information structure, discourse -

13th EUROPEAN SUM M ER SCHOOL I N L OGI C, L ANGUAGE AND I NFORM ATI ON

(ESSLLI2001)

INFORMATION STRUCTURE,DISCOURSE STRUCTURE AND

DISCOURSE SEMANTICS

WORKSHOP PROCEEDINGS

EDITED BY

IVANA KRUIJFF-KORBAYOVA AND MARK STEEDMAN

AUGUST 20th–24th 2001THE UNIVERSITY OF HELSINKI , HELSINKI , FINLAND

Workshop ProgrammeCommittee

IvanaKruijf f-Korbayova (co-chair;TheUniversityof theSaarland)Mark Steedman(co-chair;TheUniversityof Edinburgh)

JohnBateman(Universitat Bremen)DanielBuring (Universityof California,LosAngeles)ElisabetEngdahl(Goteborg University)BartGeurts(Universityof NijmegenandHumboldtUniversityBerlin)EvaHajicova (CharlesUniversity, Prague)David Traum(Universityof SouthernCalifornia)Maria Vilkuna(ResearchInstitutefor theLanguagesof Finland)BonnieWebber(TheUniversityof Edinburgh)

i

Workshop Programme

Session1: Monday, August20th

17.00–17.35 OpeningMark Steedmanand IvanaKruijf f-Korbayova: Two Dimensionsof InformationStructure in Relationto DiscourseStructure andDiscourseSemantics (p. 1)

17.35–18.10 Rick Nouwen:Complementsetreference (p. 115)

18.10–18:45 LaurentPrevot: TopicStructure in RouteExplanationDialogues (p. 145)

Session2: Tuesday, August21th

17.00–18.10 Invited talk by Enric Vallduvı: Fragmentsin InformationPackaging (p. 15)

18.10–18:45 Svetlana McCoy: Connecting Information and Discourse Structure Levelsthrough“K ontrast”: EvidencefromColloquial RussianParticles-TO, ZHE,andVED’ (p. 85)

Session3: Wednesday, August22th

17.00–17:35 CarlaUmbach:RelatingContrastandContrastiveTopic: a focus-semanticanal-ysisof “b ut” (p. 175)

17.35–18:10 NoboKomagata:EntangledInformationStructure: Analysisof Complex SentenceStructures (p. 53)

18.10–18:45 IvanaKruijf f-Korbayova andBonnieWebber:InformationStructure andtheSe-manticsof “otherwise” (p. 67)

Session4: Thursday, August23th

17.00–18.10 Invited talk by Livia Polanyi: TheRelationshipof DiscourseStructure to Infor-mationStructure (p. 11)

18.10–18:45 MassimoPoesioand Barbaradi Eugenio: Discourse Structure and AnaphoricAccessibility (p. 129)

Session5: Friday, August24th

17.00–17:35 KatherineForbes,Eleni Miltsakaki,RashmiPrasadAnoopSarkar, Aravind JoshiandBonnieWebber:DiscourseParsingwith a LexicalizedTree-AdjoiningGram-mar (D-LTAG) (p. 17)

17.35–18:10 JenniferSpenader:Presuppositionor Anaphora: Constraints on Choiceof Fac-tiveComplementsin SpokenDiscourse (p. 161)

18:10–18:45 ClosingDiscussion

iii

Preface

Thecall for papersfor this workshopsetasits aim to provide a forum for recentresearchon IS-sensitive discourseprocessingand to facilitatecontact,exchangeof ideasandcooperationbetweenresearchersin computationalandtheoreticallin-guisticsandlogicians,whoareinvestigatingtheseissues.Themissionof thework-shopwasdescribedasfollows:

Informationstructure (IS) is construedbroadlyhereascomprisingstructuralandsemanticpropertiesof utterancesrelatingto thediscoursestatusof their con-tent, the actualandattributedattentionalstatesof the discourseparticipants,andthe participants’prior andchangingattitudes(knowledge,beliefs,intentions,ex-pectations,etc.). This broadview of IS is meantto subsumenotionslike focus,presupposition,givenvs. new, themevs. rhemeandthevariousdichotomiessuchastopic vs. commentor focus,groundor backgroundvs. focusetc.

While discourse structure (DS) is more difficult to define, there is at leastagreementthat coherentdiscourse(multi-sentencedialogueor monologictext) ismorethanasequenceof propositions,justassentencesaremorethansequencesofwords. In discourse,bothexplicit andimplicit devicessignify links betweensen-tences,betweengroupsof sentences,andbetweenelementswithin sentences,andin turn, carryadditionalelementsof discoursesemantics.We arethustaking DSbroadly, to cover all aspectsof theinternalorganisationalstructureof a discourse.DS thussubsumesnotionssuchassegmentation,relationsbetweensegments(in-formationalandintentional),anaphoricrelations,modalsubordination,discoursetopic, thematicprogression,etc.

UnderstandingIS in light of DS andvice versais justified on morethanjusttheoreticalgrounds:experiencewith applicationssuchastranslatingtelephony andinteractive query-answeringmakesit painfully clearthat a theoryrelatingIS andDS is essentialfor accurateNaturalLanguageProcessing.Fortunately, formal ac-countsaddressingtheseissueshavestartedto emergeandsome,to beembodiedincomputationalmodelsof discourseprocessing.Furtherdevelopmentandadapta-tion into practicalsystemswill requireexpertisefrom linguistics,logic andcompu-tation.ESSLLI providesanappropriateforumfor fosteringcollaborationbetweeninterestedresearchersacrosstherelevantareas.

We hopethat the workshopparticipantswill transcendthe difficulties causedby proliferatingterminologies,andconcentrateinsteadon investigatingthe inter-actionsbetweenIS andDS,with thegoalof understandinghow sentence-level se-manticdevicesthatmakeupIS symbioticallyservetheneedsof discoursecohesion

v

Paperpresentedat theESSLLI2001WorkshoponInformationStructure, DiscourseStructure andDiscourseSemanticsEditedby IvanaKruijf f-Korbayova andMark Steedman.

andcoherence.We seekcontributionsadvancingbeyond descriptive frameworkstowardsanexplanatoryaccountof how IS andDS,in whatever framework, interactto referto andupdateadynamicallyevolving representationof discoursecontext.

In responseto the call for papers,we received 24 submissions,which havebeenreviewed by the programmecommittee,and of which we have decidedtoaccepteight for presentationand threeas a reserve. An importantcriterion inselectingamongthe extremelystrongsubmissionswasthe extent to which theyaddressedlinkagesbetweentheareasof informationstructure,discoursestructureanddiscoursesemantics.In addition to the authorsof the submittedpapers,wehave invited two distinguishedresearchersto give keynotespeeches,onefrom theareaof informationstructure,andonefrom theareaof discourseunderstanding.

We have aimedto assemblea coherentset of papers,addressingclosely re-latedissuesfrom variousperspectives,suchthat theexchangeof ideasamongtheworkshopparticipantscouldbemaximized.Theprogrammeof theworkshopcon-sistsof fivesessions,eachcontainingtwo to threepaperpresentationsfollowedbydiscussion.Here,webriefly overview thesessions:

Session1(Monday, August 20th) In an introductorytalk, Mark Steedmanat-temptsto lay out the evolution of IS in variousapproaches.Then, two papersfollow which representdifferentperspectives:

Rick Nouwen’s contribution representstherecentoptimality-theoreticstrandinformal discoursesemantics/pragmatics;he addressesthe problemof interpretingpluralpronounswith quantifiedantecedents,andformalizestheresolutionof theseanaphorsin termsof the marked-unmarked oppositionbetweenthe referencesetandthecomplementset.

Laurent Prevot’s contribution is couchedin the SegmentedDiscourseRepre-sentationtheoryapproach;basedon an investigationof a speechcorpusof routeexplanationdialogues,it is proposedto considerin parallelthedialoguestructure(reflectingrhetoricalandintentionalstructuring)andthetopic structure(reflectingtheconstructionof acommongroundbetweentheparticipants).

Session2 (Tuesday, August 21th) Enric Vallduvı addressesthe differencebe-tweenfragmentaryvs. full answersto questions,in termsof combininginforma-tion structureandanotionof accessibilityfor propositionaldescriptionsin context.

SvetlanaMcCoyalsodiscussestherelationbetweeninformationstructureandthecognitive statusof informationin context; sheexplainsthedistribution of sev-eralRussianparticlesin termsof cognitive statusandanotionof “kontrast”,whichsheconcludesservesto connectthelevel of clauseanddiscoursestructure.

Session3 (Wednesday, August 22th) Threepaperspoint to a tight relationbe-tweenthe IS-partitioningof an utteranceandthe relationsbetweenutterancesindiscoursecontext:

vi

Carla Umbach proposesananalysisof “but” whichcombinesits IS-sensitivityandits discoursefunctionof markingdenialwith respectto anissueunderdiscus-sion; shediscussessomedifferencesin distribution between“and” and “but” inoverinformative responsesto questions,andconcludesthat “but” marksa changeof theissueunderdiscussionin theresponsew.r.t. thequestion.

NoboKomagata comparesalternative analysesof IS-partitioningin complexsentences,andsupportstheview that IS is a non-recursive, matrix-level partition-ing. With regard to NLP systems,he concludesthat information structureanddiscoursestructureprocessingneedto proceedin parallelat somepoint.

Ivana Kruijff-Korbayova and BonnieWebberdiscussexampleswherethe in-terpretationof “otherwise” as a discourseanaphorrequiresaccessto conditionsderived on thebasisof the IS-partitioningof the relevantantecedent.It is arguedthat informationstructureaffectsdiscoursecontext by influencingwhat contextsareavailablefor theinterpretationof subsequentutterances.

Session4 (Thursday, August 23th) Livia Polanyi discussesdiscoursestructureandits relationto informationstructure,from both thespeaker’s andthehearer’sperspective, andarguesfor anintegratedapproachto discoursecontinuity.

MassimoPoesioand Barbara di Eugenio addressthe relationshhipbetweendiscoursestructureandanaphoricreference:usinga corpusof tutorial dialogues,they evaluatethecommonlyacceptedcorrelationbetweenanaphoricaccessibilityanddiscoursestructure.Their studyconfirmsan earlierproposalthat intentionalrelationsarerelevant for modellingattentionalstate,while informationalrelationsare not. But they also found evidencethat attentionalstatecannotbe properlymodelledasa stack,but only asa list.

Session5 (Friday, August 24th) KatherineForbeset al. presentan implemen-tationof a parsingsystemfor a lexicalisedTree-AjoiningGrammarfor discourse,integratingsentenceanddiscourselevel processing.Thesystemis basedontheas-sumptionthatthecompositionalaspectsof semanticsat thediscourse-level parallelthoseat thesentence-level. Thiscouplingis achievedby factoringaway inferentialsemanticsandanaphoricfeaturesof discourseconnectives.

JenniferSpenader’s empiricalstudyof factive verbsin spokenEnglishhasre-vealedthatanoverwhelmingmajority of themoccurwith full propositionalcom-plementsexpressingnew information,thatneedsto beaccommodated.Shecom-parespresupposedcomplementswith abstractobjectanaphora,anddraws implica-tionsfor ageneraltheoryrelatingpresupposition,anaphoraandaccommodation.

Finally, in theclosingdiscussion,anattemptis madeto summarizetheimpli-cationsof thecontributionspresentedat theworkshop.

IvanaKruijff-Korbayova andMark SteedmanJuly 2001

vii

Tableof Contents

Intr oduction: Mark SteedmanandIvanaKruijf f-Korbayova:Two Dimensionsof InformationStructure in Relationto DiscourseStructure andDiscourseSemantics 1

Livia Polanyi, Martin vandenBerg andDavid Ahn: A Noteon theRelationshipof DiscourseStructure to InformationStructure 11

Enric Vallduvı: Fragmentsin InformationPackaging 15

KatherineForbes,EleniMiltsakaki,RashmiPrasadAnoopSarkar,Aravind JoshiandBonnieWebber:DiscourseParsingwith a LexicalizedTree-AdjoiningGrammar(D-LTAG) 17

JeanetteGundel,MichaelHegarty, KajaBorthen:Informationstructure andpronominalreferenceto clausallyintroducedentities 37

NoboKomagata:EntangledInformationStructure: Analysisof ComplexSentenceStructures 53

IvanaKruijf f-Korbayova andBonnieWebber:InformationStructure andtheSemanticsof “otherwise” 67

ValeriaMolnar: Contrastfroma ContrastivePerspective 99

SvetlanaMcCoy: ConnectingInformationandDiscourseStructure Levelsthrough“K ontrast”: EvidencefromColloquialRussianParticles-TO, ZHE,andVED’ 85

Rick Nouwen:Complementsetreference 115

MassimoPoesioandBarbaradi Eugenio:DiscourseStructure andAnaphoricAccessibility 129

LaurentPrevot: TopicStructure in RouteExplanationDialogues 145

JenniferSpenader:Presuppositionor Anaphora: Constraintson Choiceof FactiveComplementsin SpokenDiscourse 161

CarlaUmbach:RelatingContrastandContrastiveTopic: a focus-semanticanalysisof “b ut” 175

Klausvon Heusinger:IntonationalPhrasingandDiscourseSegmentation 189

ix

Intr oduction

MARK STEEDMAN

Informatics,Universityof Edinburgh,2 BuccleuchPlace,EdinburghEH89LW, U.K.

[email protected]

IVANA KRUIJFF-KORBAYOVA

Computerlinguistik,Universitat desSaarlandes,Saarbrucken,Germany

[email protected]

Two Dimensionsof Inf ormation Structure in Relation toDiscourseSemanticsand DiscourseStructure

Thefollowing introductionis intendedto do no morethanto establishsomebasicdistinctionsanddefinitionswithin which we believe all the theoreticalpositionsrepresentedat the workshopcanbe comparedand/orreconciledand their oftencloserelationshipsto eachotherbeunderstood.

1 SomeTentativeDefinitions

The title of the workshopimplies a three-way distinction betweenInformationStructure,DiscourseSemantics,andDiscourseStructure.While onoccasionthesetermshave beenusedinterchangeablyor in overlappingsenses,we believe that itis usefulto distinguishthemasfollows:� Information Structureis a SentenceInter nal partition of the information

in an utteranceaccordingto its relation to the discoursecontext underdi-chotomiessuchastopic/comment,theme/rheme,given/new, focus/background,etc.SuchcategoriesareessentiallyReferential in nature.� DiscourseSemanticscentrallyconcernsthenatureof thecontextual model,andtheentitiesin it to which InformationStructuralcategoriesrelate,in theform of alternative sets,filecards,presupposedpropositions,etc.

1


� DiscourseStructureconcernstheInter -clausalrelationsof explanation,elab-oration,exemplification,andillocutionary force that hold betweensucces-sive utterancesof a discourseor dialog,supportinginferenceaboutthedo-mainandpurposesof thediscourse.

2 A Computational Model

We have tried in settingup the workshopto keepcomputationalmodelsof dis-courseanddialoguein focus. In particularwe believe it maybeusefulto think ofthesedistinctionsin termsof someextensionto GroszandSidner’s ((GroszandSidner(1986)))computationalaccount.Specifically:� Informationstructureseemstobelongin whatthey call “linguistic structure”—

essentially, thegrammar.� Its DiscourseSemanticextensionsseemto concernthe “AttentionalState”component.(Althoughthestructuralrepresentationof thisstateasastackoffocusspacesseemsquestionable,andtheontologyof typesthatmayinhabitfocus spacesmust be greatly extended—seePoesioand Di Eugenio,andForbesetal. below.)� Otherrhetoricalandillocutionary aspectsof DiscourseStructureappeartorelateto their “IntentionalStructure”component.

3 The Two Dimensionsof Inf ormation Structure

The terminologythat is usedto describeInformationStructureandits semanticsis simultaneouslyvarious,and under-formalized. Yet it seemsthat all defini-tions have someelementsin common. They all draw at leastoneof the follow-ing distinctions:(i) a “topic/comment”or “theme/rheme”distinctionbetweenthepart of the utterancethat relatesit to the purposeof the discourse,and the partthat advancesthe discourse;(ii) a “given/new” distinction,betweenpartsof theutterance—actually, words—whichcontribute to distinguishingthe contentfromotheralternativesthatthecontext makesavailableandthosepartsthatarecommonto all of them.

Thereare differencesamongthe theoriesof course. Some,like Halliday’s((Halliday(1967))),view thesetwo distinctionsasorthogonal,applyingat indepen-dentlevelsof structure.Othersin theBolingeriantradition,like Vallduvı, McCoy,Molnar, andourselves in the presentvolumeview themasdifferentaspectsof asinglelevel of structure.An importantissuethat further differentiatestheBolin-geriantheoriesis thatof whetherpitch accentcorrespondsto a singlecontrastivenotionof focusapplyingto boththemeandrheme,or whether“contrastive focus”is a distinct notion,applyingto explicitly mentionedentitiesandassociatedwithtopic or themealone(seeMolnar, andUmbach,below). Many of thesequestions

2

await fuller empiricalandcorpusbasedstudies:we welcomethe studiesof thiskind by Spenader, McCoy, Prevot, andPoesioandDi Eugenio.

Therearefurthersimilarities: while someof the theoriesleave theassociateddiscoursesemanticsat anintuitive level, the theorieswhich do addressformal se-manticissuesall tendto usesomeversionof “update”semantics,suchasthefile-cardsemanticsandsegmentedDRSsof theKampo-Heimiansynthesis(seePrevot,vonHeusinger, andGundeletal.,below), or theAlternativesemanticsof RoothandBuring (seeNouwen,Komagata,Umbach,andKruijf f-Korbayova andWebber, inthepresentcollection).

The diagramon the next pagedisplayingour view of the influencesandter-minologicaldependenciesin theoriesof InformationStructureandtheassociatedDiscourseSemanticsmayhelpto make thesecommonalitiesmoreobvious.

4 An Example that ExemplifiesBoth Dimensions

Theterminologythatweuseourselvesmayserveto illustratethesetwo dimensionsatwork. (Wetakethetermsthemeandrhemefrom FirbasandBolinger, ratherthanHalliday, for our topic/commentdistinction,anduseDahl’s backgroundandfocusfor theHallideangiven/new distinction,andapplytheseelementsin a singlelevelof informationstructure.)

(1) Q: I know thatMary likesthemanwhowrotethemusical.But whodoessheADMIRE?

A: (Mary ADMIRES) (thewomanwho DIRECTED themusical)� �� L�

H � LH%� �� H �� LL%� �� Background Focus� �� Background Focus Background� ��

Theme Rheme

In Englishwe have claimedthat InformationStructureis homomorphicto In-tonationStructure,andthat thecontourdescribedin Pierrehumbert’s (Pierrehum-bert(1980))notationasL+H* LH% is oneof the “themetunes”that identify theintonationalphraseMary admiresasthe themeof this utterance,while H* LL%is a “rhemetune”. Within both the themeandthe rheme,the presenceof oneormorepitch accentsidentifieswordswhich contribute to distinguishingthat themeand/orthatrhemefrom otherthemesandrhemesthatthecontext affords.Thuswefollow (Bolinger(1965))and(Lambrecht(1994))in viewing the role of accentinEnglishasasingleundifferentiatedcontrastive meaningapplyingto bothinforma-tional components.

Wethink of thatcontext asanupdatabledatabaseincludingtwo Roothianalter-nativessets,respectively calledthe“ThemeAlternativeSet”andthe“RhemeAlter-nativeSet”—cf. “AlternativeSemantics”(Rooth(1985);Rooth(1992)),(Buring(1995)).

In CombinatoryCategorialGrammar(Steedman(2000)),InformationStructureandtheassociatedstructuredmeaningsareassociateddirectlywith surfacesyntac-tic derivationalstructure,uniting InformationStructure,IntonationStructure,andSurfaceSyntaxin asinglemoduleof Grammar.

