken
utterance? Doesthediscoursecontextaffectinterpretationpreferences? Iseye-trackingavalidmethodol-
ogyforinvestigatingthesequestions?
Ourdatasuggestthataccentaloneis notsufficientto switch referenceto a lesssaliententity. Rather,
we find that (i) the type of inferred discourse coherence relation, and (ii) the ability to locally resolve
the presuppositionof contrastevokedby the accentinfluence the interpretationof accentedpronouns. In
addition,ourdatatellussomethingaboutthetime-courseofincrementalinterpretationofutteranceswith
accentedsubjectpronouns.Wefindthatbothpotentialantecedentsareevokedimmediatelyuponhearingthe
accentedpronoun.Apreferenceforonereferentovertheotheronlyemergesoncesubsequentpropositional
informationisencounteredwhichlendssupportfortheinferreddiscourserelation.
1.1 Constraints onaccented pronoun interpretation
Anaphoricexpressionssuchaspronounsmustbeinterpretedwithreferencetothepreviousdiscoursecon-
text. Theantecedentofa anaphorisgenerallysomesaliententitymentionedrecentlyinthediscourse. A
greatnumberoflinguistictheorieshaveproposedmechanismsbywhichsuchreferenceresolutioncanef-
fortlesslyoccurinlanguagecomprehension.However,amajorityofthesestudieshaveinvestigatedoff-line
judgments, and have focused on pronoun interpretation in written language, or on unaccented pronouns
in spoken language. In this study, we focus our attention on the on-line interpretation of intonationally
prominent,or‘accented’,pronounsinspokendiscourse.
Earlyresearchonaccentedpronounsreliedonintrospectivejudgmentsaboutpronouninterpretationin
sequencesofconjoinedclauses, suchasthoseshownin(1a)and(1b)above(e.g.[Gleit61,AJ70,Lak71],
etc.). ThegeneralconclusionfromthesestudiesissummarizedbyAkmajianandJackendoff’sobservation
that“contrastivestressoneitherapronounornounwillprohibitcoreference”[AJ70,p.124].Inotherwords,
givensomeheuristicwhichdetermineswhichentityinadiscoursemodelislikelytobetheantecedentof
(i.e. ‘coreferentwith’) a pronominal,puttingan accentonthat pronominalwill cuethat theantecedentis
infactsome‘other’entity.Thisobservationasstatedreflectsourintuitionthatanintonationallyprominent
pronounreferstoanentitywhichisnottheonethatwewould‘expect’,thoughitdoesnotspecifydetails
ofhowtogoaboutdeterminingwhichreferentinthediscoursethis‘otherguy’is.
1.1.1 Theparallelfunctionstrategy
Followinguponthisearlyhypothesis,Solanandotherssuggestedthatunaccentedpronouninterpretationis
drivenbyaheuristiccalledtheparallelfunctionstrategy,andthat“contrastivelystressingthepronounina
1Followingtheusualconventionintheoreticallinguistics,capitalizationwillbeusedinexampleshereasashorthandforindicating
thatthewordbearsaprominentaccent.Whattypeofaccentthisrepresentsisanopenquestionwhichwearecurrentlyinvestigating.
2ThisexampleofaccentonasubjectpronounistakenfromOehrle’s [Oeh81]extensionofAkmajianandJackendoff’s [AJ70]
originalexamplesinwhichtheaccentoccursonanobjectpronoun.
3
sentencehastheeffectofunderminingtheparallelfunctionstrategy... [thatis,ithas]theeffectofshifting
preferredantecedents”[Sol83, p.163]. Theparallelfunctionstrategyisageneralheuristicfirstproposed
bySheldon[Shel74],bywhichlisteners/readersinterpretanunaccentedpronountobecoreferentwiththe
entity which was mentioned in the same grammaticalposition (e.g. subject, object, etc.) in the previous
clause. Given this strategy, Solan’s claim is that placing intonationalprominenceon the pronounresults
inaninterpretationinwhichthepronounnowreferstoanentitywhichisinanothergrammaticalposition
(see also[Smyth92,Smyth94]). As Solannotes, “stressinga pronouninformsthe hearerthatthe speaker
intendsitsantecedenttobesomethingunexpected”[Sol84,p.176]. Here,whatis‘expected’isdefinedby
theparallelfunctionstrategy.