3

..B

urin

g 19

95

(Hal

liday

& H

asan

197

6)

nucl

eus/

focu

skn

own/

unkn

own

Firb

as 1

964,

196

6th

eme/

rhem

e

Hal

liday

196

7th

eme’

/rhe

me’

give

n/ne

w (

orth

ogon

al)

back

grou

nd/fo

cus

Haj

icov

a, P

arte

e, &

Sga

ll 19

98

pres

uppo

sitio

n

Cho

msk

y 19

65B

olin

ger

1965

them

e/rh

eme,

acc

ent

pres

uppo

sitio

n/fo

cus

Kar

ttune

n 19

68

Cho

msk

y 19

70/J

acke

ndof

f 197

0

Kar

ttune

n &

Pet

ers

1979

pres

uppo

sitio

n/fo

cus

(alte

rnat

ive

set)

Roo

th 1

985

topi

c/fo

cus

C/Q

alte

rnat

ives

set

Sel

kirk

198

4

topi

c/fo

cus,

Val

lduv

i 199

0lin

k/ta

il/fo

cus

topi

c/co

mm

ent (

orth

ogon

al)

topi

c/co

mm

ent

back

grou

nd/fo

cus

cont

ext b

ound

/unb

ound

cont

ext d

epen

dent

/inde

pend

ent

Dah

l 196

9

Mat

hesi

us 1

929

(Rus

sell

1905

)

topi

c/co

mm

ent

(Str

awso

n 19

50, 1

954)

(Grim

es 1

975)

(Man

n &

Tho

mps

on 1

987)

(Bro

wn

1983

)

Ste

edm

an 1

991

them

e/rh

eme,

Cha

fe, C

lark

,G

unde

l, P

rince

Kay

197

5gi

ven/

new

topi

c/co

mm

ent

give

n/ne

w’ (

orth

ogon

al)

Val

lduv

i & V

ilkun

a 19

98th

eme/

rhem

e,0/

kont

rast

Hen

drik

s 19

99lin

k/ta

il/fo

cus

pres

uppo

sitio

n/na

rrow

focu

s,K

rifka

, Kra

tzer

wid

e fo

cus

(Win

ogra

d, W

oods

)

topi

c/fo

cus,

Sga

ll 19

67

cont

ext b

ound

/unb

ound

(Sac

ks, S

cheg

loff

& J

effe

rson

197

4)

(str

uctu

red

mea

ning

s,D

RT

)

Kam

p, H

eim

)(C

ress

wel

l, vo

n S

tech

ow

(Mon

tagu

e 19

73)

Gro

sz, J

oshi

& W

eins

tein

)(P

ierr

ehum

bert

& H

irsch

berg

,

(Pol

anyi

and

Sch

a 19

83)

(Gro

sz &

Sid

ner,

Web

ber)

4

5 SomeQuestionsfor the Workshop

In termsof thiswehopeuncontentioussystematizationof thetermsof referenceoftheworkshop, thefollowing questionscanbeofferedfor considerationamongthemany othersraisedby thepresentersat theworkshop:� Whatis theDiscourseSemanticsof InformationStructure?� WhatpartdoesInformationStructureplay in DiscourseStructure?� How doesInformationStructurerelateto IntonationStructurein Englishand

relatedlanguages?� How doesInformationStructurerelateto SyntaxandSemanticsin the lan-guagesof theworld?� Your questionhere

6 SomeAnswersthat Emerged� Whatis theDiscourseSemanticsof InformationStructure?

Thereseemedto be two flavors of semantics:Forbeset al. , Korbayova &Webber, andMcCoy usedformsof Alternative Semantics.Prevot; Vallduvı,andPoesiowerein theKampo-HeimianStructuredMeaningtradition. Thisappearsto bea mainly notationaldifferenceandthesystemsseemin somesenseequivalent.� Whatpart doesInformationStructure play in DiscourseStructure?Many DiscoursestructuralandDiscourseSemanticphenomenadiscussedattheworkshopappearedto dependatleastin partononinformation-structuralrepresentationsin the context, including anaphora,presupposition,andthemeaningof connectiveslike but, otherwiseandalthough, anddiscoursepar-ticleslike theRussianzhe, ved’. (Polanyi; Spenader;Poesio,Nouwen;Kor-bayova & Webber;Umbach)� How doesInformation Structure relate to Intonation Structure in Englishandrelatedlanguages?Several papersclaimedor werecompatiblewith the view that IntonationStructurein English is identical to an underspecifiedform of InfomationStructure,and in particular that intonationalphrasescorrespondto infor-mationunits. (Vallduvı; Umbach;Prevot; McCoy; Komagata;Steedman.)Therewere interestingquestionsraisedaboutdialectvariation,the consis-tent specificationof informationstructuralrole by prosodictones,and thestatusof All-Themeutterances(Kamp;Gundel.)

5

� How doesInformationStructure relateto SyntaxandSemanticsin the lan-guagesof theworld?Therewasgeneralagreementthat thesamepair of orthogonaldistinctions–in Vallduvıand Vilkuna’s terms, theme/rhemeand Kontrast/ground—wereat work in other languages,marked by otherdevicessuchasmorphology,verbalparticles,word-orderetc. (Prevot; Komagata;McCoy; Molnar; vonHeusinger). Therewas an openquestionhow doesUmbach’s quæstiofitin—is it a theme?

7 Further Questionsthat Emerged� How areContextual Representationsto beManaged?

Contexts, whetherviewed ascollectionsof alternative setsor of DRSsorfilecardshaveacomplex structure.Alternativerepresentationsof AttentionalStatewerediscussed,including thoserelatedto computationaldialogman-agement,suchasthatproposedin theTRINDI framework, whichmaintainslists of QUDs. ThemesandQUDs seemclosely related. The questionofrepresentingchangingsalienceof discourseobjectswasalsodiscussed.� TheRoleof AccomodationAccomodationwascentralto anumberof proposals.Whattypesof dicourseentity canaccomodate,andwhenis accomodationblocked?� TheRoleof InformationStructure PartitioningWhatis thelinguisticunit of discoursesstructure?It seemedclearthethemesandrhemescouldbediscontinuouswithin asingleutterance.Thequestionofwhetherany suchunitswereobligatoryremainsunresolved.Thequestionofhow to reliably identify informationalstructuralunits in real texts remainedproblematic,but anumberof testswereproposedincludinglookingateffectsof systematicallyalterning intonation(and equivalent parametersin otherlanguages)and the effect on the interpretationof IS-dependentdiscourseconnectiveslike although, focusparticles,etc.� Whatis thebestempiricalagreementfor investigatingthesequestions?Therewasgeneralagreementthatdialog,ratherthanextendedexposition,wasthedomainthatreallystretchedthetheories.

Bibliography

Bolinger, Dwight. 1965.Formsof English. Cambridge,MA: HarvardUniver-sity Press.

Brown,Gillian. 1983.“ProsodicStructureandtheGiven/New Distinction.” InAnneCutler, D. RobertLadd,andGillian Brown, eds.,Prosody:ModelsandMeasurements, 67–77.Berlin: Springer-Verlag.

6

Buring, Daniel. 1995.The 59th Street Bridge Accent: On the MeaningofTopicandFocus. Ph.D.thesis,Universitat Tubingen.Publ.as(?).

Chafe,Wallace.1976.“Givenness,Contrastiveness,Definiteness,Subjects,TopicsandPointsof View.” In CharlesLi, ed.,SubjectandTopic, 25–55.New Tork: AcademicPress.

Chomsky, Noam.1965.Aspectsof the Theoryof Syntax. Cambridge,MA:MIT Press.

Chomsky, Noam.1971. “Deep Structure,SurfaceStructure,and SemanticInterpretation.” In Danny Steinberg andLeonJakobovits, eds.,Semantics,183–216.Cambridge:CambridgeUniversityPress.

Clark, Herbert,andSusanHaviland. 1977.“ComprehensionandtheGiven-New Contract.” In R. Freedle,ed.,Discourseproductionandcomprehen-sio, 1–40.Norwood,NJ: Ablex.

Cresswell,M.J.1973.LogicsandLanguages. London:Methuen.

Dahl, Osten.1969.Topic andFocus:a Studyin RussianandGeneral Trans-formationalGrammar. Goteborg: ElandresBotryckeri.

Firbas,Jan.1964.“On DefiningtheThemein FunctionalSentenceAnalysis.”TravauxLinguistiquesdePrague, 1, 267–280.

Firbas, Jan. 1966. “Non-thematic Subjects in ContemporaryEnglish.”TravauxLinguistiquesdePrague, 2, 229–236.

Grimes,Joseph.1975.TheThreadof Discourse. Mouton.

Grosz,Barbara,Aravind Joshi,and Scott Weinstein.1995. “Centering: AFramework for Modeling theLocal Coherenceof Discourse.” Computa-tional Linguistics, 2, 203–225.

Grosz, Barbara,and CandaceSidner. 1986. “Attention, Intention and theStructureof Discourse.” ComputationalLinguistics, 12,175–204.

Gundel,Janet.1974.TheRoleof Topic and Commentin Linguistic Theory.Ph.D.thesis,Universityof Texas,Austin.

Hajicova, Eva, BarbaraH. Partee,andPetrSgall.1998.Topic-FocusArticu-lation, Tripartite Structures,andSemanticContent. Dordrecht:Kluwer.

Halliday, Michael.1967.“Noteson Transitivity andThemein English,PartII.” Journalof Linguistics, 3, 199–244.

Halliday, Michael,andRuqaiyaHasan.1976.Cohesionin English. Longman.

Heim, Irene.1982.Thesemanticsof definitesandindefinitenounphrasesinEnglish.. Ph.D.thesis,Universityof Massachussetts,Amherst.

Hendriks, H. 1998. “Links without Locations.” In F. Hamm and E. Zim-mermann,eds.,Linguistische Berichte, Sonderheft9/1998: SEMANTIK.Opladen:WestdeutscherVerlag.

7

Jackendoff, Ray. 1972. SemanticInterpretation in Generative Grammar.Cambridge,MA: MIT Press.

Kamp, Hans.1984. “A Theory of Truth and SemanticRepresentation.” InJeroenGroenendijk,TheoJanssen,andMartin Stokhof,eds.,Truth, In-terpretation,andInformation, 1–41.Dordrecht:Foris.

Karttunen, Lauri. 1968. “What Makes Definite Noun PhrasesDefinite?”Tech.rep.,RandCorporation.

Karttunen,Lauri, andStanley Peters.1979.“ConventionalImplicature.” InChoon-Kyu Oh andDavid Dinneen,eds.,SyntaxandSemantics11: Pre-supposition, 1–56.New York: AcademicPress.

Kay, Martin. 1975.“SyntacticProcessingandFunctionalSentencePerspec-tive.” In RogerSchankandBonnieNash-Webber, eds.,Theoretical Issuesin Natural Language Processing, (Supplement), 12–15.MIT, June.

Kratzer, Angelika.1991.“The Representationof Focus.” In Arnim von Ste-chow andDieter Wunderlich,eds.,Semantics:An InternationalHand-bookof Contemporary Research, 825–834.Berlin: deGruyter.

Krifka, Manfred.1991.“A CompositionalSemanticsfor Multiple FocusCon-structions.” Cornell WorkingPapers in Linguistics, 10,127–158.

Lambrecht,Knud. 1994. InformationStructure and SentenceForm: Topic,Focus, and the Mental Representationsof Discourse Referents. Cam-bridge:CambridgeUniversityPress.

Mann,William, andSandraThompson.1987.“Rhetoricalstructuretheory:Aframework for theanalysisof texts.” Tech.Rep.RS-87-185,InformationScienceInstitute,Marinadel Rey, CA.

Mathesius,Vilem. 1929. “Functional Linguistics.” In Josef Vachek, ed.,Praguiana: SomeBasicandLessWell-knownAspectsof thePragueLin-guisticsSchool, 121–142.Amsterdam:JohnBenjamins.Publ.1983.

Montague,Richard.1973.“The ProperTreatmentof Quantificationin Ordi-naryEnglish.” In Jaakko Hintikka, J. M. E. Moravcsik, andPatrick Sup-pes,eds.,Approaches to Natural Language: Proceedingsof the 1970Stanford Workshopon Grammar and Semantics, 221–242.Dordrecht:Reidel.Reprintedin Montague1974,247-279.

Pierrehumbert,Janet.1980.ThePhonology and Phoneticsof English Into-nation. Ph.D.thesis,MIT. Distributedby IndianaUniversity LinguisticsClub,Bloomington.

Pierrehumbert,Janet,andJulia Hirschberg. 1990.“The Meaningof Intona-tionalContoursin theInterpretationof Discourse.” In Philip Cohen,JerryMorgan, andMarthaPollack, eds.,Intentionsin Communication, 271–312.Cambridge,MA: MIT Press.

8

Polanyi, Livia, andRemko Scha.1983.“The Syntaxof Discourse.” Text, 3,261–270.

Prince,Ellen. 1981.“Towardsa Taxonomyof Given-New Information.” InPeterCole, ed., Radical Pragmatics, 223–256.New York: AcademicPress.

Rooth,Mats.1985.Associationwith Focus. Ph.D.thesis,Universityof Mas-sachusetts,Amherst.

Rooth,Mats. 1992.“A Theoryof FocusInterpretation.” Natural LanguageSemantics, 1, 75–116.

Russell,Bertrand.1905.“On Denoting.” Mind, 14,479–493.

Sacks,Harvey, EmanuelSchegloff, and Gail Jefferson.1974. “A SimplestSemanticsfor the Organizationof Turn–Taking for Conversation.” Lan-guage, 50,696–735.

Selkirk,Elisabeth.1984.PhonologyandSyntax. Cambridge,MA: MIT Press.

Sgall,Petr. 1967.“FunctionalSentencePerspective in a Generative Descrip-tion.” PragueStudiesin MathematicalLinguistics, 2, 203–225.

von Stechow, Arnim. 1991. “Focusingand BackgroundingOperators.” InWernerAbraham,ed.,DiscourseParticles: DescriptiveandTheoreticalInvestigationsontheLogical, SyntacticandPragmaticPropertiesof Dis-courseParticlesin German, 37–84.Amsterdam:JohnBenjamins.

Steedman,Mark. 1991.“StructureandIntonation.” Language, 67,262–296.

Steedman,Mark.2000.“InformationStructureandtheSyntax-PhonologyIn-terface.” LinguisticInquiry, 34,649–689.

Strawson,Peter. 1950.“On Referring.” Mind, 59,320–344.

Strawson,Peter. 1954.“A Reply to Mr. Sellars.” PhilosophicalReview, 63,216–231.

Vallduvı, Enric. 1990.TheInformationComponent. Ph.D.thesis,Universityof Pennsylvania.

Vallduvı, Enric,andMariaVilkuna.1998.“On RhemeandKontrast.” In PeterCulicover andLouiseMcNally, eds.,SyntaxandSemantics,Vol. 29: TheLimitsof Syntax, 79–108.SanDiego,CA: AcademicPress.

Webber, Bonnie.1988.“TenseasDiscourseAnaphor.” ComputationalLin-guistics, 14,61–73.

Winograd,Terry. 1972.UnderstandingNatural Language. Edinburgh: Edin-burgh UniversityPress.

9

Invited Talk:A Note on the relationship of discoursestructur eto information structur e

L IV IA POLANYI

FXPAL, 3400Hillview, Palo Alto, CA 94304

[email protected]

MARTIN VAN DEN BERG

FXPAL, 3400Hillview, Palo Alto, CA 94304

[email protected]

DAVID AHN

Departmentof ComputerScience,Universityof Rochester, Rochester, NY 14627

[email protected]

1 Intr oduction

Althoughit is generallyacceptedthata sentencesinformationstructure(IS) is de-terminedby its relationshipto previous text, thequestionof how to establishtheappropriatediscoursecontext for IS assignmentis neverraised.Analysesof IS nor-mally assumethatthatsentenceis a questionandthetargetsentenceis theanswerto that question[1]. The assumptionis alwaysthat the prior context for a targetsentenceis the immediatlyprecedingsentence.While this assumptionmay be aconvenientconventionfor investigatingthesubtleraspectsof intrasentialIS, seri-ousproblemsarisein analyzingactualtextswhenthecritical context for assigningIS is notpartof theimmediatelyprecedingsentence.For thesecases,it isnecessaryto have a theoryof how discoursestructure(DS) constrainsthechoiceof possiblecontexts for determininingsententialIS. This noteaddresseshow discoursestruc-tureandinformationstructure(IS) arerelatedandsketchesanintegratedapproachto thephenomenaof discoursecontinuity.

In brief, our claim is that DS constrainsthe setof possibleconstituentsin adiscoursethatcanprovide therelevantcontext for structuringinformationin a tar-getsentence,while IS critically constrainsDS ambiguity. To developanargument

11


to supportthis claim, we rely uponthe Linguistic DiscourseModel (LDM) asatheoryof DS [2], andfollow Steedmanin assumingthatsentenceIS consistsof athemeandrheme,eachfurthersubdividedinto backgroundandfocus[3, 4].

2 DiscourseStructure and Inf ormation Structure

UndertheLDM, therepresentationof adiscourseis constructedincrementallyus-ing informationin thesurfacestructureof incomingutterancestogetherwith dis-courseconstructionrulesandinferenceover themeaningof the utterancesto re-cursively constructanopen-righttreeof discourseconstituentunits(DCUs). Thistreeindicateswhich unitsareaccessiblefor continuationandanaphoraresolution.All nodesonanLDM treearefirst classobjectscontainingstructuralandsemanticinformation;terminalnodescorrespondto thestringsof thediscourse,while non-terminalsare constructednodeslabeledwith a discourserelation (coordination,subordination,or n-ary). IS is representedat terminalsandnon-terminalsaswell.A C-nodeinheritsthegeneralizationof thethemesof its constituentnodesandoftheir rhemes.An S-nodeinheritstheIS of its subordinatingdaughterdirectly.

Ourinitial hypothesis,illustratedby theexampletext andaccompanying chart1,is that the attachmentis (1) a C-nodeif the themeof the main clauseof the newsentencematchesthematicinformationavailableat theattachmentpoint or (2) anS-nodeif the themeof themainclauseof thenew sentencematchesrhematicin-formationavailableat theattachmentpoint2.

In analyzinga discourse,discoursesyntaxassignseachincomingsentenceitsplacein the emerging discoursetree. In currentapproaches,lexical information,syntacticandsemanticstructure,tenseandaspect,andworld knowledgeareusedto infer theattachmentpointandrelation(c.f. [5]). However, afterexploiting theseresources,attachmentambiguitiesoftenstill remain. Given thatnormallanguageusersseldomexperiencediscourseattachmentambiguities,additionalsourcesofinformationmustbeusedin attachmentdecisions.We believe that the IS of boththeincomingsentenceandaccessibleDCUs providesinformationcritical for dis-ambiguation.Theproblemof identifying thetargetDCU thatprovidesthecontextfor IS assignmentfor an incomingsentenceis analogousto anaphoraresolution:thetargetunit mustbealongtheright edgeof thetreeandthereforeaccessible[6].

Froma discourseperspective, the IS of anincomingsentencedividesit into atheme,which mustbelinkedbackto theprecedingdiscourse,anda rheme,whichneednot be. Establishinga link betweenthe themeof the main clauseof a newsentenceand information available at an accessiblenodein the tree determines

1Here,for thesake of presentationalsimplicity, constituentsareassumedto besentences.how-ever, undertheLDM, themuchmorefinely-grainedDCU segmentationconventionsenablesubor-dinateclausesto serve asattachmentpointsfor themainclausesof subsequentsentences.

2Herewe will notdiscussn-arynodesthatareusedto representthestructureof dicsoursegenresas well as conversationaladjacency structuresand logical relationsfurther. It is enoughfor ourpurposesto mentionthatthey follow moread-hoc,thoughwell-defined,rules.

12

thesentencesattachmentpoint. The typeof attachment,coordination,subordina-tion, or n-ary, reflectsthethemesrelationto theIS of theDCU representedat theattachmentnode.

3 Analysis of an exampletext

In theexample,themesaremarkedwith a θ; rhemesareunmarked. Wordsreceiv-ing stressarein SMALL CAPS.