1.1.2 Determiningwhichreferentis‘expected’or‘salient’
Accordingto the parallelfunctionstrategy, an ‘expected’referentis thatwhich is in the same grammati-
cal position as the pronounin question. Recent studies in theoretical and computationallinguistics have
proposedanalternativemeanstodefinewhichreferentsare‘expected’inadiscourse,ormoreimportantly
here, whichreferentsare ‘unexpected’antecedentsof pronominalforms. Attention-drivenstudiesof dis-
course coherence and pronoun resolution have proposed that entities in the discourse model are ranked
according to their attentional ‘salience’, and that this ranking determines the ‘expectedness’ for corefer-
encewithasubsequentpronominal(see,forexample,theproposalsassociatedwithCenteringTheory,e.g.
[GJW95,WJP98]). Inthisapproach,thesetofsaliententitieswhicharepotentialreferentsofapronominal
forminclause isdefinedasthoseentitieswhicharerealizedintheimmediatelyprecedingclause .3
We will call thUisiset of referents the ‘salient subset’ (aka. the ‘forward-looking center list’ in CenUtei(cid:0)ri1ng
Theory).Thesaliencerankingoftheentitiesinthissalientsubsetisoftendeterminedbytheirgrammatical
positionintheclause,withsubjectsbeingmoresalientthanobjects,etc.[Chafe76,WJP98]. Giventhistype
ofranking(andanunderlyingpreferencefordiscoursestobecoherent),a subjectin willbea more
‘expected’antecedentofapronominalin thanwouldbeanobject.4 NotethatintheUcai(cid:0)se1ofpronominals
insubjectpositionin , asin(1)above,Uthiepredictionsoftheparallelfunctionstrategyandthesalience
ranking/coherencehypUoithesisareidentical:thepreferredantecedentofanunaccentedsubjectpronounwill
bethegrammaticalsubjectrealizedintheprecedingclause.5
Thissimplesalience ranking(accompaniedby thepreferencefordiscoursecoherencevia centercon-
tinuation)can explainthe ‘default’interpretationof a greatmajority of written pronounsand unaccented
pronounsinspokenlanguage. However,thesedefaultpreferencesdonotholdwhenthepronounreceives
intonational prominence, as described above. To account for these cases, there have been a number of
proposals within the attention-driven framework. Among the proposals, Cahn states that “when a pitch
accentisappliedtoapronominal,itsmaineffectisattentional,ontheorderofitemsin[thesalientsubset]”.
[Cahn95, p. 291] (see also [Cahn]). Nakatani suggests that accented pronounsmark a “shift in attention
awayfromthecurrentdiscoursecentertoanewdiscourseentitythatwasindeedsalientintheimmediate
discourse context”[Naka93, p. 166](see also [Naka97a, Naka97b]). Terken remarksthat “accentedpro-
nouns... signalthattheintendedentityisnotthemostaccessibleentityatthatmoment. Thatis,weexpect
theoccurrenceofanaccentedpronouninsituationswherethepronounviolatestheprominenceranking.”
[Terk93]. Andfinally, Kameyamaproposesthat“a focusedpronountakes thecomplementarypreference
of the unstressed counterpart” [Kame99, p. 315]. What all of these proposals have in common is that a
candidateset ofcurrentlysaliententitiesis defined, andisrankedaccordingto relativesalience oracces-
3Weusetheshorthand‘clause’heretodenotetheunitofstructureoverwhichthesetofsaliententitiesisdefinedandupdated.
However, therehasbeensomedebate about whattheexact nature ofthis unitshouldbe: asentence? atensedclause? etc.(see
[Kame98,Milt]andalsoSection3.4.5below).
4Thisparentheticalaboutcoherenceisincludedherebecauseinthemostprominentattention-drivenanalysisofdiscourseinterpre-
tation,CenteringTheory,pronouninterpretationisbasedontwointeractingconstraints: (i)therankingofthesalientsubset,and(ii)
thepreferenceforadiscoursetobecoherent. Thatis,readers/listenersprefertocontinuecenteringthesamediscourseentityacross
utterancespairs.Therefore,resolutiondependsonmorethanjustthesalienceranking.