1. (Japanesepeopleoccasionallychooseto eat)θ NOODLES. 2. (Noo-dlesareUSUALLY eaten)θ for LUNCH or a light SNACK. 3. Depend-ing on the SEASON, (noodlesmight be served)θ in a HOT SOUP orCOLD like a salad. 4. (When noodlesare served in a hot SOUP,)θVEGETABLES, TOFU, andMEAT areALSO foundwithin thesoup. 5.Several TYPES of noodles(are eatenIN JAPAN.)θ 6. (UDON)θ areTHICK , WHITE noodlesmadefresh from wheatflour and are USU-ALLY served with a hot SOUP. 7. (SOBA)θ are THIN BUCKWHEAT

noodleswhich are FIRMER thanudon. 8. (They canbe served in aSOUP like UDON,)θ but are USUALLY served asa COOL dish in theSUMMER. 9. (RAMEN)θ arevery thin, CURLY wheatnoodlesservedasa QUICK mealor a LATE night SNACK. 10. (Noodlesareeaten)θ asa VARIATION for thedaily MEAL.

Sentence# 1 2 3 4 5Theme Japanese Noodles Noodles Noodles Eaten

people hot IN JAPAN

eat eaten served soupRheme NOODLES LUNCH SEASON VEGETABLES TYPES

SNACK HOT SOUP TOFU

COLD MEAT ALSO

Attachment N/A 1 2 3 (S2– (S3–4))Relation N/A S S S C

Sentence# 6 7 8 9 10Theme UDON SOBA The RAMEN Noodles

SOUP � �[UDON]

Rheme THICK WHITE THIN USUALLY � VERY VARIATION

USUALLY BUCK- COOL CURLY �SOUP WHEAT SUMMER QUICK LATE MEAL

Attachment 5 6 7 (C 6–(S7–8)) (C (S2 � )–(S5 )) (S5 ))

Relation S C S C C

As the chart indicates,(1)-(4) exhibit theme-rhemechaining,resultingin nestedsubordinations.For (5), theappropriatecontext for IS assignmentis provided by(2), with a theme-themelink resultingin a coordination.Therhemeof (5) inten-sionally introducesa setof typesof noodlespickedup asthethemealternative set

13

for (6), (7), and (9). The themefocus for eachof thesesentences(udon, soba,ramen) is presupposedto belongto this set. Thesesentencesarethereforecoor-dinatedto eachotherandsubordinatedto (5). Processing(8) demonstratesthatboth DS andIS may operateautonomously. The IS of (8) is determinedprimar-ily by theconjunctionbut whichactswith thepossibilitymodalin its first conjunct(whichprovidesanaccessiblesetof possibleworldsastherhemealternativeset)toconstructa theme-rhemepair, while discourseattachmentof (8) fulfills anaphoraresolutionrequirementsratherthanIS. For (10), (5) providestheappropriatecon-text for theIS assignment.Thetheme-themelink resultsin acoordinationthatpopsthestateof thediscourseseveral levels.

4 Conclusion

Althoughtheassignmentof IS to a sentencedependson theDS,andtheconstruc-tion of theDSmaydependontheIS of theunitsinvolved,thedependency betweenIS andDS is complementaryandnot circular. For thespeaker, theDS providesasetof possiblecontexts for continuationwhile IS assignmentis independentof DS.For thehearer, theIS of a sentencetogetherwith DS instructsdynamicsemanticshow rhematicinformationshouldbeusedto updatethemeaningrepresentationofthediscourse(c.f. [7]). Thus,therelationshipbetweenDS andIS reflectsthedif-ferentbut deeplyrelatedtasksof speaker andhearerin acommunicative situation.

References

[1] Vallduvı, Enric. (1994)InformationPackaging:A Survey, WordOrder, ProsodyandInformationStructureProjectReport.Universityof Edinburgh.[2] Polanyi, L. andR. Scha(1984)A syntacticapproachto discoursesemantics,COLING’84.[3] Steedman,M. (1991)Structureandintonation,Language,68.[4] Steedman,M. (to appear)Informationstructureandthesyntax-phonologyin-terface,Linguistic Inquiry, 31:4.[5] Lascarides,A. andN. Asher(1993)Temporalinterpretation,discourserelationsandCommonsenseEntailment,LinguisticsandPhilosophy, 16:5.[6] Polanyi, L. andM. v. d. Berg (1999)Logical structureanddiscourseanaphoraresolution,ACL99 Workshopon Discourse/DialogueStructureandReference.[7] Polanyi, L. andM. v. d. Berg (1996)Discoursestructureanddiscourseinter-pretation,10thAmsterdamColloquium.A numberof text typesaredistinguishedby beingaboutoneperson.In anobituaryor letterof recommendation,for example,nootherpersonmayevenbe

14

Invited Talk:Fragmentsin Inf ormation Packaging

ENRIC VALLDUVI

DeptdeTraduccio,UniversitatPompeuFabra

[email protected]

Paperson informationstructureoftenmakeuseof question-answerpairslike(1) and(2) to illustratethesententialpartitionbetweenrheme(focus,new info) andground(theme,presupposition,old info, topic):

(1) a. Who likesbeer?

b. [R JOHN] [G likesbeer.]

(2) a. WhatdoesJohnlike?

b. [G Johnlikes][R BEER.]

The ideais that thequestionis theexponentof someunderspecifieddescrip-tion in a context or informationstate.In (1a),for instance,aninstantiationfor theexperiencerparameteris needed.Therhemein (1b) is thestructuralelementthat”answersthequestion”in thesensethatit expressesavaluefor this uninstantiatedparameter, while thegroundsimplyexpressesthepiecesof meaningin thisdescrip-tion thatarealreadyavailablefrom context. Let uscall thepiecesof meaningthatmake a contribution to context the”update(potential)”andthepiecesof meaningthatarealreadyavailablein context the”base”.

Thefactis oftenneglectedthat(1b)and(2b)arenotdiscourse-naturalanswersto (1a) and(2a). In fact, themostnaturalanswersto (1-2a)arethosein (3) (andalsothosein (4), seebelow):

(3) a. JOHN.(answerto (1a))

b. BEER.(answerto (2a))

(4) a. JOHNdoes.(answerto (1a))

b. He likesBEER(answerto (2a))

A full accountof informationpackagingshouldtake this dialogicalevidenceinto account. The differencebetween(1-2b) and (3) on the other is clearly notin their updatepotential. Still, we needto accountfor the fact that (3) is a morenaturalanswerto (1-2a)than(1-2b).

15


The proposaldefendedhereis that the differencebetween(1-2b) and (3) isin the natureof the base,namely, in the accessibilitystatusof the base. In (1-2) the immediate,explicit questionsmake thebasefully accessibleat the time ofutterance,so a groundis not needed. This follows from a view of groundsascontext-anchoringelements.Thegroundis thestructuralexpressionof a base,butsuchexpressionis only neededwhentheanchoringof theupdateto agivenbaseisnot self-evidentat thetime of utterance.In question-answerexchanges,wherethebaseis fully accessiblefrom context, groundsaresuperfluous;this is why rheme-only fragmentutterances,which expressonly theupdate,aresocommonasdirectanswersto questions.Of course,thegroundmayberedundantlyusedasananchorin contextswhereit is notneeded,but thisoftengivesriseto secondaryinterpretiveeffects,sometimestriggeredby theviolationof Griceanmaxims.

Whatweneedisatheorythatallowsusto establishsomenotionof accessibilityfor propositionaldescriptionsin context, akinto thenotionof accessibilityassumedin accountsof anaphoraresolution.Groundsareunnecessary/disfavored whenthebaseis within somelocalattentionaldomain:thelink betweentheupdatepotentialexpressedasrhemeandthe basein the context proceedsas the resultof certaindefault dynamicsof dialogue.However, groundsbecomenecessarywhenthebaseof agivenupdateis outsidethis local attentionaldomain:theunificationof updateandbasecannottakeplacewithouttheexplicit anchoringeffectedby theground.Inasense,then,rheme-onlyfragmentsareto rheme-groundsentenceswhatpronounsareto full definitedescriptionsin a theoryof referentialaccessibility.

Theanswersentencesin (4) above behave morelike sentencefragmentsthanfull answerswith respectto informationpackaging(their dialogic distribution iscloserto (3) than(1-2b)).Clearly, theexamplesin (4) arenot syntacticfragments,but their behavior is accountedfor if they aretreatedasinformationalfragments;in otherwords,theanswersin (4) do not have a ground. The weakpronounandthe verbalelementsthat appearin (4) are thereexclusively for morphosyntacticreasonsanddo not carry out theanchoringfunction thatgroundscarry out. Thisleadsusto proposetoo that(5b) is asmuchof aninformationalfragmentas(6b) is(although(5b) is obviously notasyntacticfragment):

(5) a. How doeshefeelaboutBill?

b. He LOVEShim.

(6) a. Who doesJohnlove?

b. BILL.

Both (5b) and(6b) areanswersthatdo not needa groundto make theconnectionexplicit betweentheirupdatepotentialandthepertinentbasein context.

16

D-LTAG System- DiscourseParsing with aLexicalizedTreeAdjoining Grammar

KATHERINE FORBES ELENI M ILTSAKAKI RASHMI PRASAD

Universityof Pennsylvania Universityof Pennsylvania Universityof Pennsylvania

[email protected] [email protected] [email protected]

ANOOP SARKAR ARAVIND JOSHI BONNIE WEBBER

Universityof Pennsylvania Universityof Pennsylvania Universityof Edinburgh

[email protected] [email protected] [email protected]

ABSTRACT.

We presentan implementationof a discourseparsingsystemfor a lexicalized Tree-AjoiningGrammarfor discourse,specifying the integration of sentenceand discourselevel process-ing. Our systemis basedon the assumptionthat the compositionalaspectsof semanticsatthe discourse-level parallelthoseat thesentence-level. This couplingis achieved by factoringaway inferentialsemanticsandanaphoricfeaturesof discourseconnectives. Computationally,this parallelismis achieved becauseboththesentenceanddiscoursegrammarareLTAG-basedandthesameparserworksatbothlevels.Theapproachto aLTAG for discoursehasbeendevel-opedby (Webber& Joshi1998;Webberetal. 1999b)(amongothers)in somerecentpapers.Oursystemtakesadiscourseasinput,parsesthesentencesindividually, extractsthebasicdiscourseunits from thesentencederivations,andreparsesthe discoursewith referenceto the discoursegrammar.

1 Intr oduction

All work ondiscoursestartsfrom thepremisethatdiscoursemeaningis morethanthesumof its parts(i.e., its constituentsentencesor clauses).Thequestionis howto get there. Work in the tradition of RhetoricalStructure Theory(RST) (Mann& Thompson1988)– bothin interpretation(Marcu2000)andgeneration(Mellishet al. 1998)– views theadditionalmeaningsolely in termsof discourserelationsthathold betweenadjacenttext spans,treatingdiscourseconnectivesassignallingtypesof discourserelations.How thebasictext spansareassignedan interpreta-tion, andhow thatinterpretationmightcontribute to discoursemeaningapartfromdiscourserelations,is largely ignored.

Not so in more formal work on discoursesemantics(Gardent1997; Polanyi& vandenBerg 1996;Scha& Polanyi 1988;Schilder1997;vandenBerg 1996),which takesseriouslyacompositionalprocessof deriving discoursemeaningfrom

17


themeaningof its parts.However, thiswork (1) onlymakesuseof twomechanismsfor deriving discoursemeaningfrom themeaningof its parts– compositionalse-manticsandinference– and(2) treatstheprocessby whichdiscoursederivescom-positionalaspectsof meaningasbeingentirely separatefrom how clausesdo so.Both of thesearethe focusof theapproachdevelopedin (Webber& Joshi1998;Webberetal. 1999b).In thisapproach,it is arguedthatcertainaspectsof discoursemeaningarebetterseenasderiving from anaphoricandpresuppositionalpropertiesof lexical items,andthatthis is facilitatedthrougha uniform lexicalisedtreatmentof both clausalsyntaxandsemanticsanddiscoursesyntaxandsemantics.Thispaperpresentsan initial implementationof a discourseparsingsystem(D-LTAG)thatdraws on theinsightsof this latterapproach.

Our motivation for using this approachis to explore the hypothesisthat theboundarybetweensentencelevel structureanddiscourselevel structureis not asharpone.Sentencelevel structuresupportscompositionalsemanticseventhoughthereareothersemanticaspects,suchasanaphoricrelations(e.g., intrasententiallinks for pronounreference)and inferential interpretation(e.g., interpretationofcompoundnouns)thatneedto beaccountedfor. In thesameway, discourselevelstructureis alsoviewed assupportingcompositionalaspectsof semantics,whileallowing for otherinterpretive componentsto beaddedon for a completeseman-tics for discourse– e.g.,for determininganaphoricandinferential interpretation.Thus,we pursuethe idea that the formal device usedfor deriving the structuraldescriptionsat both levels is the same,while noting that at the discourselevel,thedevice mayhave lessgenerative power. In addition,we alsoillustratethat thedescribedarchitecturefor discourseparsingallows for a smoothtransitionfromsentencelevel to discourselevel processingandfor the useof a singleparseratbothlevels.

In Section2, we discussthe LTAG framework for discoursedescription,asoutlinedin (Webber& Joshi1998).Section3 presentsadiscussionof ourmethod-ology for determiningthe structureand semanticsof discourseconnectives, ac-companiedwith a casestudyof the discourseconnective however. In Section4,we describethe architectureof our system,anddiscussvariousissuesthat aroseduring the implementation. Section5 discussessomeof the advantagesof oursystem,in particular, with respectto the closelink betweensentencelevel anddiscourselevel semantics.In Section6, we compareour systemwith someotherapproaches,in particularwith thosethat usesomevariantof TAG for describingdiscoursestructure,suchas (Gardent1997)and (Schilder1997),and thosethatattemptto automatethederivationof discoursestructure,suchas(Marcu2000)s.

2 The Framework: A Lexicalized TreeAdjoining Gram-mar for Discourse

TheD-LTAG systemis basedon theapproachto a lexicalizedTAG for discourse,asdescribedin (Webber& Joshi1998). A LTAG for discoursepositstwo kinds

18

of elementarytrees:initial trees,which encodepredicate-argumentdependencies,andauxiliary trees,which arerecursive andmodify and/orelaborateelementarytrees.All structuralcompositionis achievedwith two operations,substitutionandadjunction. Clausesconnectedby a subordinatingconjunctionform aninitial treewhosecompositionalsemanticsis determinedby thesemanticrequirementsof thesubordinateconjunction(the predicate)on its arguments(the clauses).Auxiliarytreesareusedfor providing further informationthroughadjunction.They canbeanchoredby adverbials,by conjunctionslike and, or mayhave no lexical realiza-tion. Furthermore,a discoursepredicatemay take all its argumentsstructurally,as in the caseof subordinatingconjunctions,or anaphorically, by makinguseofeventsor situationsavailablefrom thepreviousdiscourse,asin thecaseof then.1

This division betweenthecompositionalpartof discoursemeaning(projectedbythe treestructures)andthenon-compositionalcontributionsdueto generalinfer-encingandanaphorais a key insightof theapproachto anLTAG for discourse.Itsimplifiesthestructureof discourseandextendscompositionalsemanticrepresen-tationsfrom thesentencelevel to thediscourse.

Figure (7.1a)shows one initial tree in the grammar.2 We treat connectivesanchoringthis treeasdiscoursepredicateswhich requiretwo clausalarguments.In general,suchtreesareanchoredby subordinatingconjunctionssuchasbecause,whenetc.A corollaryof thestructureof elementarytreesin thediscoursegrammaris that discourseconnectives are allowed discourseinitially only if they anchoran initial tree. A secondinitial tree is shown in Figure (7.1b). As suggestedinprevious work (Webber& Joshi1998),this treereflectsdependenciesin parallelconstructionsandis projectedby pairsof connectivessuchason theonehand...on the other hand.... (As notedin this previous work, both membersof the pairneednotberealizedin thesurfacestring.)

Sr

S1↓ Conn◊ S2↓

Sr

Conn1◊ S1↓ Conn2◊ S2↓

(a) (b)

Figure7.1: Initial Treesin D-LTAG

As in thesentence-level grammar, treesin theD-LTAG grammararegroupedinto treefamilies, whicharetakento reflectsurfaceclauseordervariationrealizedwith preposedand postposedsubordinateclauses.Treesbelongingto the same

1Our useof the term anaphora doesnot includeanaphoricrelationssuchas thoseestablishedby pronounsanddefinitedescriptions.Accountsof theserelationshave beenactively pursuedinotherdiscourse-orientedsemantictheoriessuchasDRT (Kamp1981)andDynamicsemantics(Groe-nendijk& Stokhof1991).Obviously, afull accountof thephenomenonof anaphorain discoursewillhave to take theseinto account.But they arenot ourpresentconcern.

2In all the elementarytreesshown in the paper, “ � ” marksthe anchorof the tree, “ ” marksthesubstitutionnodes,and“ � ” markstheadjunctionnodes.Subscriptsareusedto distinguishnon-terminalnodeswith thesamelabel.

19

family sharethesamepredicate-argumentdependencies.Onesuchtreefamily isshown in Figure7.2,anchoredby connectiveslike because.

Sr

S1↓ Conn◊ S2↓

Sr

Conn◊ S1↓ S2↓

(a) (b)

Figure7.2: TreeFamily in D-LTAG

Thesecondtypeof elementarytreesconsistof auxiliary trees, which introducerecursionandserve to extendor modify a descriptionof the previous discourse.Therearetwo kindsof auxiliary trees,shown in Figure7.3.

Sr

Sf* Conn◊ S1↓

Sr

Conn◊ Sf*

(a) (b)

Figure7.3: Auxiliary Treesin D-LTAG

The treein Figure7.3(a)is anchoredby connectivesthat simply continuethedescriptionconveyed by thestructureto which the treeadjoins. Otheraspectsofmeaningthat relatethe two argumentsarederived anaphoricallyor inferentially(e.g., basedon the relationshipbetweenthe tense/aspectof the two arguments(Hitzemanet al. 1995; Kehler 1994; Kehler 2000; r Lascarides& Asher1993;Webber1988).Theanchorof this treecanalsoremainlexically unrealized,whenit is usedto connectadjacentclauseswithout overt connectives, suchas “Marywalkedtowardsthecar. Thedoor wasopen”. Thetreein Figure7.3(b)is selectedby connectiveswhosefirst argumentis resolved anaphoricallyandthesecondar-gumentis theinterpretationof theclausethey adjointo. Wesaymoreaboutthis inthenext section.

3 Determining TreeStructuresfor DiscourseConnectives

In theprevious section,we definedtheelementarytreesincludedin theD-LTAGgrammar. The next crucial stepin lexicalizing a Tree-AdjoiningGrammaris de-terminingwhich treesor family of treesareselectedby a discourseconnective. Inprevious work (Webberet al. 1999b,Webberet al. 1999c),it wasshown that theconnectivesthen, for exampleandotherwisearebesttreatedasanaphoric,anchor-ing treesof type7.3(b). But for someotherconnectives,suchashowever, it waslessclearwhetherthey arestructuralor anaphoric.

In whatfollows,wetake theview thatthelexicalizationof treesis anempiricalquestionandwe describethe methodologywe adoptto determinethe structures

20

lexicalizedby connectives.Whenin doubtaboutthestructureof acertainconnec-tive,we startwith thehypothesisthattheargumentsof theconnective arerealizedstructurally. This is because,from a computationalpoint of view, it would belessinterestingto startwith theassumptionthatargumentsareresolvedanaphorically.Assumingthatall connectivesfind their argumentsanaphoricallywould probablybe adequateto characterizeall predicate-argumentrelationshipson the discourselevel. However, it wouldnot shedmuchlight on thoseaspectsof structuralorgani-zationthatarerelevantto languagestructureandpresumablycontributeto theeffi-ciency of theinferencingprocessesrequiredin the interpretationof thediscourse.(This motivation is inspiredby (Joshi& Kuhn1979)).Predicateswhich find theirargumentsstructurallydefinea domainof locality for structuraldependenciesandconstrainthe interpretationof discoursein a computationallyefficient way, as isthecasefor verbpredicatesat thesentencelevel syntax.