5Thisstudyinvestigates theinterpretation ofpronominals insubject position only. Inthecaseofobject pronominals, suchas
JohnhitBillandthenGeorgehithim,thepredictionsoftheparallelfunctionstrategyandthesalienceranking/coherencehypothesis
differ: the‘expected’ antecedent oftheobjectpronounistheprevious objectaccordingtotheparallel functionstrategy, butisthe
previoussubjectaccordingtotheattention-driven hypothesis. Wearecurrentlyconductingexperimentswhichexaminetheon-line
interpretationofaccentedandunaccentedobjectpronominals.
4
sibility (in practice, grammaticalrole). The preferredantecedentof an accented pronounis some salient
entity containedwithin thisset, butcruciallyis notthe mostsalient entity(accordingto the defaultrank-
ing). ThisisessentiallythesameclaimmadebyAkmajianandJackendoff[AJ70], Solan[Sol83,Sol84],
Smyth[Smyth92, Smyth94]andothers, thoughcouchedin adifferentframework. Intheattention-driven
approaches,thedefinitionofthesalientsubsetandtherankingofentitieswithinitismadeexplicit.
1.1.3 Kameyama’saccountofaccentedpronouninterpretation
To clarify the predictions made by the attention-based theories, we will briefly outline Kameyama’s
[Kame99]accountofaccentedpronouninterpretation.Thisisperhapsthemostexplicitandwell-documented
ofthecurrentproposals(andisgenerallyrepresentativeoftheassumptionsaboutaccentedpronouninter-
pretationmadeby[Cahn,Cahn95,Terk93,Naka93,Naka97a,Naka97b],thoughnotnecessarilyintheexact
details).
Kameyamaclaimsthataccentedpronouninterpretationresultsfromaninteractionofthesemanticfocus
associated with the pitch accent, and thedefinitionand rankingof the salientsubsetused forinterpreting
unaccentedpronouns.Shestatesthat“astressedHEpresupposesaconstraint Fthatthereisacontextually
determinedsetofentities( = x x F E whereEisthedomai(cid:24)nofindividuals)withatleast
f
twomembers—thedenota[t[iHonEo]f]HEf( j 2 )a(cid:18)ndagtleastonemorecontrastingindividual”[Kame99,p.
o
308].Thisconstraintisduetotheseman[[tHicEfo]]cusinterpretationofthe(narrowfocus)pitchaccentitself,as
describedbyRooth[Rooth92]. Kameyamaimplicitlyassumesthatthetypeofintonationalprominenceon
a‘stressedpronoun’isthesameasanarrowfocus,or‘contrastive’,pitchaccent.6 Thishypothesisiscon-
sistentwiththecharacterizationofaccentedpronounsinthepreviousliteratureas‘contrastivelystressed’
(e.g.[AJ70,Sol83,Smyth94]).7
The“contextuallydeterminedsetofentities”whichKameyamareferstohereisfunctionallydefinedas
thesetofsaliententitiesinthelocalattentionalstate(i.e.theentitiesrealizedintheimmediatelypreceding
clause). Thissalientsubsetisthesameforbothaccentedandunaccentedpronouns—theonlydifference
being the relative salience ranking within the set. She proposes that accented pronoun interpretation is
driven by a preference ranking which is ‘complementary’to the ranking for unaccented pronouns. This
meansthatthesaliencerankingofpossibleantecedentsforaccentedpronounsinequivalenttothereverse
oftherankingusedforinterpretationofunaccentedpronouns. Kameyamasuggeststhataccentedpronoun
interpretationproceedsviathesequenceofcomputationsdescribedbelow(paraphrasedhere,see[Kame99,
p.315]forfulldetails). Consideragaintheexamplein(2).
(2) JohnhitBill. ThenHE...
Determinethesalientsubsetbasedonthelocalattentionalstateof :
(cid:15) John,Bill Ui(cid:0)1
Defterminethgesaliencerankingfortheunaccentedpronoun(‘default’)case:
(cid:15) John Bill
Cofmpute>compglementarypreference(i.e.re-rank):
(cid:15) Bill John
Difschar>gethepgresupposedconstraintofcontrast Cfortheutterance .
(cid:15) (cid:24) Ui
By this account, interpretation involves re-ranking of entities in the salient subset, and choosing the
mostsaliententityofthatnewlyrankedsetasthepreferredantecedentoftheaccentedpronoun.8 Notethat
if the initialsalience rankingis basedonlyupongrammaticalrole, suchan accountpredictstheincorrect
interpretationofanaccented(orunaccented)pronouninobjectpositioninparallelstructuressuchasJohn
6Thefunctionanddistributionofnarrowfocuspitchaccentshasbeendiscussedextensivelyintheintonationliterature:seeRooth
[Rooth92],Bolinger(e.g.[Bol61]),Ladd(e.g.[Ladd80]),amongmanymanyothers.
7Notehoweverthatinotherrecentstudies,thereissomedebateaboutwhattheexactnatureofthisintonational prominenceis.
CahnsuggeststhatthepitchaccentmustbeL+H*[Cahn,Cahn95],whileNakataniclaimsthattheshiftinpreferredinterpretation
occurswithH*accentsaswell(e.g.[Naka97b]).Wearecurrentlyconductingexperimentswhichexaminethisquestionindetail.
8Notethatthisproposalpredictsthatthelowest-ranked entities inthedefaultorderofsaliencewillbecomethehighest-ranked
entitiesinthere-rankedset.ThissuggeststhatthepreferredantecedentoftheaccentedpronouninasequencelikeJohnhitBillusing
thebatownedbySam.ThenHE...willbeSam.Thispredictionhasyettobeempiricallytested.
5
hit Bill then George hit him/HIM. That is, the previoussubject should be the preferredantecedentin the
unaccentedcase,whilethepreviousobjectshouldbepreferredintheaccentedcase. Bothofthesepredic-
tionsresultinanincorrectinterpretation. Pronouninterpretationinstrictlyparallelsequencesisageneral
problemencounteredbyapproacheswhichconsideronlygrammaticalroleindeterminingsalienceranking
(e.g.mostoftheimplementationsofCenteringTheory). Toaccountforthis, Kameyamahasproposedan
additional property sharing constraint, similar to the parallel function strategy, which she claims comes
intoplayinthe(default)saliencerankingstep(see[Kame86,Kame99]).9 Wewillreturntodiscussionsof
pronouninterpretationinstrictlyparallelvs.non-parallelsequencesinSections3.4.3,3.4.4and4below.
Itis importantto askatthispointwhetherKameyama’sproposedcomputationsforaccentedpronoun
interpretation can be used in a psycholinguistic modelof on-line interpretation. Kameyama is careful to
note that that “no sequential order is assumed” among the computations [Kame99, p. 308], but we can
assumethatsomestepsdoprecedeothers—forexample,thesaliencerankingprobablydoesprecedethe
dischargingofthepresupposition Cfor . Howmightthisprocessworkon-line? Onecanimaginethat
uponparsingthe clause , lis(cid:24)tenerswUililhavein memorythe saliententitieswhichwere justencoun-
tered, and they may evenUbne(cid:0)1able to form a hypothesisaboutthe salience rankingof these entities at this
point,basedon(atleast)knowledgeaboutgrammaticalroles. Then,uponhearingthediscourseconnective
thenandthepronounHEin thefollowingclause , listenersmaybecuedtoinitiate thecomplementary
preferencecomputation.ThismayoccurrapidlyasUthieaccentisperceived.Thefinalstepininterpretationis
todischargethepresupposedconstraintofcontrastfortheutterance . IfwefollowKameyama’sdescrip-
tion,thisstepnecessarilycannotbecomputedimmediatelyaftertheUacicentedpronounisperceived.Thisis
becausethepresupposedconstraintwhichisdischargedindicatesthatthereis“a contextuallydetermined
setofpropositions[ouremphasis]obtainedbyinstantiatingasetabstractionwiththealternativevaluesof
thefocusedelement”[Kame99,p.308].Whilethe‘alternativevalues’maybeimmediatelyavailable(after
parsingthepreviousclauseandtheaccentedHE),thepropositioncarriedbythetargetclause maynot
be known in full until the whole clause is parsed. That is, since the contrast set is obtained bUyi instanti-
ating hit George(x) x F with the alternativevaluesof (see [Kame99, p. 308]for full
f
detailsf)10, one must fijrst2knogw what the proposition carried[b[Hy E]]is.2TFherefore, this step, as worded by
Kameyama,musttakeplaceafter hasbeenparsed.Ofcourse,Uwihetherornotalistenercanbuildon-line
an incrementalhypothesisof theUpriopositionalcontentof an utterance is one of the ultimate questionsin
sentence processing(see e.g. [TT95] fora review). If listenerscan glean information(albeitincomplete)
aboutpropositionalcontenton-line, then it is highly likely that they may also be able to proceedon-line
withthedischargingofthepresupposedconstraintofcontrastwhichKameyamadescribes. Insuchacase,
the contrast set would be computed based on incomplete information and may be ultimately wrong and
subjecttosubsequentrevision. We willreturntothisissueoftheincrementalnatureofaccentedpronoun
interpretationinthediscussioninSection4.