On empiricalgrounds,the diagnosticwe useto test if a connective is struc-tural is crossedstructuraldependencies.ThecurrentXTAG grammarfor Englishdoesnot leadto crosseddependenciesasthey seemto beunnecessaryat thesen-tencelevel for English.We make a similar assumptionfor thediscourselevel andconcludethata connective definesa domainof structurallocality only whensuchdomainsdonot crosstreenodes.3

A CaseStudy: However For theconnective however, ourfirst assumptionis thathoweveranchorsthestructuralauxiliarytree,shown in Figure7.3(a).Regardingitssemanticcontribution, we follow (Knott 1996)and(Lagerwerf1998)in assumingthat however presupposesa defeasiblerule holding betweena generalizationofits first argumentanda generalizationof thenegationof its secondargument,andassertsthat the rule fails to hold in this case(see(Webberet al. 1999a)for aformalizationof therule). To investigateif bothargumentsarerealizedstructurally,weconductedacorpusstudyof theconnective. Weidentified71tokensof howeverfrom theBrown corpusandlocatedthe two argumentsof theconnective for eachtoken. In 69outof the71instances,thepositionof bothargumentsin thediscoursewasconsistentwith the structuralhypothesis:oneargumentwas realizedin thesentencecontainingtheconnective,andthesecondargumentwasrealizedeitherintheimmediatelyprecedingsentence(58 instances)or in animmediatelyprecedingchunkof text (11 instances).In both cases,attachmentto the previous discoursedid not yield crossingof treenodes.

Theremaining2 caseswereof two kinds.One,exemplifiedin (1), involvedanargumentthatwasnot directly realizedin thepreviousdiscourse.Rather, thepre-supposeddefeasiblerulecouldonly beseenasholdingbetweenrathercomplicated

3However, until we have accumulatedampleempiricalevidence,suchconclusionsaretentativeand subjectto revision. Also, it would be very interestingto investigatelanguageswhich allowcrosseddependenciesat the sentencelevel (e.g. Dutch) andexaminewhetherin thoselanguagescrosseddependenciesarealsopermissibleon thediscourselevel. Our conjectureis thatthis will notbethecase.

21

generalizationswhichwouldhave to be inferred from thetwo arguments.Herewetake the defeasiblerule to be somethinglike “If the speaker/writer makesan ap-parentlynegative commentabouta book,thenhis/herattitudeis negative towardsit.”

(1) a. If this new Bible doesnot increasein significanceby repeatedreadingsthroughouttheyears,it will not survive theagesashastheKing JamesVersion.

b. However, an initial perusalandcomparisonof someof the famouspassageswith thesamepartsof otherversionsseemsto speakwell of theefforts of theBritish Biblicalscholars.

In the othercase,however appearedto make no semanticcontribution to thediscourse,otherthansimplecontinuation.This is shown in (2).

(2) a. It is in this spirit which explainssomeof theanomaliesof AmericanCatholichighereducation,in particularthewastefulduplicationapparentin someareas.

b. I think for exampleof threewomen’scollegeswith pitiful enrollments,clusteredwithina few milesof a majorCatholicuniversity, which is alsoco-educational.

c. This is notanisolatedexample;

d. this aspectof thetotal picturehasbeencommenteduponoftenenough.

e. It wouldseemto representespritdecorpsrun riot.

f. Apart, however, from the questionof wastefulduplication,thereis anotheraspectofthe“f amily business”spirit of Catholichighereducationthatdeservescloserscrutiny.

While it is clearthat (2f) attacheshigherup to the structurecontaining(2b)-(2e), it is lessclear what the semanticcontribution of however is to the inter-pretationof the discourse.However hereseemsto be actingsimilar to the dis-coursemarkernow(e.g.,“Now, apartfrom thequestionof wastefulduplication...”)(Hirschberg & Litman 1987),reinforcingthe IRU cue(i.e., “apart from theques-tion of wastefulduplication”)asa signalof returningto (2a)aftera conceptuallyembeddedsegmentwasclosedoff at (2e).4

To summarize,the corpus-basedstudy for the connective however providesconsiderablesupport for the hypothesisthat it finds its argumentsstructurally.However, asindicatedby themorecomplex examples(1) and(2), furtherempiricalstudieswill berequiredbeforea definitive conclusionis reached.

4 SystemDescription and Implementation

In thissection,wedescribeourinitial implementationof adiscourseparsingsystembasedonalexicalizedTree-AdjoiningGrammarfor discourse.Discoursestructure

4InformationallyRedundantUtterances(IRUs) arecharacterizedasrepetitionsof propositionsalreadyavailable in the discourse. (Grosz& Sidner1986) have shown that IRUs correspondtoembeddedsegments.(Walker 1993)arguesthat,with respectto a well definedtask,IRUs areusedasa discoursestrategy to improve the efficiency of completinga task. Thedistribution of IRUs inWalker’scorpusindicatesthatIRUs functionasmarkersof returningto asuperiorsegment.Seealso(r Forbes& Miltsakaki2001)for a discussionon thecollaborationof IRUs with othercuesderivedfrom CenteringTheoryto signaltheboundariesof embeddedsegments.

22

is derivedin two passesof parsing.In thefirst pass,thesentencesin thediscourseareparsed,whereasdiscourseparsingis donein thesecondpass.Without losingsight of the key ideasof the theoryof an LTAG for discourse,this two passim-plementationachievesa considerablesimplificationover a singlepassof parsing,especiallyin termsof the parsingtime andspacerequirementsthat would resultfrom usingboththesentence-level andthediscourse-level grammaratonce.

Figure7.4: D-LTAG: SystemDescription

Tree

Extractor

TreeMapper

Discourse Input

Generation

and

Tree Selection

Clausal

Derivations

Tree Database

Clausal Tree

LEM

Derivation Structure

for

Discourse

Representations

LEM

Input Discourse

(Sentence Parsing) (Discourse Parsing)

Figure7.4 shows the overall architectureof the system.The input discourseis submittedto LEM, theParser, whichparseseachsentencein thediscoursewithreferenceto thesentencegrammar. Theoutputderivations(onederivationeachforeachof the sentences)arethensubmittedto the TreeExtractor , which extractsthebasicdiscourseconstituentunitsfrom eachsentencederivation. Thebasicdis-courseunitsconstitutetheelementarytreeslexicalizedby discourseconnectivesinthesentence-level grammar, andthederivation(andderived)structuresassociatedwith the clausalunits.5 In the next step,the sentence-level elementarytreesan-choredby theconnectivesaremappedby theTreeMapper to their correspondingelementarytreesin the discoursegrammar. The discoursegrammar, asspecifiedin Section2, is containedin theTreeDatabase. Theoutputof theTreeMapper,togetherwith the clausalunits andthe input discourse,is thenusedto constructa discourseinput representationthat consistsof a sequenceof lexicalizedtrees(treeselection),with theextractedconnectivesandclausalunitsasthelexicalizingelements.6 Finally, thediscourseinput, theTreeDatabase,andtheclausaltreerep-resentationsaresubmittedto thesameParser, which providesderivationsfor thegivendiscourse.

5In this paper, we assumethatclausalunitscorrespondto theminimal tensedclause.Thetensedclauseis furthertakento includeall sententialcomplements,relative clausesandparticipialclauses.In someotherdiscourseworks,suchas(Polanyi 1996),agreaterrangeof propositionalelementsareregardedastheminimal unitsof discourse.

6Eachextractedclausederivation is taken to beanatomicunit in thediscoursegrammar, muchlike a singlelexical item.

23

In therestof this section,we describethedifferentcomponentsof thesystemin greaterdetail,anddiscussvariousissuesthataroseduringtheimplementation.PARSER (LEM).The parseris a chart-basedhead-cornerparser(Sarkar2000).The sentence-level grammarusedby the parseris the XTAG grammar(XTAG-Group2001),a wide-coveragegrammarof Englishdevelopedat theUniversityofPennsylvania.7 For eachsentence,eachsubsequentphaseof thesystemassumesthat thereis exactly onederivation per sentence.Since,in general,therecanbemany ambiguitiesfor eachsentencein thediscourse,theparserpicksonederivationpersentenceto passon to subsequentprocessing.In thesystemdescribedin thispaper, theparserproducesa singleparsefor eachsentenceby usingheuristicsthat(a) decidewhich elementarytreeto assignto eachword, and(b) pick the lowestattachmentbetweenthesetrees. In future work, we plan to experimentwith twoothermethodsto dealwith ambiguity: (1) usingthe parserasa statisticalparser(2001) whereit reportsthemostprobableparsebasedontrainingtheparseronthePennTreebank,and(2) representingthemany parsesassociatedwith eachclausalunit in thesentencein acompactform (aparseforest)andrepresentingtheseastheelementaryunitsin thediscourse.TREE EXTRACTOR. The taskof this componentis to extract, from eachsen-tencederivation, the clausalderivationsandany elementarytree(s)anchoredbydiscourseconnectives.TheExtractorfirst doesatop-down traversalof thesentencederivation,andidentifiesthepartof thederivationassociatedwith any connectives.Identificationof theconnectivesis doneagainstadatabasecontaininga list of pos-siblediscourseconnectivesaswell astheelementarytree(s)anchoredby eachofthemin thesentencegrammar.

The useof both lexical and structuralinformation is necessaryto correctlyidentify thediscourseusagesof connectivesin thesententialderivations. That is,neitherkind of information by itself is sufficient for identification. On the onehand,many elementsthat function asdiscourseconnectives canalsohave otherfunctions:and functionsasa discourseconnective whenit conjoinsclauses,asin“The dogbarkedandMary smiled”, but it canalsoconjoinnounphrases(amongotherphrasalcategories),asin “Lana atecheeseandcrackers”. As a result,if weusedonly thelexical appearanceof theelementsastheidentificationcriterion,thentheandwhichconjoinsnon-sententialcategorieswould beincorrectlytreatedasadiscourseconnective. Knowledgeaboutthe elementarytreesassociatedwith thediscourseusageof and is thereforenecessaryto rule thisout. On theotherhand,itis notsufficient to only usestructuralinformationto identify discourseconnectives.For example,the sentence-level grammardoesnot make a structuraldistinctionbetweensententialadverbsthat are discourseconnectivesandthosethat arenot:the elementarytreein Figure7.5 canbe lexicalizedboth by however, which is aconnective, aswell asby always, which is not a connective. Identificationof theformer- andnot thelatter- structurecanthereforebedoneonly if thelexicalizingelementsarealsoused.

7For a recentevaluationof theXTAG grammar, see(Prasad& Sarkar2000).

24

Sr

Adv◊ Sf*

Figure7.5: Elementarytreeanchoredby anadverbin thesentence-level grammar

After theidentificationof theconnectives,theclausalderivationsaredetachedin the sentencederivation at the substitutionand/oradjunctionnodesof the con-nective elementarytree.Theresultof thisprocedureis shown in Figure7.6for thederivationof thesentencewhileshewaseatinglunch, shesawa dog.8

saw

she dog while

she was lunch

a eating

,

Extractor

saw

she dog

a

(i)

(iii) while

eating

she was lunch

(ii)

,

Figure7.6: Resultof TreeExtractorappliedto derivationof while shewaseatinglunch, shesawa dog.

Theaboveprocedurecanbeshown to work onall derivationsin whichconnec-tivestake clausesastheir argumentsin thesentence-level grammar. In thesurfacestring, this correspondsto connectivesappearingat clauseboundaries.However,connectivescanalsoappearin clause-medialpositions,asin Example3. Thecon-nective thenadjoinsto theverbphrase(VP node)in theclause.

(3) Susanwill thentake dancinglessons.

Thoughsuchclause-medialconnectivesarepositedastaking clausesastheirargumentsin thediscourse-level grammar, webelievethattheirclause-internalsyn-taxshouldbevisibleatthediscourse-level description,asit is probablyanindicatorof InformationStructure(IS).9 TheExtractorachievesthis by only makinga copyof the derivationsfor theseconnectives, and by replacing- in the derivation ofthe clause- the lexical occurrenceof the connective by an index, to indicateitsclause-internalposition. The resultof this procedurefor example(3) is given inFigure7.7. � then� in 7.7(i) representstheclause-medialconnective index left bytheExtractor.

8In derivation structures,dottedlines indicatesubstitutionand solid lines indicateadjunction.Also, note that eachnode is labeledby the lexical items, but theseonly serve as labels for theelementarytreewith which they areassociated.

9The hypothesiswe arepursuingis that a clause-medialconnective flagsmaterialto its left asbeingacontrastive theme(Steedman2000a)– cf. Section5.

25

dancingwill

then

take

lessonsSusanExtractor

(i)take

lessonsSusan

dancingwill

{then}

(ii) then

Figure7.7:Resultof Extractorappliedtoderivationof Susanwill thentakedancinglessons

Thus,theoutputof theExtractoris, for eachsentence,asetof elementarytreesanchoredby connectivesin thesentencegrammar, andasetof clausalderivations.For theexamplediscoursegiven in (4), (5) shows thepartsof thediscourseinputcorrespondingto theextracteddiscourseconstituentunits. (b), (f), (i) and(k) aretheextractedconnectives,andtherestaretheclausalderivations.

(4) a. Mary wasamazed.

b. While shewaseatinglunch,shesaw adog.

c. She’d seena lot of dogs,but thisdogwasamazing.

d. ThedogbarkedandMary smiled.

e. Thenshegave it a sandwich.

(5) a. marywasamazed

b. while

c. shewaseatinglunch

d. shesaw a dog

e. she’d seena lot of dogs

f. but

g. this dogwasamazing

h. thedogbarked

i. and

j. marysmiled

k. then

l. shethengave it a sandwich

TREE MAPPER. Theconnective-lexicalized elementarytreesthatareextractedfrom thesentencederivationsaresubmittedto theTreemapper, which mapstheirsentence-level structuraldescriptionsto theirdiscourse-level structuraldescriptions(taken from theTreeDatabase).This is a crucial stepin thediscoursederivationbecauseit is involvedwith determiningwhatkindsof contribution(s)a givencon-nective makes to discoursemeaning,that is, what it contributesthroughcompo-sitional semantics,throughanaphoraandthroughinference.Furthermore,ashasbeenpointedout in Section3, determiningthe discoursestructuresanchoredbyconnectivesis anempiricalmatter. A majorpartof thefuturework in this project

26

is to fully determinethismappingwith corpusbasedwork on thebehavior of con-nectives.Wecontinuehereby assumingthemappingsshown in Figure7.8 for theexamplediscourse(4).

rSS

Conn S1 S2

while

rSS

S f *Sc

P S1

while

(a)

rSS

S f * Conj S1

but

rSS

S f * S1

but

Conn(b)

rSS

S f * S1Conn

and

rSS

S f * Conj S1

and

(c)

then

VPr

VPf *AdvS f *Conn

rSS

then

(d)

Figure7.8: Elementarytreemappingsfor connectivesin discourse(4)

DISCOURSEINPUT GENERATION. In thenext phaseof thesystem,theclausalderivationsarefirst convertedintoelementarytreerepresentations,whicharetreatedassingularatomicunits that canserve astheargumentsof thediscourseconnec-tives.Theseclausalunits,theinputdiscourse,andtheconnective elementarytreesgeneratedby the TreeMapperarethenusedto generatea discourseinput repre-sentationthat is essentiallya sequenceof lexicalizedtrees,wherethe lexicalizingelementsaretheconnectivesandtheclausalunits.Becauseof theextractionof thediscourseunitsfromthesentencederivations,andthetreemappingof thestructuresof connectives,treeselectionambiguitiesat thediscourselevel areminimized,anddiscourseparsingthussimplified.

Thesequenceof lexicalizedtreesis orderedwith referenceto thesurfaceorderof the input discourse(compare(4) and (5)), except for the clause-medialcon-nectives. Theseareplacedbeforethe clausefrom within which they arecopiedout. This doesnot, however, disruptthe surfacestring order: the clause-internalindex of theseconnectives, left by theTreeExtractor, succeedsin preservingthesententialsurfacestringrepresentation(seeFigure(5i)).

This phasealsoincludesan insertionalgorithmto insert treeswith an emptylexicalanchor(whichmaystill carrysomefeatureinformation)into theinputrepre-sentation.Recallfrom Section2 thatthegrammarcontainsanauxiliary treethatisusedto continuethedescriptionby adjoiningto thepreviousdiscourse(henceforth,continuationauxiliary trees)(Figure7.3a).Thisauxiliary treemaybeanchoredbyconnectiveslike andandor, or remainlexically unrealized.In theextractedunitsshown in (5), thereareonly 2 overt connectivesthatcananchorthisauxiliary tree:andandbut. This meansthat the lexically emptytreesneedto be insertedat theappropriatepositionsin theinput representation.Theinsertionalgorithmdoesthisby referringto the treelabelsfor eachof the units in the (thusfar created)inputrepresentationandby following a few simpleinsertionrules.Weusethelabel“E”to indicatea null anchor. Alternatively, thesetreescanbe taken to be lexicalizedby thesentence-finalpunctuationmarkers.

27

she was eating lunch

a dogshe saw

she’d seena lot of dogs

this dog was

amazing

the dog barked

Mary smiled

amazedmary was

she gave it a

sandwich

E

while

but

E

E

and

E

then

Sf

Sr

amazedmary was

Sr

Sf

S1 S2

she was eating lunch

a dogshe saw

Sr

Sf

Sf

Sr

this dog was

amazing

Sr

Sf Sr

Srthe dog barked

fS

Mary smiled

Sr

Sf

she gave it a

sandwich

she’d seena lot of dogs

Conn

E

while

Conn

E Conn

but Conn

E Conn

and

Conn

E then

(a) Derived Tree

(b) Derivation Structure

Figure7.9: DerivedTreeandDerivationStructurefor ExampleDiscoursein (4)

Thesequenceof lexicalizedtreesafter insertionof thelexically emptytreesisthenparsedwith thesameparser(LEM) thatwe usedto parseeachsentencein thediscourse.Sincethetreesareuniquelyselectedby theconnectivesandtheclausalunits,theresultingdiscourseparsecontainsno ambiguitiesthatarecausedby treeselection.10 However, thesystemdoescontainattachmentambiguitiescausedbythe continuationauxiliary trees. In the currentapproach,thesemay be resolvedwith an inferential component,or by statisticalmethods. For presentpurposes,we pick a uniquederivationout of all theparseswhich satisfiesthefollowing twocriteria: (a) adjunctionin initial treesis only allowed at the root node; and (b)for all otherpermissibleadjunctions,only lowestadjunctionis allowed. Giventhesimplegrammarpositedin the system,thesetwo criteria aresufficient to yield auniquederivation.

Thederivedtreeandderivationstructurefor theexamplediscoursein (4) afterdiscourseparsingareshown in Figure7.9.

We have alsotestedour systemon connective rich sectionsof theWall StreetJournal(WSJ) from the PennTreebank(Marcuset al. 1993). In order to avoidthe problemof gettingtoo many sententialderivationsfor the long andcomplexsentencestypically found in this corpus,we usedthesinglederivationsproducedby LEXTRACT (Xia et al. 2000),which takestheTreebankandTreebank-specificinformationandproducesderivation treesfor thesentencesannotatedin theTree-

10This resultobtainsbecausethediscoursegrammarassumedhereis quitesimple,with discourseconnectivesprojectingasingleelementarytree.Wenotethatuponfurtherempiricalinvestigationofthebehavior of individual connectives,this maynot turnout to bethecase.

28

bank.For theWSJdiscoursesegment(taken from Section21 of theWSJcorpus)given in Example6, the derived tree andderivation structureareshown in Fig-ure7.10.Thediscourseconnectivesin thetext areshown in bold.