1.1.4 Whenaccentedpronounsdonotshiftattention
Althoughamajorityofpreviousaccountsdescribeaccentedpronounsasshiftingthecenterofattentiontoa
lesssaliententityinthediscoursecontext,therearemanycasesinwhichtheattentionisnotshifted.Instead,
theaccentservestocueacontrastbetweenthesaliententityandsomeotherunspecifiedsetofentities(e.g.
seediscussionsin[Prev95,Prev96]).Considertheexamplegivenin(3).
(3) Jackisaphysicist.HE...
[Kame99,p.317]
Here,nativespeakersunambiguouslyinterprettheaccentedpronounasreferringtoJackandnotsome
‘other guy’, even when Jack is the most salient entity in the current discourse context. These sorts of
examplesareclearcounterexamplestothemanyclaimsthataccentedpronounsshiftattentionawayfrom
themostsaliententity.Howthencanweresolvethisapparentcontradiction?
9ButseethediscussioninSection4belowofKehler’s[Kehl01]unifiedaccountofdiscoursecoherenceandpronouninterpretation
inbothparallelandnon-parallelsequences.
10Forthisexample,thiscomputationwouldresultinthesetofpropositions JohnhitGeorge,BillhitGeorge .
f g
6
Kameyama suggests that these cases can also be accounted for by the same mechanisms used in in-
terpreting accented pronounswhich shift attention. The key is the size of the salient subset of discourse
entities. Kameyamaproposesthat these cases (of no shift) occur when there is onlyone salient entity in
thelocalattentionalstate. Giventhis,theirinterpretationfallsoutfromtheproposedstepsofinterpretation
outlinedabove. Thatis,ifthesalientsubsetcontainsonlyoneentity,thenre-rankingthesetresultsinthat
singleentityremainingthemostsalient,andthusthepreferredantecedentoftheaccentedpronoun. Inthis
way,Kameyamapredictsthatthesinglesaliententitywillbethepreferredantecedentofbothanunaccented
pronounaswellasanaccentedpronoun. Shenotesthatintheaccentedpronouninterpretation,thereisan
additionalpresuppositionofcontrast:incaseswherethesalientsubsetisasingleton,“atleastonecontrast-
ingindividualisaccommodated”whenthisconstraintisdischarged[Kame99,p.315]. However,theexact
natureofthisaccommodationremainsunspecified.
Kameyama’saccountunifies the apparentdiscrepancyin interpretationof the two contrasting classes
ofaccentedpronouns:thoseinwhichtheattentionisshiftedtoaless-saliententity,andthoseinwhichthe
attentionis notshifted. In thispaper, we will restrictourfocusto cases in which the attention is shifted.
Thatis,wewillexaminecasesinwhichthereismorethanonesaliententityintheimmediatecontext.
1.2 Building adiscourse context
Most of the early descriptions of accented pronouns describe intuitions about coreference only in very
parallelclause sequenceslikeJohnhit Billand then HE hit George, presentedin isolation (e.g.[Gleit61,
AJ70,Lak71,Oeh81,Sol83,Sol84],andalsomorerecentlyby[Smyth94,BST98]). Inmorerecentwork
onpronounresolutionwithinthecontextofcomputationallinguisticsandartificialintelligence,researchers
have generalized the use and interpretation of accented pronounsbeyond strictly parallel structures (e.g.