(6) a. The pilots could play hardballby noting they arecrucial to any saleor restructuringbecausethey canrefuseto fly theairplanes.11

b. If they were to insist on a low bid of, say $200a share,the boardmighn’t be ableto obtaina higheroffer from otherbiddersbecausebanksmight hesitateto financeatransactionthepilots oppose.

c. Also, becauseUAL chairmanStephenWolf andotherUAL executiveshave joinedthepilots’ bid, theboardmightbeableto excludehim from its deliberationsin orderto befair to otherbidders.

Sr

they were to insiston a low bid of, say$200 a share

S1

Sf

the board mighn’t be able to obtain a higher offer fromother bidders

banks might hesitateto finance a transac-tion the pilots oppose

S2

Sr

Sf

S1

UAL chairman StephenWolf and other UAL executives have joined the pilots’ bid

S2

Sf

The pilots could playhardball by noting they were crucial to any sale or restruc-turing because theycan refuse to fy the airplanes.

SConn

E

Conn

if

Conn

because

Conn

E

r

Conn

Conn

because

Also

the board might be able to exclude him from its deliberations in order to be fair toother bidders

The pilots could playhardball by noting they were crucial to any sale or restruc-turing because theycan refuse to fly theairplanes.

they were to insiston a low bid of, say$200 a share

the board mighn’t be able to obtain a higher offer fromother bidders

banks might hesitateto finance a transac-tion the pilots oppose

if

because

UAL chairman StephenWolf and other UAL executives have joined the pilots’ bid

the board might be able to exclude him from its deliberations in order to be fair toother bidders

E

because

Also

E

(b) Derivation Structure

(a) Derived Tree

Figure7.10:DerivedTreeandDerivationStructurefor WSJDiscoursein (6)

5 Discussion

5.1 Deriving DiscourseSemantics

In (Joshi& Vijay-Shanker 1999)and(Kallmeyer & Joshi2001),an approachtocompositionalsemanticswasprovidedfor theLTAG grammar. Thecompositional

11Note that “because”in this sentencehasnot beentreatedasa connective because,initially, wehave ignoredsententiallyembeddedconnectives. How they contribute to discoursestructureandmeaningremainsanimportanttopic for futurework. Seealsofn 5.

29

semanticswasdefinedwith respectto thederivation treestructureandnot thede-rived tree. Thederivation treeis a recordof thehistoryof compositionof theel-ementarytrees.Semanticrepresentationsareassociatedwith theelementarytreesandthesearecomposedvia unification. In D-LTAG, the intuition is thata similarprocesswill becarriedoutat thediscoursestructurelevel usingthederivationtreeof theD-LTAG grammar. Thedetailsof this compositionalsemanticsfor D-LTAGhavenotbeenworkedoutyet. However, in general,thefinal representationwill beessentiallya flat structure,asis thecasefor thesemanticsof LTAG.

While eachelementaryLTAGtreeis associatedwith asemanticrepresentation,this representationdoesnot have to reproducethehierarchicalstructureof theel-ementarytree. Theelementarytreeis thusconsideredasa “semanticunit”. Thisview of representingsemanticsdirectly correspondsto theso-calledflat represen-tation, which is a conjunctionof formulas,wherehierarchicalstructureappearsonly whenneeded,for example,for operatorson predicates.Sucha “flat” repre-sentationis alsomotivatedin thecontext of generation,whereonewantsto startwith a representationof theinput which makestheminimal commitmentto struc-ture. Detailson associatinga flat semanticswith a derived TAG treecanalsobefound in (1997) and(Stoneet al. 2001). In (7) below, we roughly illustratethesemanticformula associatedwith the examplediscoursein (4). We refer to thesemanticsof theconnective treesby thenamesof theconnectives,anduse“&” torepresentthe semanticsassociatedwith theauxiliary treeassociatedwith contin-uation/elaboration.Thenumberedargumentsof thesetreesarelabeledeither“S”for statesor “E” for events,dependingon their semanticcontent.

(7) S1& while(S2,E3)& but(S4,S5)& (E6 & E7) & after(E7,E8)

5.2 DiscourseConnectives,Inf ormation Structure, and DiscourseSe-mantics

While the currentstudy doesnot directly addressinteractionsbetweeninforma-tion structure(IS) and discoursestructure/semantics,we note that a lexicalisedgrammar for both sentencesand discourseallows semanticmaterial from boththe lexicon andconstructedphrasesto project into both sentence-level meaninganddiscoursemeaning.In particular, this allows sentence-level IS distinctionstobe projectedinto discourse-level descriptions. We illustrate this by consideringclause-medialadverbialdiscourseconnectives.

Many adverbial connectivesdisplaya wide variability with respectto thepo-sition they arefound in the sentence.This variability, while usuallynot alteringthecompositionalsemanticsof thesentence,appearsto correlatewith IS. In par-ticular, we arepursuingthe hypothesisthat a clause-medialconnective indicatesthatmaterialto its left servesasa contrastivetheme. Thesimplestcaseis giveninExamples8 and9.

(8) Mary smiled.However, Johnfrowned.

(9) Mary smiled.John,however, frowned.

30

In (8), the clause-initialposition of however is, by itself, neutralabout thepartitionof thesentenceinto themeor rhemeandaboutwhetheror not the themeis contrastive. In (9), on the other hand,the clause-medialposition of howevercorrelateswith stresson “John” and appearsto convey that Johnand Mary areelementsof an alternative set (in the senseof (Rooth1992))– that is, that Johnservesasacontrastive theme.

This comesout morestronglyin Examples(10) and(11). (10) is infelicitousbecausemedialhowever flagsthesubjectascontrastive theme,but thissubjectis acoreferential(unstressed)pronounandcannotserve asacontrastive theme.Exam-ple (11), on theotherhand,is fine, asthepositionof however flagstheadverbialthenas the contrastive theme(presumablyin an alternative setwith the time ofMary smiling).

(10) � Mary smiled.Thenshe,however, frowned.

(11) Mary smiled.Then,however, shefrowned.

Our claim hereis just that, by having elementslexicalisedboth with respectto sentenceanddiscourse,we canrepresentin thesameway their contributionstoboth,aswell asinter-relationsbetweenthem. For example,in (11), not only doestheclause-medialpositionof however, flag thenasacontrastive theme(in contrastwith alternatives provided in the discourseor the speechsituation),but the de-feasiblerule presupposed(or conventionallyimplicated)by however (Knott 1996;Lagerwerf1998)involvesthatspecific“inertial” property– i.e., if someonesmiles,they will continueto doso.Howeverassertsthatit failsto hold,andwhathappenedthenis thesourceof thefailure. While we have not yet exploredthis with respectto LTAGandD-LTAG,(Bierner& Webber2000)and(Bierner2001)illustratehowanotherlexicalisedgrammar, CombinatoryCategorialGrammar(CCG)(Steedman2000b),canbeusedto expressbothassertionalandpresuppositionalcomponentsof meaningassociatedwith thesentenceandwith discourse,and(Steedman2000a)shows how onecancomputebothIS-partitioning,its prosodyandits semanticsinlockstepwith otheraspectsof meaning.

6 Comparisonwith RelatedApproaches

Recently, (Marcu2000)developeda systemfor identifying rhetoricalrelationsonunrestrictedtext. His systemtrains on a corpusannotatedwith rhetoricalrela-tions andutilizes correlationsof surface-basedfeatureswith RST relationsto as-sign rhetoricalstructureto unseentext. Our systemis a cleardeparturefrom thisapproachin two significantways:a) we developa systemthatactuallyparsesdis-courseallowing thesemanticsto bebuilt compositionallyfrom thesentenceto thediscourselevel, andb) discourseconnectivesarenotviewedasnamesof relations,insteadthe semanticsof the connectives form only a part of the compositional

31

derivationof discourserelations.12

(Gardent1997)usesa variantof Feature-basedTreeAdjoining Grammarstoconstuctthe structureof discourseand the semanticsderived from it. (Schilder1997) extendsGardent’s formalism to handleworld and contextual knowledge,proposinganon-monotonicreasoningsystemto achieve that.Despitethissimilar-ity of the above works with our approach,both systemsdiffer significantlyfromoursin thefollowing way. Gardent’s system(alsoSchilder’s) builds thesemanticsof discoursecompositionallybut only after the semanticsof the input segmentsandtherhetoricalrelationconnectingevery two segmentsis identified. However,it is not clearhow thesemanticsof theinput segmentsarecomputedsince,appar-ently, the sizeof the input segmentrangesfrom tensedclauses(’We weregoingto take Johnasa lawyer’), to complex sentences(’As we found out, eitherhe ison sick leave’) or even fragments(’Too honestfor his own good, in fact’).13 Inour approach,we do not assumepre-processingor segmentationof thetextual in-put. Theoutputfrom thesentencelevel parseris theinput to thediscourseparser,building up thesemanticscompositionallyfrom thesentencelevel to thediscourselevel. Likewise, rhetoricalrelationsarenot assumednor picked out from a pre-viously definedsetof relations. We are interestedin thoseaspectsof discourseinterpretationthatarederivedcompositionally, factoringaway non-compositionalsemanticcontributions,i.e. inferencingbasedon world-knowledgeandanaphoricpresuppositions.

7 Conclusions

Building on earlierwork, we have developedandimplementeda systemfor dis-courseparsingbasedon a lexicalizedTree-adjoiningGrammarfor discourse,inwhich the discourseconnectivesarethe predicates,andthe clausesarethe argu-mentsof theseconnectives. The systemtakesa discourseasits input, parsesthesentencesindependently, extracts“discourse”connectivesandclausalunits fromtheoutputderivationsof thesentences,andreparsesthediscourseinputby submit-ting fully lexicalizedtreesto thesameparser.

We have motivateda corpusstudyof discourseconnectives in order to fullydeterminethesemanticcontribution they make to discourse,andthus,to alsode-terminetheelementarytreetype(s)they lexicalize in thediscoursegrammar. Thegrammarthusdevelopedwill serveasacrucialcomponentof theimplementedsys-temwhichusesthis informationafterextractingtheconnectivesfrom thesentencederivations,in order to createlexicalizedelementarytreesat the discourse-level.

12In otherwords, in our view, the ’name’ of a rhetoricalrelationis ultimately derived from thecompositionalsemanticsof our system,andothernon-compositionalaspectsof discoursemeaning,i.e. the inferentialcomponent.Theuseof ’rhetoricalrelations’in discourseintepretationseemstoconflatethosetwo distinctaspectsof meaning,namelycompositionalandinferential.In oursystem,we teasethetwo apartandderive thecompositionalpart.

13Theexamplesarefrom (Gardent1997),pp.7.

32

Thesubmissionof thelexicalizedtreesastheinput for discourselevel parsingsim-plifiestheparsingprocessconsiderably, andthissimplificationis achievedbecausethesystemintegratessentence-level processingwith discourse-level processing.

Acknowledgements

We thankMiriam Eckert, AlexandraKinyon, Alistair Knott, BangaloreSrinivasandFeiXia for helpfuldiscussionsduringthedifferentstagesin thepreparationofthis paper. We alsothankthreeanonymousreviewersfor their comments,whichhave helpedin improving thecontentandpresentationof thispaper.

Bibliography

Bierner, Gann(2001).AlternativePhrasesandNatural Language InformationRetrieval. In Proc.of the39th ACL. Toulouse,France.

Bierner, Gann& BonnieWebber(2000).InferencethroughAlternativeSetSe-mantics. Journalof LanguageandComputation,1(2):259–274.

Forbes,Katherine& Eleni Miltsakaki (to appear).AutomatedIdentificationofEmbeddedStructure in DiscourseSegmentation. In PennWorkingPapersinLinguistics,2001.

Gardent,Claire(1997).DiscourseTreeAdjoiningGrammars. ClausReport89,Saarbrucken: Univeristat desSaarlandes.

Groenendijk,Jeroen& Martin Stokhof(1991).DynamicPredicateLogic. Lin-guisticsandPhilosophy, 14:39–100.

Grosz,BarbaraJ. & CandaceL. Sidner(1986).Attention,Intentions,and theStructure of Discourse. ComputationalLinguistics,12(3):175–204.

Hirschberg, Julia& Diane.J.Litman (1987).NowLet’s Talk AboutNow: Iden-tifying CuePhrasesIntonationally. In Proc.of the25thACL, pp.163–171.

Hitzeman, Janet, Marc Moens & Claire Grover (1995). Algorithms forAnalysingtheTemporal Structure of Discourse. In Proc.of EACL, pp.253–260.Dublin, Ireland.

Joshi,Aravind & Steve Kuhn(1979).CenteredLogic: TheRoleof Entity Cen-teredSentenceRepresentationin Natural Language Inferencing. In Proc.ofthe6th IJCAI, pp.435–439.

Joshi,Aravind & K. Vijay-Shanker (1999).CompositionalSemanticswith Lex-icalizedTree-AdjoiningGrammar(LTAG): How much UnderspecicationisNecessary?In H. C. Blunt & E. G. C. Thijsse(Eds.),Proc.of the ThirdInternationalWorkshoponComputationalSemantics,Tilburg, Netherlands,pp.131–145.

Kallmeyer, Laura& Aravind Joshi(to appear).FactoringPredicateArgumentand ScopeSemantics:UnderspecifiedSemanticswith LTAG. In JournalofLanguageandComputation,2001.

33

Kamp,Hans(1981).A Theoryof TruthandSemanticRepresentation. In J.Groe-nendijk,Th. Janssen& M. Stokhof(Eds.),FormalMethodsin theStudyofLanguage,pp.277–322.Amsterdam:MathematischCentrumTracts.

Kehler, Andrew (1994).Temporal Relations: Referenceor Discourse Coher-ence. In Proc.of the32nd ACL, StudentSession,pp. 319–321.LasCrucesNM.

Kehler, Andrew (2000). ResolvingTemporal Relationsusing TenseMeaningandDiscourseInterpretation. In MartinaFaller, StefanKaufmann& MarcPauly (Eds.),FormalizingtheDynamicsof Information.CSLI publications.

Knott, Alistair (1996).A Data-DrivenMethodology for Motivatinga Setof Co-herenceRelations, (Ph.D.thesis).Edinburgh: Universityof Edinburgh.

Lagerwerf,Luuk (1998).CausalConnectiveshavePresuppositions. TheHague,TheNetherlands:Holland AcademicGraphics.PhD Thesis,CatholicUni-versityof Brabant.

Lascarides,Alex & Nicholas Asher (1993). Temporal Interpretation, Dis-course Relationsand CommonsenseEntailment. Linguistics and Philoso-phy, 16(5):437–493.

Mann,William C. & SandraA. Thompson(1988).RhetoricalStructure Theory.Toward a FunctionalTheoryof Text Organization. Text, 8(3):243–281.

Marcu,Daniel(2000).TheRhetoricalParsingof UnrestrictedTexts: A Surface-BasedApproach. ComputationalLinguistics,26(3):395–448.

Marcus,Mitchell P., BeatriceSantorini& Mary Ann Marcinkiewicz (1993).Building a Large AnnotatedCorpusof English: ThePennTreebank. Com-putationalLinguistics,19.

Mellish, Chris, Mick O’Donnell, JonOberlander& Alistair Knott (1998).AnArchitecture for OpportunisticText Generation. In Proc. of the 9th Intl.Workshopon NLG, pp.28–37.Ontario,CA.

Polanyi, Livia (1996).TheLinguisticStructure of Discourse. TechnicalReportCSLI-96-200:CSLI.

Polanyi, Livia & Martin H. vandenBerg (1996).DiscourseStructure andDis-courseInterpretation. In P. Dekker & M. Stokhof(Eds.),Proc.of the10thAmsterdamColloquium,pp.113–131.Universityof Amsterdam.

Prasad,Rashmi& Anoop Sarkar(2000). ComparingTest-SuiteBasedEval-uation and Corpus-BasedEvaluationof a Wide-Coverage GrammarForEnglish. In Using Evaluationwithin HLT Programs:Resultsand TrendsLREC’2000SatelliteWorkshop.Greece,pp.7–12.

Rooth,Mats (1992).A Theoryof FocusInterpretation. NaturalLanguageSe-mantics1, pp.75–116.

Sarkar, Anoop (2000).Practical Experimentsin Parsing usingTreeAdjoiningGrammars. In Proc.of TAG+5,France,May 25–27.

34

Sarkar, Anoop (2001).ApplyingCotraining Methodsto StatisticalParsing. InProc.of the2ndNAACL. Pittsburgh,PA.

XTAG-Group,The(2001).A LexicalizedTreeAdjoiningGrammarfor English.TechnicalReportIRCS01-03:Universityof Pennsylvania.

Scha,Remko & Livia Polanyi (1988).An AugmentedContext FreeGrammarfor Discourse. In Proc.of theCOLING’88, pp.573–577.Hungary.

Schilder, Frank(1997).TreeDiscourseGrammar, or How to GetAttachedto aDiscourse. In Proc.of theTilburg ConferenceonFormalSemantics,Nether-lands,January1997.

Steedman,Mark (2000a).InformationStructure and theSyntax-Phonology In-terface. Linguistic Inquiry, 34:649–689.

Steedman,Mark (2000b).TheSyntacticProcess. CambridgeMA: MIT Press.

Stone,Matthew & ChristineDoran(1997).SentencePlanningas DescriptionusingTreeAdjoiningGrammar. In Proc.of ACL, pp.198–205.

Stone, Matthew, Christine Doran, Bonnie Webber, Tonia Bleam & MarthaPalmer (2001).Microplanningfrom CommunicativeIntentions: SentencePlanningusingDescriptions(SPUD). Submittedto ComputationalIntelli-gence.

van denBerg, Martin H. (1996).DiscourseGrammarand DynamicLogic. InP. Dekker & M. Stokhof(Eds.),Proc.of the10thAmsterdamColloquium,pp.93–111.Universityof Amsterdam.

Walker, Marilyn A. (1993).InformationalRedundancyandResourceBoundsinDialogue, (Ph.D.thesis).Universityof Pennsylvania,CIS

Webber, Bonnie(1988).TenseasDiscourseAnaphor. ComputationalLinguis-tics,14(2):61–73.

Webber, Bonnie & Aravind Joshi (1998). Anchoring a Lexicalized Tree-Adjoining Grammarfor Discourse. In ManfredStede,Leo Wanner& Ed-uardHovy (Eds.),DiscourseRelationsandDiscourseMarkers:Proceedingsof theConference,pp.86–92.Somerset,New Jersey: ACL.

Webber, Bonnie,Alistair Knott & Aravind Joshi(1999a).Multiple DiscourseConnectivesin a LexicalizedGrammarfor Discourse. In 3rdIWCS,Tilburg,TheNetherlands.

Webber, Bonnie,Alistair Knott, Mathew Stone& Aravind Joshi(1999c).Dis-course Relations: A Structural and PresuppositionalAccountusing Lexi-calisedTAG. In Proc.of the36thACL, CollegeParkMD., pp.41–48.

Webber, Bonnie,Alistair Knott, Mathew Stone& Aravind Joshi(1999b).Whatare Little Treesmadeof: A Structural andPresuppositionalAccountusingLexicalizedTAG. In Proc.of the36thACL, CollegePark,MD, pp.151–156.

35

Xia, Fei, MarthaPalmer& Aravind Joshi(2000).A Uniform Methodof Gram-mar Extractionandits Applications. In Proc.of theJointSIGDAT Confer-enceonEmpiricalMethodsin NaturalLanguageProcessingandVeryLargeCorpora(EMNLP/VLC). HongKong.

36

Inf ormation structur eand pronominal referenceto clausally intr oducedentities

JEANETTE GUNDEL

Universityof Minnesota

[email protected]

M ICHAEL HEGARTY

LouisianaStateUniversity

[email protected]

KAJA BORTHEN

NorvegianUniversityof ScienceandTechnology

[email protected]

ABSTRACT. Clausallyintroducedentitiesin Englisharemorefrequentlyaccessibleto referencewith a demonstrative pronounthanwith thepersonalpronounit. This factcanbeexplainedontheassumptionthatsuchentitiesaretypically activated,but notbroughtinto focus,immediatelysubsequentto their introductioninto a discourse.However, clausallyintroducedentitiesare,infact,sometimesreferencedwith it immediatelysubsequentto theirintroduction.An examinationof thediscourseenvironmentsin whichthis is possibleprovidesinsightsinto thevariousfactors,includinginformationstructure,whichcanboostthesalienceof anentityandbring it into focus.

1 Intr oduction

Whenentitiesare introducedinto a discourseby a clause(or othernon-nominalexpression),they areaccessibleto immediatesubsequentreferencewith demon-strativepronouns,but comparatively lessaccessibleto referencewith personalpro-nouns,asnotedby Webber(1988)andotherssince1. Thus,Webber(1991)foundthatof 96pronominalreferencesreferringto theinterpretationof oneor morepre-viousclausesin written Englishtexts,only 15 usedthepersonalpronounit, whiletherestweredemonstrative thisor that. Personalpronounstendto favor reference

1Our examplesherewill befrom English,althoughsimilar restrictionson pronominalreferenceto clausallyintroducedentitiescanbefoundin otherlanguages.

37


to nominallyintroducedentities,andreferenceto aclausallyintroducedentitywithit is oftenimpossible,or at leasthighly infelicitous,asillustratedin thefollowingexamples.

(1) a. Therewasasnake on my desk.That scaredme.

b. Therewasasnake onmy desk.It scaredme.(Borthen,Thorstein,andGundel1997)

(2) a. Max destroyedhis leaf collectionlastnight. That wasdumb.

b. Max destroyedhis leaf collectionlastnight. It wasdumb

(3) a. SimplifiedEnglishdisallows theuseof passive, progressive, andper-fective auxiliary verbs,amongother things. This requiresengineers...

b. SimplifiedEnglishdisallows theuseof passive, progressive, andper-fective auxiliary verbs,amongother things. It requiresengineers...(Gundel,Hedberg, andZacharski1993)

(4) a. “We believe her, thecourtdoesnot,andthat resolvesthematter,” Mr.Montanarellisaidtodayof Ms. Lewinsky’s testimony thatshehadanindependentrecollectionof thedate.

b. “We believe her, the court doesnot, and it resolves the matter,” Mr.Montanarellisaidtodayof Ms. Lewinsky’s testimony thatshehadanindependentrecollectionof thedate.(New York Times, May 24,2000)

(5) a. Clonedhumansmight show higherratesof canceror otherdiseases,but we’d only find out by cloning themandwaiting to seeif disasterstrikes.Noneof this means,however, that... eventhathumancloningisn’t goingon right now. (Talbot,Margaret.February4, 2001. New YorkTimesMagazine, Section6, p.45.)

b. .... # Noneof it means,however, ...

(6) A: I readsomewherethat the poodleis oneof the most intelligent dogsaround.

B: well uhm..Idefinitelywouldn’t disputethat. (SwitchboardCorpus,Dia-log 2019)

B’: ??well uhm..Idefinitelywouldn’t disputeit .

(7) A1: Soyoufired her?

B: We’re goingto doa lot morethanjust fire her.

A2: Whatdoesthat mean?(from thesoapopera“The Bold andtheBeautiful”)

A2’: # Whatdoesit mean?2

2Note that it in this example,aswell asthe precedingone,would be infelicitous even if stressfallson theverb.

38

The demonstrative pronounin (1)-(7) refersto someentity (a situation,fact,act,etc.) introducedby apreviousclause.By contrast,thepronounit is eithermostnaturallyinterpretedasreferringto anentity introducedby anominal,asin (1)-(4),or it rendersthe sentenceinfelicitous in thecontext whenthereis no appropriatenominalantecedent,asin (5)-(7). In (2), for example,thepronounit is mosteasilyinterpretedasreferringto Max’s leafcollection,nothisactof destroying it. And in(7) , A2’ is infelicitousbecauseit , unlike that, cannotrefer to B’s statement“wearegoingto do a lot morethanjust fire her.”

In this paper, we arguethat factsregardingthedistribution andinterpretationof this/thatvs. it referringto clausallyintroducedentitiescanbeexplainedwithinthe theory of referenceand cognitive statusproposedby Gundel,Hedberg, andZacharski(1993andearlierwork). Approachedin thecontext of this framework,thesefactsalsoprovide insightsinto themoregeneralquestionof how variouslin-guistic factors,including informationstructure,promotethesalienceof discourseentitiesandbring theminto theaddressee’s focusof attention.

2 The GivennessHierar chy

Gundelet al (1988,1993)proposethat determinersandpronounsconstrainpos-sible interpretationsof nominalformsby conventionallysignalingthememoryorattentionstatusthat the intendedreferentis assumedto have in the mind of theaddressee.Gundeletal identify six cognitivestatuses.Thearrayof statuses,calledtheGivennessHierarchy, is presentedin Figure1.

infocus � activated � familiar � uniquely

identifiable � referential � typeidentifiable

(it)

�� thatthis

this N

� �� (thatN) (theN)

(indefinitethis N)

(a N)

Figure1. TheGivennessHierarchy(GH) andassociatedformsin English.

Statuseson thehierarchycorrespondto memoryandattentionstates,rangingfrom mostrestrictive, “in focus”, to leastrestrictive, “type identifiable.” An empir-ical claimof thetheoryis thatall languageshavewaysof codingcognitive statuseswith individual determinersandpronouns,andthatsuchformswill beusedappro-priatelyonly if thestatusthey conventionallycodewithin thelanguageis satisfiedin thegivencontext of use.Theformsthusserveasprocessingsignalswhichassisttheaddresseein restrictingpossibleinterpretations.

Thestatusesarein a unidirectionalentailmentrelation. If somethingis in fo-cus(centerof attention),it is necessarilyactivated(in working memory); if it isactivated,it is necessarilyfamiliar (in memory);if it is familiar, thentheaddresseecanassociatea uniquerepresentation;if theaddresseecanassociatea uniquerep-resentation,thenit is referential;andif it is referential,it mustbetypeidentifiable.

39

The theory thuscorrectlypredictsthat a given cognitive statuscanbe appropri-atelycodedby a form whichexplicitly signalsthatstatus,but alsoby formswhosemeaningsareentailedby thatstatus.In thelattercase(e.g.useof adefinitearticlefor areferentthatis in focus)theform is simplyunderspecifiedfor cognitive statusof theintendedreferent.

Theuseof underspecifiedformshaslimits, however, becauseof interactionofthe GivennessHierarchywith generalpragmaticprinciplesinvolved in languageproductionand understanding(seeGrice, 1975, Sperberand Wilson, 1986/95).Theimplicationalnatureof theGH givesriseto “scalarimplicatures”,in thesenseof Horn (1972),which further restrictthe distribution andinterpretationof refer-ring forms(seeGundelet al., 1993,GundelandMulkern,1998).For example,inEnglish,the indefinitearticle is rarelyusedif thestatusis higherthanreferential,resultingin associationof theindefinitearticlewith unfamiliarity. Useof theindef-inite articletypically implicatesby thefirst partof theQuantityMaxim (makeyourcontribution asinformative aspossible)thatconditionsfor usingamorerestrictiveform arenot met sincethe addresseeis not ableto uniquelyidentify an intendedreferent.Anotherresultof interactionof theGivennessHierarchywith theQuan-tity Maxim is thatmostin-focusreferentsarenotcodedwith demonstratives,eventhoughthey couldbe;anddemonstrativesoftenimplicatea focusshift.

As seenin Figure1, Gundelet al proposethatdemonstrative pronounsin En-glishcodethestatus“activated”,whereasthepronounit codesthemorerestrictivestatus“in focus”. This permitsanexplanationof factslike thosein (1)-(7), if theclausallyintroducedentiitiesin theseexampleshavebeenactivatedbut notbroughtinto focus.In thefollowing section,weexaminefactorsthatcontributeto bringinganentity into focus,includingtherole thatinformationstructureplaysin determin-ing thecognitive statusesof referentsintroducedby clauses,andthusthenominalformswhichcanbeusedto referto theseentities.

3 What brings an entity into focusof attention?

3.1 Syntactic structure

Theframework outlinedabovemakespredictionsabouttheappropriatenessof dif-ferentpronominalforms dependingon whetheror not the intendedreferentcanbe assumedto be in focusfor the addressee.Although the theoryitself doesnotpredictwhat bringsan entity into focus,Gundelet al. (1993)suggestthat “theentitiesin focusatagivenpoint in thediscoursewill bethatpartially-orderedsub-setof activatedentitieswhich are likely to be continuedastopicsof subsequentutterances.” Membershipin this set is partly, thoughnot wholly, determinedbysyntacticstructure.For example,subjectsanddirect objectsof matrix sentencesaremorelikely to bring anentity into focusthanelementsin subordinateclausesandprepositionalphrases(cf. theCenteringAlgorithmsof Grosz,Joshi,andWe-instein,1995a,b).Theseassumptionsmake it possibleto explain factsaboutthedistribution andinterpretationof demonstrativesandunstressedpersonalpronouns

40

(including it) suchasthoseillustratedin (8) and(9).

(8) a. My neighbor’s Bull Mastiff bit agirl on abike.

b. It’ s/That’s thesamedogthatbit Mary Benlastsummer.

(9) a. Searsdeliverednew sidingto my neighborswith theBull Mastiff.

b. #It’ s/That’s thesamedogthatbit Mary Benlastsummer.

Sincethe Bull Mastiff is introducedin matrix subjectposition (and is mostlikely alsothe topic) in (8a), it is broughtinto focus,andcanthereforebeappro-priatelyreferredto with eitherthat or it in (8b). Thepronounit is possiblein (8b)becausethe intendedreferentis in focus. The pronounthat is possiblebecauseanything in focusis alsoactivated,i.e. in working memory. But in (9), wheretheBull Mastiff hasbeenintroducedin a moreperipheralposition,it is activatedbutnotbroughtinto focus.Therefore,only referencewith that is possible.

Thisaccountcanbenaturallyextendedto factslike thosein (1)-(7) if wemaketherelatively uncontroversialassumptionthatentities(indirectly) introducedby awholeclause,or sequenceof clauses,will beactivated,but aremuchlesslikely tobebroughtinto focusthanentitiesintroducedby majorthematicargumentsof theverb. For example,in (2), at theconclusionof A’s utterance,theactof destroyingthe leaf collectioncanbeassumedto beactivated,sinceit wasjust introducedintheprecedingsentence,but not in-focus;thefocusof attentionaftertheutteranceisprocessedis onthereferentsof themajorargumentsin (2A), specifically, Max andthe leaf collection. Similarly, in (5), thecomplex situationconsistingof potentialdrawbacksto humancloning is renderedactivatedby the first paragraph,but wecanassumethat it is not renderedin focusgiven thehighersalienceconferredbythis passageon clonedhumans,ratesof cancer, andotherreferentsof mainclausearguments.

A factor propositionintroducedby anNPwithin aclauseis alsomorelikely tobebroughtinto focusthanonewhich is introducedby thewholeclause.Compare(10)with theexamples(1)-(7)above, for example.

(10) a. Then,Maria broughtup anotherfact. It sentshiversdown my spine.

b. Max thenintroducedanew proposition.But it wasrejected.

3.2 Semanticand pragmatic factors

Conditionswhichappearto boostthesalienceof entitiesalsoincludelessovertfac-torssuchascovert arguments,presuppositionsandprior beliefs,andeveninquisi-tive looks,all of which cancauseanentity to be“reprocessed”,andthusbroughtinto focus,evenwhenit is overtly mentionedonly once(seeBorthen,1997,Gun-del,Borthen,andFretheim,1999).

In (11),abaselinecasefor comparison,thespeaker, uponclausallyintroducingthefactthatlinguistsearnlessthancomputerscientists,canassumethatthis factisrenderedactivated,but not in-focus,for thehearer, leadingto a preferencefor thatover it in thefollow-up referenceto this fact.

41

(11) a. I hearlinguistsearnlessthancomputerscientists,andthat’s terrible.

b. ??I hearlinguistsearnlessthancomputerscientists,andit ’s terrible.

In (12), in contrast,thefollow-up referenceis madeby anotherspeaker, whichresultsin somewhatmorecomplicatedinferencesregardingthecognitive statusofthefactat issue.

(12) A: I just readthatlinguistsearnlessthancomputerscientists.

B: (i) That’s terrible! (ii) It ’s terrible!

At the completionof A’s utterance,B canassumethat the fact that linguistsearnlessthancomputerscientistsis at leastactivatedfor A. In responseB(i), B’suseof that signalsthe assumptionthat this fact hasbeenactivated,but possiblynot broughtinto focus,by A’s utterance,therebyinviting A to infer that the factis news to B. In responseB(ii), B signalsthe assumptionthat the fact is in focusfor A, or oughtto be,consistentwith it beingacceptedbackgroundinformationfordiscoursein the relevant socialcircle; this invites A to infer thatB alreadyknewthefact.

In (13) below, thepropositionthatB hasa dentalappointmentis clausallyin-troducedby A’s utterance.If themereutteranceof a sentencedoesnot bring theexpressedpropositioninto focus, this would explain why (13)B’ soundsunnatu-ral, given that it requiresthereferentto be in focus,whereasthat merelyrequiresactivation.

(13) A: You have adentalappointmentat noon.

B: That’s true.

B’: ??It’s true.

B”: It’s true,then.

But (13) B” is noticeablymore acceptablethan (13)B’. Following Gundelet al. (1999), we suggestan explanationof this fact, drawing on a relevance-theoreticapproachto thepragmaticsof languageunderstanding(SperberandWil-son,1986/95). The word then in B” functionsas an interpretive particle whichconveys themeaningthat thecontentof thesentenceit is appendedto follows byway of inferencefrom somethingthe addresseejust said. The responseby B in(13)B” meansessentially, “Givenyour assertionthat I have a dentalappointmentat noon,thenI cantake it asconfirmedthat I have a dentalappointmentat noon.”Theonly way theutterancein B” canyield contextual effectsfor A is if A’s utter-anceconfirmedthetruthof apropositionthatB hadbeenquestioning,andB knowsthatA is awareof this. Thus,thefactthatB hadadentalappointmentatnoonwasnot activatedfor thefirst time by A; rather, A’s utterancebroughtinto focusa factthatwasalreadymutuallymanifestto bothA andB beforehand,therebylicensingtheuseof it in B”.

Saliencecanalsobeboostednon-linguistically. For example,theexchangein(14)below is fully naturalif A givesB askepticallook duringtheindicatedpause.

42

(14) A: Why didn’t youcometo therehearsalyesterday?

B: I thoughtI told you. I hadto helpPetermove. (Pause)It’s true!

Theskepticallook communicatesA’s skepticismaboutthetruth of thepropo-sition just expressedby B, thuscausingthe propositionthat B hasto help Petermoveto bereprocessed(by bothA andB) andassuringthatit is mutuallyin focus,makingit accessibleto referencewith it.

Salienceof anentity in theenvironmentalsosufficesfor pronominalreferencewith it. If A andB are in a room togetherwith a babywho suddenlybegins towalk, A canproducetheutterancein (15),or, if A seesB watchingthebabywalk,theutterancein (16).

(15) Will you look at that! Thebaby’s walking. (Jackendoff 2001)

(16) Isn’t it great?[it = thefactthatthebabyis walking]

4 The roleof information structur e

Thecognitive status,andthereforetheaccessibilityto pronominalreference,of aclausallyintroducedentity is partlyconstrainedby theinformationstructureof theutterancein which it is introducedinto adiscourse3

In particular, informationstructureyieldssomestriking effects,but alsoa sur-prisingasymmetry, whenhigherorderentitiesareintroducedby (or within) clausalcomplements.

Entitiesintroducedby clausalcomplementsto bridgeverbs,suchasthink, be-lieve, and say, exhibit the familiar patternof being renderedactivated, but notin-focus,throughmentionby a clause.This is shown by the naturallyoccurringexamplein (17)below, aswell asby theconstructeddatain (18), testedonasmallsurvey of Englishspeakers4

(17) Ising reportedlybelievedthathis negative resultswould hold in higherdi-mensionsaswell.In bf this conjecturehewaswrong.(AmericanScientist88:385)

In this/ #it, hewaswrong.

(18) WhatdoesAlex think?3By informationstructure,we meana bifurcationof materialin anutteranceinto whathasbeen

called focus versusground,commentversustopic, or rhemeversustheme. This notion is not tobe identified with contrastive focus or with the more generaldistinction betweennew versusoldinformation. Informationstructuralfocusis alsodistinct from thecognitive status“in focus”. SeeVallduvi (1990)andGundel(1999a)for moredetaileddiscussionof relatedterminologicalandcon-ceptualissues.Wewill indicateinformationstructuralfocusby thesubscriptF .

4Theuseof it in (17)wouldbejustasinfelicitousif thePPwerenotpreposed.Thus,theinfelicityof it in (17) cannotbeattributedto its incompatibilitywith thesecondaryfocal stressit bearsin thisposition.

43

A: Alex believes[F thatthecompany destroyedtheFILE].

B: That’s false;thefile hasbeensubmittedto thedistrict judge.

B’: # It ’s false;thefile hasbeensubmittedto thedistrict judge.

When(18A) is usedwith thefocus-structureshown, to introducetheproposi-tion thatthecompany destroyedthefile, theresponseby B usingthat is muchmorefelicitousthantheresponsewith it. However, it andthatareequallygoodwhenthecomplementclauseis in theground(theme/topic)of A’sutterance,asin (19).

(19) A: Alex [F INSISTS/BELIEVES]thatthecompany destroyedthefile.

B: But that’s/it ’s false;thefile hasbeensubmittedto thedistrict judge.

Sincean entity associatedwith the ground(theme;topic)is alreadyat leastfamiliar to the addresseeprior to the utterance(seeGundel1988) inter alia), itsmentionwithin the utterancesuffices to bring it into the focusof attention,if itdoesnotalreadyhave thatstatus.

In (17)-(19), relationalgivenness/newnessand referentialgivenness/newness(in the senseof Gundel,1988, Gundel1999a,b)are coextensive. For example,the informationstructuralfocus in (18) representsa propositionthat is not onlynew in relationto thetopic (whatAlex believes),but alsoreferentiallynew to thehearer;andtheclausalcomplementin (19A) (thatthecompany destroyedthefile)representsa propositionwhich is not only given in relation to the informationalstructuralfocus;it is alsoreferentiallygivenin thesenseof beingat leastfamiliar,andprobablyalsoactivated. But materialin the informationalfocusdoesn’t haveto bereferentiallynew (seeGundel1980,Gundel1999a,Gundel1999b,Vallduvi1990,Lambrecht1994).Sowhenwe have a bridgeverbcomplementwhich is aninformationstructuralfocus,but is alreadyactivatedin thediscourse,which factorwins out? Is anentity expressedby sucha complementrenderedin-focusor doesit remainmerelyactivated?Is it accessibleto referencewith it, or only with that?Consider(20).

(20) A1: I believe thatthecompany destroyedthefile, but noteverybodydoes.

B1: WhatdoesAlex believe?

A2: Alex believes[F thatthecompany destroyedthefile].

B2: But it ’s/that’s false;thefile hasbeensubmittedto thedistrict judge.

(20B2)suggeststhatit is referentialgivenness( cognitive statusof adiscourseentity), andnotrelationalgivenness(topic-focusstructure)thatdetermineswhetherthecomplementof abridgeverbwill bebroughtinto focus.

But now flip theproblemaround.Contentin thetopic/groundof anutterancedoesnot alwayshave a high degreeof referentialgivenness.Its cognitive statusmaybemerelyfamiliar, but not necessarilyactivated. Sowhenwe have a bridgeverbcomplementwhich is groundmaterial,but new to thediscourse,which factorwinsout?Is anentity introducedby suchacomplementrenderedin-focus,because

44

it is in the ground,or merelyactivated,becauseit is new to the discourse?Is itaccessibleto referencewith it, or only with that? Consider(21) [secondarystresson murdered]:

(21) a. Alex is hopeless.

b. He [F INSISTS]thatTomwasmurdered,for example,

c. –eventhoughthere’s notashredof evidencefor that.