[Cahn,Naka93,Terk93, Cahn95,Prev95,Prev96, Naka97a,Naka97b, Kame99]). InSection1.1.3above,
weoutlinedthedetailsofonesuchgeneralaccount,inwhichpronouninterpretationisdeterminedbythe
relative salience ranking of entities in the current discourse context (in conjunction with the preference
for coherence across utterance pairs). In the sections that follow, we test this hypothesis empirically by
examiningtheon-lineandoff-lineinterpretationofunaccentedandaccentedpronounsinstructureswhich
arenotstrictlysyntacticallyparallel.
Inconstructingadiscoursecontextforourexperimentalstimuli,wewerefacedearly-onwiththecrucial
question:Inwhatdiscoursecontextsistheuseofanaccentedpronounfelicitous?Basedonthediscussions
providedby[Rooth92,Prev95,Prev96,Kame99]andothers,ourworkingassumptionisthataccentedpro-
nouns are felicitous in contexts in which there is a basis for contrast among members of a set of salient
entities.11 Inordertocreatesuch acontext,we constructednarrativesin whichthe discourseparticipants
are collaboratingon a jointaction. The discourse providesinformationaboutwhat each participantcon-
tributestothejointgoal.Theprogressionofthediscourseisdrivenbytheopenquestionsunderdiscussion,
alsoknownasQUDs(seee.g.[Rob96]),andnatureoftheQUDsiswhatsetsupthebasisforcontrast. Ex-
ample(4)showstheQUDstructureofoneofthediscoursestimuliusedinthisexperiment. Thisstructure
isrepresentativeofalltheteststimuli(seeSet1intheAppendixforalldiscourses).
(4) QUD:Whathappened?
T)hezebraandthepigwantedtowashthecartogether.
QUDs: Whatdidthezebracontribute?Whatdidthepigcontribute?
T)hezebraputabucketofsoapywaternexttothepignearthefrontofthecar.
(answersWhatdidthezebracontribute?)
QUDs: Whatelsedidthezebracontribute?Whatdidthepigcontribute?
T)henhe/HEgotoutsomesponges.
(‘he’answersWhatelsedidthezebracontribute?)
(‘HE’answersWhatdidthepigcontribute?)
...
11Wearecurrentlyinvestigatingthisquestionfurtherthroughanalysesoflarge-scalecorpora.
7
Atthebeginningofthediscourse,theopenquestionisthegeneralWhathappened?. Theintroduction
of the two discourse participants (the zebra and the pig) and the joint goal (washing the car together)
in sentence 1 motivates the subsequent QUDs: What did the zebra contribute? and What did the pig
contribute?. Sentence2proceedstoanswerWhatdidthezebracontribute?,therebyremovingitfromthe
QUDlist. Atthispointinthediscourse,weareleftwiththeopenQUDWhatdidthepigcontribute?,but
havealsoaddedanotherpossibility: Whatelsedidthezebracontribute?. Sentence3couldanswereither
ofthesequestions. Ifthesubjectpronouninsentence3isunaccented,itistakentoanswerWhatelsedid
thezebracontribute?,whileifitisaccented,itanswersWhatdidthepigcontribute?. Thatis,thesalience
ranking/coherence constraints drive the interpretation of the unaccented he — the listener assumes that
theQUDaboutzebra’scontributionisstillbeinganswered. Incontrast,theaccentedHEcuesthelistener
thatnextQUD(aboutthepig’scontribution)istobeaddressed.Theexperimentalstimuliwereconstructed
basedonthistypeofQUDstructure.Thediscoursein(5)showstheexamplefrom(4),withQUDsremoved.
AllstimuliarelistedintheAppendix.12
(5) 1. Thezebraandthepigwantedtowashthecartogether.
2. Thezebraputabucketofsoapywaternexttothepignearthefrontofthecar.
3a.Thenhegotoutsomesponges.
3b.ThenHEgotoutsomesponges.