–eventhoughthere’s notashredof evidencefor ıit.

Useof it is as felicitous as that in (21c). The information structureof (21)forcesan interpretationwhere the contentof the complementclauseis alreadyfamiliar, sothat(21b)rendersit in-focus,makingit availableto referenceusingit.Thus,presentationof a clausallyintroducedentity in thegroundof anutteranceisanotherway to promotesalience,andbring the entity into focus,even if it is, infact,new to thediscourse.

With bridgeverb complements,we thusappearto have an asymmetricsitua-tion: bifurcationinto focus/groundhasnoeffecton thecognitive statusof anentityintroducedwithin theinformationstructuralfocus5 But it canhave aneffect whenanentity is mentioned(evenintroduced)within groundmaterial,becausementionwithin thegroundnecessarilysignalsa highercognitive statusfor theentity. Thisconclusionis preliminary, however, in thatthejudgmentsaresubtle,andnaturallyoccurringdatathatwouldbeardirectlyon theissueis sparse.

5 Lexical structur eversusinformation structur e

When the bridge verb in an example like (18) is replacedwith a factive verb,demonstrative and personalpronounscan both be usedto immediatelyrefer tothepropositionexpressedby thecomplementclause,regardlessof theinformationstructureof A’sutterance,asseenin (22).

(22) A: Alex verifiedthatthecompany destroyedthefile.

B: That’s false;thefile hasbeensubmittedto thedistrict judge.

B’: It’s false;thefile hasbeensubmittedto thedistrict judge.

Thus, the contrastin (18) betweensubsequentreferencewith it versusthatis not exhibited in (22). The lexical semanticsof the factive verb enforcestheconditionthat the entity expressedby thecomplementclausebe alreadyfamiliar(or at leastcapableof beingaccommodatedasfamiliar) to theaddressee,sothatitsfurthermentionin A’sutterancerendersthisentity in-focus.

In order to understandthis fully, it is useful to note that this patternis notconfinedto complementsof factive verbs. It is alsoobtainedin complementsto

5Gundel(1999a)makesa similar observation, concludingthat mentionwithin the informationstructuralfocus(her“semanticfocus”)doesn’t necessarilybring anentity into focusof attention.

45

certainnon-factive (andnon-bridge)verbs,includingagree, emphasize, deny, anddoubt, andin complementsto thenon-factive adjectival predicatebecertain.6

(23) a. Alex andSusanagreethat the company destroyed the file. I’m sur-prisedthatthey believe it.

b. Alex andSusanagreethat the company destroyed the file. I’m sur-prisedthatthey believe that.

(24) A: Alex is certainthatthecompany destroyedthefile.

B: That’s false:thefile hasbeensubmittedto thedistrict judge.

B’: It’s false:thefile hasbeensubmittedto thedistrict judge.

As with factive predicates,thepatternin (23)-(24)is onein which it is at leastasfelicitousasthat in referringto thecontentof thecomplementclause,and,insomecases,moreso.

Thepredicatesin (23)-(24)arenot factive (in thesensemadeclearby Kiparskyand Kiparsky (1971)) since they don’t commit the speaker of the ascriptioninwhich they occur to the truth of their complementclauses.However, they sharewith factives a slightly more subtlesemanticproperty: they are felicitous whenthe proposition,fact, or situationexpressedby the complementclauseis alreadyacceptedasgivenor familiar in thediscourse(seeHegarty, 2001). Usinga situa-tion variablein thesemantics,in thecontext of DiscourseRepresentationTheory(KampandReyle, 1993),theinterpretationof thefactive ascriptionin (22) canbeexpressedby theDiscourseRepresentationStructure(DRS)shown in (25)below7

Theascriptionswith ıagreeandcertainin (23)-(24),thoughnon-factive,wouldhave identicalDRS’s,with trivial substitutionof theverbdenotations.

(25)

u � v� z� sAlex(u)Company (v)File (z)destroy (v� z� s) (wo)verify (u � λw [ destroy (v� z� s)(w) ])

In contrast,a belief ascriptionsuchas that in (18A), usinga bridgeverb, isinterpretedsemanticallyas just a relationbetweenAlex and the propositionex-pressedby thecomplementclause.A DRSfor (16A) is presentedin (26).

6Cattell (1978)noticedthat thesenon-factivespatternwith factivesin wh-extractionfrom theircomplements.SeealsoMelvold (1991),Hegarty (1992),andSchulz(1999)for discussionof thisclassof predicates.

7SubordinateDRSsareabbreviatedasformulashereto save space.For semanticrepresentationsusinga situationvariable,seeGinzburg (1995a,b),and,for similar structureswith aneventvariable,(Higginbotham1985,Higginbotham1989).

46

(26)

u � v� zAlex (u)Company (v)File (z)believe (u � λw [ � s [destroy (v� z� s)(w)] ])

Of course,theascriptionmadeby A in (18)couldexpressa propositionwhichis alreadyfamiliar to the hearer. The propertydistinguishingbridgeverbsfromthefactive andotherpredicatesdiscussedhereis not thatthecontentof thebridgeverbcomplementmust beunfamiliar, but only thatit canbe.Bridgeverbs,unlikeotherpredicatesdiscussedhere,do notassumethefamiliarity of thecontentof thecomplement.

Interrogatives patternwith factive complementswith regard to the statusofabstractentitiesmentionedby or within them. A naturallyoccurringexampleisshown in (27).

(27) Onecommonattributeof ascientistis anunusuallyacutesenseof numbersandtheir implications.A senseof numbers- why do I dwell on this observation?Perhapsit’ s be-causewewhocomefrom abackgroundof engineering...(AmericanScientist88:378)

(28) A: Alex wonderswhetherthecompany destroyedthefile.

B: It ’s not likely. Thefile containedno incriminatinginformation.

B: That’s not likely. Thefile containedno incriminatinginformation. [it/that =thatthecompany destroyedthefile]

(29) a. Max wonderswhodestroyedthefile; it hasimpededtheinvestigation.

b. Max wonderswho destroyed the file; that has impededthe investigation.[it/that = thatsomeonedestroyedthefile]

The possibility of immediatesubsequentreferencewith a personalpronounin (28)-(29) follows from the presuppositionalnatureof questions.To simplify,within DRT, thewonder-ascriptionin (28A) shouldberepresentedwith a DRSofthe form shown in (30), whereφ is an appropriaterelationbetweenAlex andthepropositionp specifiedon thepenultimateline of theDRS8

8To unsimplify, questionsare, in fact, constrainednot only by the formal semanticconditioncapturedhere,but by rich contextual conditionson what would count as a suitableanswerto aquestionin a given context. SeeGinzburg (1995a),Ginzburg (1995b),andAsherandLascarides(1998).Theimportantpoint, for presentpurposes,is thattheseaccountswould incorporate,andaddto, thepresuppositionalconditiongivenhere.Theproposalssketchedherewould thereforebea partof anaccountgivenaccordingto theserichertheoriesof theinterpretationof questions.

47

(30)

u � v� z� pAlex (u)Company (v)File (z)p � w�� λw� s [destroy (v� z� s)(w)]φ

Interpretedasin (30), thewonder-ascriptionin (28A) is a questionaboutthepropositionthat the company destroyed the file. This shouldbe the form of anysemanticaccountof thewonder-ascriptionwhich capturesthepresuppositionalityof the embeddedquestion: the propositionthat the company destroyed the filemustbeanestablisheddiscourseentity prior to theutteranceof (28A), or it mustbe accommodatedin the senseof Heim (1982). The assertive contentof (28A)shouldbe capturedin the last line of the DRS,φ. On onerealizationof φ, givenin Hegarty(2001),(28A) assertsthatAlex is in thestateof wonderwith respecttothepropositionthatp holdsof theactualworld, wo.

Thus,the presuppositionalityinvolved in the lexical structureof a factive (orrelated)predicate,andthesemanticpresuppositionality of embeddedquestions,areadditionalfactorswhichcanbring anentity into focus.In thesecases,informationstructurehasnobearingon thecognitive statusof theclausallyintroducedentity.

6 Conclusion

In this paper, we addressedthefactthatclausallyintroducedentities,immediatelysubsequentto their introductioninto a discourse,are typically accessibleto ref-erencewith a demonstrative pronoun,but not with the personalpronounit. Wefound that this factcanbeexplainedon thebasisof theobservation thatsuchen-tities aretypically activated,but not broughtinto focus,upontheir introductiontoa discourse.However, clausallyintroducedentitiesare,in fact, sometimesrefer-encedwith it immediatelysubsequentto their introduction.An examinationof thediscourseenvironmentsin which this is possibleprovidesimportantinsightsintothevarioussyntactic,semantic,andpragmaticfactorsthat canboostthe salienceof anentity andbring it into focus.

We’veshown thatinformationstructure,in thesenseof afocus-groundbifurca-tion, is onesuchfactorwhenanentityis mentionedwith abridgeverbcomplement,but only in a way which is asymmetric,dependingon whethertheentity is men-tionedwithin focal or non-focalmaterial.Whenthecomplementis focal, thereisno effect: the cognitive statusof an entity expressedby a focal complementde-pendsentirely on the referentialgivenness/newness(i.e. the cognitive status)oftheentity. But whenthecomplementis partof theground(topic/theme),theentityis broughtinto focus.

In factive complementsandembeddedquestions,thelexical natureof theem-beddingpredicateand the semanticnatureof the constructionrequirean entity

48

mentionedwith thesubordinateclauseto betreatedasreferentiallygivenindepen-dentlyof theinformation-structureof theutterance.Thissuggeststhatthesaliencepromotingeffect of informationstructureis indirect. It is thereferentialgivennessof the ground,i.e. the fact that topicsareat leastfamiliar, andnot informationstructureperse,whichcontributesto bringinganentity into focus.

Bibliography

Asher, N. andA. Lascarides(1998).Questionsin dialogue.LinguisticsandPhi-losophy21, 237–309.

Borthen,K., F. Thorstein,andJ.K. Gundel(1997).Whatbringsa higher-orderentity into focusof attention?In R. Mitkov andB. Boguraev (Eds.),Opera-tional factors in practical, robustanaphora resolution, pp.88–93.

Cattell,R. (1978).On thesourceof interrogative adverbs.Language 54, 61–77.

Ginzburg, J. (1995a).InterrogativesI. LinguisticsandPhilosophy18, 459–527.

Ginzburg, J.(1995b).InterrogativesII. LinguisticsandPhilosophy18, 567–609.

Grice, H. P. (1975).Logic andconversation.In P. Cole andJ. Morgan(Eds.),SyntaxandSemantics3, Speech Acts, pp.41–58.AcademicPress.

Grosz,B. J.,A. K. Joshi,andS.Weinstein(1995a).Providing aunifiedaccountof definitenounphrasesin discourse.In Proceedingsfor the 21stAnnualMeetingof theAssociationfor ComputationalLinguistics, pp.44–50.

Grosz,B. J., A. K. Joshi,andS. Weinstein(1995b).Towardsa computationaltheoryof discourseinterpretation.ComputationalLinguistics21, 203–25.

Gundel,J.K. (1980).ZeroNP-anaphorain Russian:acaseof topic-prominence.In Proceedingsfrom the Parasessionon Anaphora, pp. 139–146.ChicagoLinguistic Society.

Gundel,J. K. (1988).Universalsof topic-commentstructure.In M. Hammondet al. (Eds.),Studiesin syntactictypology, pp.209–239.JohnBenjamins.

Gundel,J. K. (1999a).On threekinds of focus. In P. Boschand R. van derSandt(Eds.),Focus.Linguistic,CognitiveandComputationalPerspectives,pp.293–305.CambridgeUniversityPress.

Gundel,J. K. (1999b).Topic, focusandthe grammarpragmaticsinterface.InN. J. Alexanderand M. Minnick (Eds.),Proceedingsof the 23rd AnnualPennLinguisticsColloquium, Volume6.1,pp.185–200.PennWorking Pa-persin Linguistics.

Gundel, J. K., K. Borthen, and T. Fretheim(1999). The role of context inpronominalreferenceto higherorder entititesin English andNorwegian.In P. Bouquetet al. (Eds.),ModelingandUsingContext. ProceedingsfromtheSecondInternationalandInterdisciplinaryConference, CONTEXT’99,LectureNotesin Artificial Intelligence1688,pp.475–478.Springer.

49

Gundel,J. K., N. Hedberg, andR. Zacharski(1988).Thegenerationandinter-pretationof demonstrative expressions.In Proceedingsof the XIIth Inter-nationalConferenceon ComputationalLinguistics, pp.216–221.JohnVonNeumannSocietyfor theComputingSciences.

Gundel,J. K., N. Hedberg, andR. Zacharski(1993).Cognitive statusandtheform of referringexpressionsin discourse.Language 69, 274–307.

Gundel,J. K. andA. Mulkern (1998).Quantity implicaturesin referenceun-derstanding.Pragmaticsand cognition 6: Specialissueon the conceptofreferencein thecognitivesciences, 21–45.

Hegarty, M. (1992).Adjunct extractionwithout traces.In D. Bates(Ed.), Pro-ceedingsof the Tenth West CoastConferenceon Formal Linguistics, pp.209–223.Stanford:CSLI.

Hegarty, M. (2001).Referenceto abstractentitieswithin clausalcomplements.In Proceedingsof theAnnualMeetingof theBerkeley LinguisticsSociety.

Heim, I. (1982).TheSemanticsof DefiniteandIndefiniteNounPhrases. Ph.D.thesis,Universityof Massachusetts,Amherst.

Higginbotham,J.(1985).On semantics.LinguisticInquiry 16, 547–593.

Higginbotham,J. (1989). Elucidationsof meaning.Linguisticsand Philoso-phy12, 465–517.

Horn,L. R. (1972).On thesemanticpropertiesof logical operators in English.Ph.D. thesis,UCLA.

Jackendoff, R. (2001).Referenceandtruth. PlenaryTalk, Annual MeetingoftheLinguistic Societyof America,Washington,D.C.

Kamp, H. andU. Reyle (1993).From Discourse to Logic. Kluwer AcademicPublishers.

Kiparsky, P. andC. Kiparsky (1971).Fact. In D. Steinberg andL. Jacobovits(Eds.),Semantics, pp.345–369.CambridgeUniversityPress.

Lambrecht,K. (1994).Informationstructure and sentenceform: topic, focus,andthementalrepresentationsof discoursereferents. CambridgeUniversityPress.

Melvold, J. (1991).Factivity anddefiniteness.In L. ChengandH. Demirdash(Eds.),MorePapersonWH-Movement, Volume15,pp.97–117.MIT Work-ing Papersin Linguistics.

Schulz,P. (1999).Gettingthefacts,finite complements,factiveverbsandtheiracquisition. Ph.D. thesis,Tubingen.

Sperber, D. andD. Wilson (1986/95).Relevance: Communicationand cogni-tion. Blackwell.2ndEdition.

Vallduvi, E. (1990).TheInformationalComponent. Ph.D. thesis,UniversityofPennsylvania.

50

Webber, B. L. (1988).Discoursedeixisanddiscourseprocessing.Technicalre-port,Universityof Pennsylvania.

Webber, B. L. (1991).Structureandostensionin theinterpretationof discoursedeixis.Language andCognitiveProcesses6(2), 107–135.

51

Entangled Inf ormation Structur e: AnalysisofComplexSentenceStructur es

NOBO KOMAGATA

Departmentof ComputerandInformationScience,Universityof Pennsylvania

[email protected]

ABSTRACT. While informationstructurehastraditionallybeenviewed asa singlepartitionofinformationwithin anutterance,thereareopposingviews thatidentify multiplesuchpartitionsinanutterance.Theexistenceof alternative proposalsraisesquestionsaboutthenotionof informa-tion structureitself andalsoits relationto discoursestructure.Thispapersupportsthetraditionalview by observingcontextual requirementsandlinguisticphenomenaassociatedwith informationstructurefor eachalternative.

1 Intr oduction

Traditionally, informationstructure(IS) hasbeenviewedasnon-recursive, matrix-level organizationof informationwithin anutterance(e.g.,Mathesius1975). Ac-cordingto thisapproach,evenacomplex utterancehasonly oneIS partitionascanbeseenbelow (adaptedfrom Lambrecht1994).

(1) Q: Why did youhit him?

A: [I hit him]Theme[becauseheinsultedme]Rheme.

Hereandthroughoutthispaper, theIS labels‘theme’and‘rheme’(T andR, respec-tively, in later examples)areusedinsteadof moreoverloadedterms,e.g.,‘topic’and‘focus’, mostcloselyfollowing Steedman(2000).Althoughthenotionsassoci-atedwith varioustermsmaydiffer in many respects,we try to limit ourdiscussionto theessentialpropertiesof themeandrhemeinvolving binaryinformationalcon-trastbetweenthem(cf. Communicative Dynamismof Firbas1964).

In contrastto the traditionalview, someresearchersobserve multiple IS par-titions within an utterance.For example,Kruijf f-Korbayova andWebber(2001)proposethefollowing analysis.

(2) Although[Clydemarried]T[BERTHA]R, [he]T[did not inherit a PENNY]R.

SMALL CAPITALS areusedto indicatepitch accents.In addition,Partee(1996)considersevena recursive IS suchasthefollowing.

53


(3) [WhatconvincedSusanthat[our arrest]T[wascausedby HARRY]R]T[wasarumorthat[someone]T [hadwitnessedHarry’s confession]R]R.

Naturally, theexistenceof threecompetingviews posesa challengeto thedefini-tion of IS.1 In addition, this issueis also relevant to the analysisof the relationbetweenIS anddiscoursestructure(DS). To seethis point, let usfirst classifytheabove-mentionedthreeapproachesby referringto thespanof a theme-rhemepairas‘domainof IS’.

(4) a. Traditionalview: Domainof IS = utterance

b. Kruijf f-Korbayova andWebber:Domainof IS = clause

c. Partee:Domainof IS = utteranceandclause(recursive)

Wethenobserve theideaof ‘thematicprogression’studiedby Danes(1974),whichis schematicallyshown below. T i andRi referto thethemeandrheme,respectively,of the ith utterancein adiscourse.

(5) T1 � � R1!T2 (=R1) �"� R2!

T3 (=R2) � � R3

If we assumethat DS is the organizationof discourseunits correspondingtoclauses(e.g.,GroszandSidner1986), thematicprogressionseemsto be able tocharacterizethe IS-DS relationquite well, especiallyif all sentencesaresimple.Danes’s ideacanalsobeinterpretedin thefollowing way:

(6) TheDSof adiscoursecanbedeterminedby theDSprior to thecurrent(last)utteranceandtheIS of thecurrent(last)utterance.

I would like to seethis astheprimarycontactpoint betweenthenotionsof IS andDS. However, oncecomplex sentencesare involved, the situationappearsmorecomplicated.For example,theanalysisof Kruijf f-Korbayova andWebber(2001)involving two IS’s in asingleutterancemightberepresentedasfollows,wheretheexample(2) wouldcorrespondto thelower level:

(7) T1 �"� R1!?

? #[ T2a � � R2a T2b �"� R2b ]

IS1 IS2

1As pointedout by oneof the reviewers,thesedifferentviews may reflectthe black-and-whitesituationinvolving IS. But pursuingonepositionagainstotherswithout acceptingthe mixed viewseemsessentialfor adeeperunderstandingof thesubject.Althoughoneof thereviewer statesthatitis “quite plausiblethatclauseshaveIS”, I donotthinksuch‘plausibility’ haseverbeendemonstrated.

54

With multiple IS’s in a singleutterance,the description(6) would no longerhold in its givenform. Wemayneedto askquestionssuchasthefollowing. Wouldthe DS analysisproceedin two steps,i.e., IS1 is usedto form the DS up to thatpointnot includingIS2? Or, wouldsomeform of complex IS (IS1 andIS2) beusedto form theDSupto andincludingIS2 all atonce?Theformeranalysiswouldraisea questionabouttherole of thesubordinator(at thebeginningof thefirst clause),which is supposedto connectthe two clauses.The latter analysiswould raiseaquestionabouttherelationbetweenthetwo IS’s in connectionto DS.Thesituationwouldbeevenmorecomplicatedwith theanalysisof Partee.