4. Andtogethertheystartedwashingthehoodandthefenders.
Inthecurrentstudy,weexaminelistenerpreferencesaboutwhichQUDwillbeansweredbythetarget
sentence3. Wealsodocumenttheincrementalon-lineinterpretationofboththeaccentedandunaccented
pronouns,aswellasoff-linejudgments.Wenowturntoadetaileddiscussionofthoseexperiments.
2 Experiment 1: Determining potential biases
2.1 Motivation
ThepurposeofExperiment1wastwo-fold.First,wewantedtoexperimentallydeterminewhetherthereare
anybiasesinourvisualandaudiostimuliwhichwouldcauselistenerstopreferonecharacterovertheother
intheirinterpretationofthepronounintargetsentence3. The(visualandauditory)stimuliweredesigned
with the intention that either character would be equally plausible as do-er of the target action, and this
normingexperimentassessedoursuccessindoingso.13
AnothermainmotivationforExperiment1wasto investigatewhethertheintentionalstructureweset
upforthediscourses,definedherebytheopenquestionsunderdiscussion(QUDs),affectslisteners’prefer-
encesforwhowillbethedo-erofthetargetaction. Thatis,inthediscoursegivenin(4)above,iflisteners
entertaintheQUDWhatelsedidthezebracontribute? afterhearingsentence2,theywillprefersentence
3 to describethe action of thezebra. If, on the other hand, listenersentertainthe QUDWhatdid the pig
contribute?,preferenceswillbeforthepigtobedoingtheaction. IfeitherQUDisequallyplausible,pref-
erencesshouldbemixed.Inthisexperiment,weexaminedlistenerpreferencesforagentofthetargetaction
describedbysentence3. Weincludeddiscourseslikethatin(4),andalsoincludedonesinwhichtheQUD
structurewasslightlydifferent.Section2.2.1describesthediscoursesindetail.
2.2 Materials
2.2.1 Discourses
All discourses describe a joint collaborative action between two cartoon animals. Two variants of a 3-
sentencediscoursewereconstructed,asshownin(6).
12TheteststimuliarelistedinSet1,andthefillersarelistedintheothersets.
13ThankstoMarkSteedmanforsuggestingthisexperiment.
8
(6) 1. Thezebraandthepigwantedtowashthecartogether.
2a.Thezebraputabucketofsoapywaternexttothepignearthefrontofthecar.
2b.Thezebratoldthepigtoputabucketofsoapywaternearthefrontofthecar.
3. Thensomeonegotoutsomesponges.
N1(thezebra)andN2(thepig)areintroducedasaconjoinedNPinthefirstsentence. N1remainsthe
subject of the second sentence, but the do-er of the action is varied: in (6.2a) N1 does the action, while
in (6.2b) N2 does the action. In the third sentence, the identity of the actor is unspecified. We chose
to use someone instead of the pronominalhe to refer to the agentin order to avoid preferencesbased on
discoursecoherencestrategieswhichwouldresultfromusingapronoun. Sincethe openquestionsunder
discussionafterhearingsentence2arenecessarilydifferentdependingonthevariant,wesuspectthatthis
mayinfluencelistenerjudgments.ThedifferentQUDsareshownin(7).14
(7) 2a.Thezebraputabucketofsoapywaternexttothepignearthefrontofthecar.
QUDs: Whatelsedidthezebracontribute?Whatdidthepigcontribute?
)
2b.Thezebratoldthepigtoputabucketofsoapywaternearthefrontofthecar.
QUDs: Whatelsedidthepigcontribute?Whatdidthezebracontribute?
)
2.2.2 Auditorystimuli
AllutterancesintheexperimentsreportedherewererecordedbythefirstauthorusingaShureSM10Auni-
directionalhead-mountedmicrophoneandaTASCAMTEACPA-1portableDATrecorder.Theutterances
were recorded at 48KHz then transferred to UNIX/Linux workstations and downsampled to 16KHz for
analysis and playback. Acoustic analysis was performed using Entropic Research Labs ESPS/Waves+
software,version5.3.1.
Care was taken to utter the discoursesin a uniformyet naturalmannerin order to minimize acoustic
and prosodic variability. Because of the range of text material used in the discourses (see Appendix),
the prosodic structures necessarily were not identical. However, certain relevant prosodic features were
intentionallykeptconstant:thenounphrasesreferringtoN1andN2receivedpitchaccentsinsentence1as
wellasin sentence2. Sentence3wasutteredwitha ‘hat-pattern’intonation,witha pre-nuclearH*pitch
accentonsomeone.15
2.2.3 Visualstimuli
All visual stimuli used in the experiments reported here were generated in Abode Photoshop 6.0 using
animalcharactershand-drawnbyPaulTepperandclipartavailableontheinternet.
Eachstimulusshowsascenecontaininganimalsandobjectswhichareinvolvedintheactionsdescribed
bythediscourse. Figure1givesanexampleofthevisualstimuluspairedwiththediscoursein(6)above.
In each scene, the two characters involved in the joint activity are located diagonal to one another and
equidistantfromtheobjectmentionedattheendofsentence2(here,thecar),whichisimmediatelyprevious
tothementionofsomeoneinthetargetsentence3. Therelativepositioning(i.e.left-right,top-bottom)of
these characters was balanced across items. In addition, the object described by the action in the target
sentence (here, the sponges) is located beside each of the characters. Such placement of the characters
and target object was intended to preventany bias due to one character being closer to the target object,
orbiasduetoonecharacterconsistentlybeinginaparticularregionofthescene. However,theplacement
of the object of the action in sentence 2 (here, the bucket) may in fact produce a bias toward N2 in the
discourses. In sentence 2, N1 places(or tells N2 to place) the objectin the vicinity of N2. The auditory
14SeeSet1intheAppendixforalistofalldiscoursesandvisualstimuliusedinExperiment1. Thevisualstimuliwereexactly
asshown. Thediscoursestimulihadthefollowingstructure: sentence1introducedN1andN2asaconjoinedNPsubject,followed
bytheintroductionofthecollaborativeactioninthepredicate. Sentence2occurredinboth(a)‘doing’and(b)‘telling’variants. In
sentence3,someonewasusedinsteadofhe.Sentence4wasomittedinExperiment1.Thisstructureisthesameastheexamplegiven
in(6).
15See[BE94]forafulldescriptionoftheToBI-styleintonationnotationusedinthispaper.
9
Figure1:Exampleofthevisualstimuluspairedwiththediscoursein(6).
stimulus describes this action, and the visual stimulus shows the object in its resulting location near N2.
It is possible that this positioning may cause listeners to prefer N2 to be the do-er of the targetaction in
sentence3,sinceN2isclosesttothelocationthatthelastactionoccurred. Thatis, whileweintendedfor
the auditoryand visualstimuli to be unbiasedtoward either of the two characters, this positioningof the
objectinsentence2mayproduceaslightbiastowardN2inthetargetsentence. However,thisprediction
wouldholdforbothtype(a)‘doing’and(b)‘telling’variants.16
2.3 Methods
2.3.1 Subjects
FortystudentsfromRutgersUniversityparticipatedinexchangeforcoursecredit. Allsubjectswereeither
English mono-lingualsor English-dominantbi-lingualspeakers. All reported normal or corrected vision
andnormalhearing.
2.3.2 Experimentdesignandprocedure
Sixteen discourse-scene pairs were used in Experiment 1 (see Appendix Set 1 for discourse and scene
content,andexample(6)abovefordiscoursestructure). Inaddition,16fillerpairswerealsoincluded(see
Sets2&3intheAppendixforcontentand(6)fordiscoursestructure),resultingin32stimuliintotal. Test
and filler stimuli were presented in a fixed random order. The content of sentence 2 (N1 action (2a) vs.
N2action(2b))wascounterbalancedacrosstwolists,andbothlistswerepresentedinbothascendingand
descendingorder.
Visualscenesandauditorydiscoursestimuliwerepresentedusingthepsycholinguisticexperimentation
softwareDMDX,runningonaPCdesktop.17 Themannerofpresentationwasasfollows: first,thescene
16Furtherdetails oftheplacement ofthecharacters andobjects aremorerelevant totheon-lineeye-tracking study, andwillbe
describedbelowinSection3.2.3.
17DMDXistheWindowsversionofDMASTR,authoredbyKennethForsterandJonathanForsteratUniversityofArizonaPsy-
chology.
10