Betweenthe threealternatives, this papersupportsthe traditionalview of ISthroughanequivalentproposition:thereis exactlyoneIS partitionevenfor acom-plex utterance.Thepresentpositionis alsorelatedto the idea: linguistic markingof informationstructureis a matrix-level phenomenon(Komagata1999,p. 37).Oneof the consequencesof this position is that the domainof IS is not fixed tothe unit of DS (i.e., clause). In otherwords,the IS domainwould appearentan-gled aroundclauseboundaries.However, this potentialcomplicationseemsto beinevitablefor developingaDSoutof bothsimpleandcomplex utterancesin awayconsistentwith theview (6).

Theconstructionswefocusin thispaperarecomplex structuresinvolving sub-ordinatorssuchasalthoughandbecause. The (sentence)coordinatestructureisnot discussedin this paperas it is fundamentallydifferent from the subordinatestructure(e.g.,Quirk et al. 1985,pp. 920). It canbe consideredasa sequenceof utterances,eachof which may containits own IS. Furthermore,if a multiple-clausestructureis consideredascoordinatestructure,e.g.,nonrestrictive sententialrelative clause,therecanbeaseparateIS for eachutterance.

Thesecondqualificationis thatwedo notdiscussaspecialcaseof IS partitionwithin anembeddedclausesuchasthefollowing.

(8) Q: Whatdid you think Marcelproved?

A: [I thoughtMarcelproved]T[completeness]R.

While this type of ‘non-traditional’ constituentsare fairly common(Steedman2000),they donotappearin thetypeof complex structuresdiscussedin thispaper.

Theorganizationof thispaperis asfollows. Section2 discussesproblemswith(2). Section3 pointsoutproblemswith (3). In Section4,wesupportthetraditionalview of IS, mainlyby discussingpotentialcounterexamples.

2 IS Partition within the SubordinateClause

2.1 SemanticMoti vation

The main point of Kruijf f-Korbayova andWebber(2001) is that we canexplainthesemanticsof althoughif we consideran IS partition for eachclauseasin (2).

55

Roughly, their ideais thattheconventionalimplicaturesfor althoughcanbespec-ified in termsof thealternative setsassociatedwith thethemesandtherhemes,forthetwo readingsof although, i.e.,denialof expectationandconcessive opposition.

Their approachadoptsthe framework of Steedman(2000),which is basedonalternativesemantics(Rooth1985),anddoesclarify theinterpretationof although.But thequestionhereis whethertheeffect is dueto IS. To seethispoint, let usnowexaminethefollowing example.

(9) Themarriageof Clydeto BERTHA did not let him inherit a PENNY.

This sentenceseemsto besubstitutablefor (2) in virtually any context. In particu-lar, bothsentencescontainthesamecontrastive situationinvolving Berthain rela-tion to thepropositionthatClydeis marriedto Bertha.In (9), thesubject-predicaterelationis no longerbasedon although, but dependson thecontrastive semanticsassociatedwith Berthaasin (2), which occurswithin a nounphrase(with no em-beddedclauses).I do not know of any proposalfor anIS partitionwithin this typeof nounphrases.Althoughonemaycontestthis assumption,it seemsthatsuchamovewould leadto amajorrevision of thestandardview of IS. Thus,thecontrastobservedin (9) and(2) mustberepresentedsemanticallyregardlessof thepresenceof anIS.

Sucha semanticeffect canactuallybeaccountedfor by theanalysisof Steed-man (2000), which distinguishestwo levels betweenIS and focus-background.Note that the notion of ‘focus’ here is as in (Rooth 1985) and is not the sameas‘rheme’. While ‘rheme’ is a componentof IS, ‘focus’ is a notion tightly con-nectedwith a phonologicalprominence.Further, a focuscanappearin a themeora rhemeascanbeseenbelow.

(10) Q: I know thatMarcel likesthemanwho wrotethemusical.

But whodoesheADMIRE?

A: [Marcel ADMIREStheme-focus

]T[thewomanwho DIRECTEDrheme-focus

themusical]R.

With appropriatesemanticoperations,both(2) and(9) couldbeanalyzedin asim-ilar way in termsof thecontrastive situation. This paperdoesnot show how thiscanbedoneasit is notourpoint. But it mustbesimilar to theapproachof Kruijf f-Korbayova and Webber(2001), replacingthe theme-rhemedistinction with thefocus-backgrounddistinction.

2.2 Availability of Contexts

In their paper, Kruijf f-Korbayova and Webber(2001)considera context for (2)suchas a question“Is Clyde HAPPY?” (for the concessive-opposition interpre-tation). But it is not clear whetherthis or other questionscan actually providethe right context for the proposedIS, which is assumedfor both the denial-of-expectationandconcessive-opposition interpretations.Beforeproceeding,weneed

56

a few notes.The questiontest is useful in many cases.But it is alwayspossibleto respondto a questionindirectly. In sucha response,the IS cannotnecessarilybe identifiedbasedon thequestion.For a moreprecisediscussionof IS, we mayneedto considera moreformal approachsuchastheonedevelopedby Steedman(2000).But thispaperremainsopenin this respect.

Onepoint we canstill make is thatanisolatedquestioncanbeusedin supportof anIS in a direct responseto thequestion,asin (1). Thefollowing example(thedenial-of-expectation interpretation),in conjunctionwith an additionalutteranceprior to thewh-question,canshow acertainIS.2

(11) Q: I know Clydemarriedoneof thoserich women.But whathappenedtohim afterthewomandied?

A: [AlthoughClydemarriedBERTHA]T1, [he]T2 [did notinheritaPENNY]R.

Notethatit is not crucialthattherearetwo, discontiguousthemesabove.However, it seemsdifficult to demonstratean IS partitionwithin thesubordi-

nateclauseeven with a directquestionintendedto singleout a rhemein suchanenvironment.As anattempt,let usconsiderJapanesein which a wh-word canbeplacedin a subordinateclausefreely asin thefollowing example(grammaticalla-bels: TOP: topic/thematic,NOM: nominative, ACC: accusative, COP: copula,Q:question).

(12) Dare-ga Ken-o tasuketa-kara Naomi-ga koreta-no?who-NOM Ken-ACC helped-because Naomi-NOM could.come-Q

“Naomi wasableto comebecauseKenis helpedby whom?”

An interestingpoint is that even this type of questioncannotgive rise to an ISpartition within the subordinateclause. To seethis, we usethe propertythat nopartof therhemecanbeomittedin a response.

(13) A1: #Erika.

A2: Erika-ga tasuketa-kara.Erika-NOM helped-because“BecauseErika helped(him).”

It is not possibleto respondto sucha questiononly with the constituentcorre-spondingto thewh-word; theentiresubordinateclauseis neededin theresponse.

2Oneof thereviewerspointedout thattheconcessive-oppositionmayleadto separateutteranceswith theirown IS’s. Then,it would beanalogousto thefollowing typeof parallelstructurewith twoutterancesandtwo IS’s.

(1) Q: What’shappenedto JohnandMary?

A: [John]T [went to theZOO]Rand[Mary]T [went to theMUSEUM]R.

If this canbeshown, theIS analysisof (2) maybepossiblefor concessive opposition.However, wenotethat(Kruijf f-Korbayova andWebber)assumethesameIS analysisfor thebothinterpretations.

57

Note that it is perfectlyfine to respondto certainwh-questionswith just a noun.While thisdemonstrationis notaproof, it appearsconsistentwith ourposition.

Next, welook at intonationin Englishasanadditionalsupportfor theargumentthat the IS analysisin (2) is hardto comeby. As a non-native speaker of English,I cannotjudgetheappropriatenessof intonationin English.3 But I conjecturethattheexamplesbelow canbeusedin supportof thecurrentposition.

First, Steedman(2000)presentsanexamplewherethe responseto a questionintroducesanimplication.

(14) Q: DoesMarcel loveopera?

A1: MarcellikesMUSICALSH*

.

A2: MarcellikesMUSICALSL+H*

.

Here, (A1) is analyzedas a rheme(with the rhemetune), a completeutterance;(A2) is analyzedasa theme(with thethemetune),with animplied rheme.For thecaseof (A2), if the respondentthinks that the inquirer would not understandherimplied rheme,shewould have uttereda moreexplicit responsedependingon theimplication,asin thefollowing.

(15) A2a: As MarcellikesMUSICALSL+H*

, helovesOPERAH*

.

A2b: AlthoughMarcellikesMUSICALSL+H*

, hedoesn’t love OPERAH*

.

Structurally, (A2b) is parallelto (2). Now, let ussupposethatsomecontext allowstheIS in (2). Theintonationpatternof (2) wouldbeasfollows (applyingSteedman2000):

(16) Although[Clydemarried]T[BERTHAH*

]R, [he]T[did not inherit a PENNYH*

]R.

My conjectureis that this intonationpatternis inconsistentwith the hypotheticalcontext (atleastfor denialof expectation).Ontheotherhand,theintonationpatternof (15A2b) seemspossiblefor (2) for somecontext suchas(11),but is inconsistentwith theproposedIS.

In summary, lackof aconvincing context for theIS analysisin (2) is aproblemfor Kruijf f-Korbayova and Webber(2001). In addition, the proposedsemanticmotivation is not sufficient becauseit is neededindependentof IS. Oneadditionalquestionconcerningexample(2) is whetherthereis anotherlevel of IS connectingthetwo clausesleadingto a recursive IS, which is notdiscussedin their paper.

3 Recursive IS and Tripartite Structure

This sectiondiscussesthe recursive analysisof Partee(1996)asseenin (3). Hermotivationfor thismove is to mediatetwo analysesinvolving presupposition.The

3It would bepossibleto evaluatenative speakers’ intuition usingsynthesizedspeech/intonationapplyingtheideasdiscussedin Prevost(1995).

58

first is an analysisof Hajicova (1984), who arguesthat the distinction betweenpresuppositionand‘allegation’ (potentialpresupposition,seebelow) is affectedbyIS. The secondis that of Heim (1982), who adopts‘tripartite structure’,a typeof semanticstructurecloselyassociatedwith quantifierscopes,for heranalysisofpresupposition.

In orderto discusstheissueat hand,let usobserve anexamplefrom Hajicova(1984)(the possibility of the presuppositionrelevant to the currentdiscussionisalsoshown to theright of eachsentence).

(17) a. This time John’s COUSIN causedourvictory. $ Wewon.

Negation: This time John’s COUSIN didn’t causeour victory. $ Wewon.

b. This time John’s cousincausedour VICTORY. $ Wewon.

Negation: this time John’s cousindidn’t causeour VICTORY. % Wewon.

The proposition“we won” is a presuppositionof (a) but only an ‘allegation’ of(b) becauseits negationcannotentail theproposition.Hajicova’s argumentis thatwhenthe presupposition-triggering materialis in the theme(rheme),it resultsina presupposition(allegation). Partee(1996)attemptsto connectthis analysistoHeim’s analysisof presuppositionalongtherecursive tripartitestructure.Accord-ing to Partee,thepresupposition/allegation distinctioncanbeobservedrecursivelyjustasthetripartitestructureis.

But the presupposition/allegation distinction can be observed within a nounphraseaswell, asshown below.

(18) a. theRECORD of ourarrest$ Wewerearrested.

Negation:no RECORD of ourarrest$ Wewerearrested.

b. therecordof our ARREST $ Wewerearrested.

Negation:no recordof our ARREST % Wewerearrested.

Thus,the distinctionmustbe analyzedindependentof IS asin the previous sec-tion. As pointedoutby oneof thereviewers,thisdistinctioncanbemadein termsof the notionsof CB (contextually-bound)vs. NB (non-bound)(e.g.,Sgall et al.1986).This situationseemsto correspondto thepoint thattheanalysisof Kruijf f-Korbayova andWebber(2001) could be castwithin the foreground-backgrounddistinctionof Steedman(2000).

The samerecursive idea is discussedeven further in Hajicova et al. (1998).Althoughthey statethatan IS canbeassociatedwith eachembeddedclause,thispoint wasnever emphasizedin their earlierwork (e.g.,Sgallet al. 1986). In fact,Parteewrites that shelearnedthis point only after 1991 (Hajicova et al. 1998,p. 95). In addition,anotherpaperof Hajicova et al. (1995)on a computationalanalysisof IS completelyleavesout complex structures.This seemsto imply that

59

thenotionof recursive IS might not beasgeneralor essentialasarguedin Partee(1996)andHajicova etal. (1998)

Further, the growing trend in ‘interpreting’ IS is to adoptan informal, pro-ceduralview (Vallduvı 1990)or a formal, dynamicview (Steedman2000). Therecursive approachesto IS have not discussedthis aspectof IS analysisyet. Thesamecommentalsoappliesto anotherrecursive approachof Hoffman(1995).

As in thecaseof (2), it hardlyseemspossibleto comeup with a questionthatdirectly confirm suchan IS. Again, this doesnot rejectPartee’s proposal.But asbefore,themotivationcannotbethepresupposition/allegation distinctionbecauseit is independentof IS. From the discussionin this and the previous sections,Iwould like to presenttwo conjectures.First, thesemantic/pragmaticeffect thatcanbeobservedentirelywithin a nounphrasecannotbeanIS effect. Second,IS onlyprovidesa boundon thedomainof tripartitestructure,a muchweaker view of therelationbetweenIS andtripartitestructure.

4 Subordinatevs. Coordinate Structures

We begin this sectionwith potentialcounterexamplesto themainpoint of this pa-per. Then,I will discussexamplesin Japanese,German,andin Englishin supportof ourposition.

4.1 Potential Counterexamples

Althoughthewritten form of Englishis not very rich in markingIS, topicalizationandfocusmovementcanbeconsideredto mark IS (Prince1984). BonnieWeber[p.c.] pointsout thatthepresenceof oneof theseconstructionswithin analthough-clausecanbe a counterexampleto the proposedposition,andprovided with thefollowing examplesfrom theBritish NationalCorpus(BNC).

(19) a. The shapeseemedto be looking througha book, althoughwhat thebookwasHenrycouldnot tell. [ASS 676]

b. His motherwasalwaystelling him thatit wasimportantfor teacherstogive, althoughwhat they weresupposedto give shedid not say. [HR831]

c. Althoughwhat thatcouldpossiblybe,I have no idea,Melissathoughtto herselfassheputdown thephone.[GVP 1963]

First, I agreethat topicalization/focus movementis a weakform of IS mark-ers(Komagata1999),but will questionthestatusof thealthough-clausesobservedin theseexamples.Next, in mostof the potentialcounterexamplesincluding (19a � b), thealthough-clausefollows themain clause.Theonly exceptionknown tome is (19c). I would not provide a detailedexplanationfor this exampleat thispoint except for pointing out the following: the presenceof topicalization/focusmovementin the although-clausesuggeststhat thereis an IS division within the

60

although-clause,but thereis no further, direct evidenceof doubleIS in this ex-ample;I conjecturethat the themain clausein (19c) is an afterthought(andthusa part of the theme). The positionthat sentence-initialalthough-clausesarethe-matic is alsorelatedto thefollowing. While not categorical, thefirst componentsin an utterancetendsto be thematic,asHalliday (1967)pointedout. In addition,accordingto Quirk et al. (1985,p. 919),oneof thesemanticcharacterizationsofsubordinationis that thesubordinateclausepresentsinformationasif it is presup-posedasgivenratherthanassertedasnew. This descriptionseemsto applyto themajority of although-clauses,especiallywhenthesubordinateclauseprecedesthemainclause(57%of all theinstancesinvolving analthough-clausein theACLDCIcorpusfrom LDC). As a consequence,this paperwill still be left with onepoten-tial counterexample(but not a numberof counterexamplesascommentedby a re-viewer). Theremainingwiderangeof potentialcounterexampleswill beaccountedfor in a fairly systematicmannerin thefollowing subsections.Notethattheabovepotentialcounterexamplesdo notdirectlysupporttheparticularIS analysisin (2).

Thereareotherrelatedconstructions,whichareoccasionallymisunderstoodasIS markers.In particular, bothit-cleftsandpseudocleftsdonotnecessarilymarkanIS partition,esp. in embeddedenvironments.Accordingto thefindingsof Prince(1978),Collins (1991),andDelin (1995), it-clefts serve heterogeneousfunctionsof markingIS, contrastiveness,andreferentialstatus.Also dueto Prince(1978)and Collins (1991), the free relative part of a pseudocleftis either ‘evoked’ or‘inferrable’, which is analogousto thereferentialstatusof thedefiniteexpression.As definiteexpressionscanappearin themesandrhemesnotmarkinginformationstructure,pseudocleftcannotbeadirectIS marker.

4.2 Mor phology in Japanese

Next, we explore potentialcounterexamplesin Japanese.This languagehasanexplicit thememarker, i.e., theparticlewa (Kuno1973),which is usefulfor ana-lyzing informationstructure.A caveat is that the samemorphemealsoservesasthecontrastivenessmarker. However, thecontrastive caserequiresthat therebeaphonologicalprominencewithin thenounphrasethatis suffixedwith wa (summa-rized in Komagata1999). Thus,the instancesof wa suffixed to a non-prominentnounphrasecanbeconsideredthematic.

Oneof thepropertiesof thethematicwaobservedby severalJapaneselinguistsincludingKuno(1973)is thatthematicwa cannotoccurin theembeddedenviron-ment. This is in accordancewith our proposition. But Noda(1996)discussesaclassificationof subordinateclausesinto the following categories,which includepotentialcounterexamples.

(20)Type Examples ga waStrong (re)ba“ if ”, toki “when”, koto “ that” yes no

node(focal) “because”Weak node“since”, ga “but” yes yes

61

AlthoughNodasaysthatthethematicwacannotbeusedin strongsubordinateclauses,healsolists severalexamplesthatheconsidersexceptional.But thewa-marked subjectsin theseexamplesare either the subjectof the main clauseorcontrastive. Thus,they arenotexceptionsatall.

In addition,Nodalists several examplesof weaksubordinateclausesthat in-volve a thematicwa. Onecaseinvolves the conjunctive particlega “but” as thesentenceconnector. But theseexamplesarebetterclassifiedascoordinatestruc-tures.Theothercaseinvolvesa ‘subordinator’particle,but lacksthemainclauseasfollows:

(21) memorii-wa ... mottomo anzen-na basyo-dakara-da.memory-TOP most safe place-because-COP

“Becausethememoryis thesafestplace(for virusesto stay).”

ThesubjectNPsuffixedwith wa is consideredasthethemeof thisutterance.Eventhoughthis sentencecontainsa subordinator-like particle,it is an independentut-terance.Then,it is no surprisethat thereis an IS partition within it, andthus,itis not a counterexampleto thecurrentposition. In this case,thesubordinator-likeparticledakara “because” is betterconsideredadiscourseconnector.

4.3 Syntax in German

We next investigatethe obwohl “although”-clause in German,which is particu-larly interestingbecauseonly thematrixclauseexhibits theverb-second(V2) phe-nomenon.

While theobwhol-clausetypically hastheverb-finalpattern(i.e., subordinateclause),Gunthner(1996)observesthe growing tendency of theV2 configurationin theobwhol-clausein spokencolloquialGermanasshown below.

(22) A: DU ich brauchenkleinenSTIFT

“hey I needasmallpencil”

B: momentmal

“just asecond”

A: obwholNE eigntlichweißichsauchAUSwendigglaubich

“althoughno actuallyI do know it by heart”

Gunthnerarguesthatthis typeof obwhol-clausesform acoordinatestructureratherthana subordinatestructure.Then,obwholcanbeconsideredasa discoursecon-nector. Gunthneralsoanalyzestheconditionfor usingsubordinateobwhol-clausesas follows: the relation betweenthe main clauseand the obwhol-clauseis veryloose(or independentillocutionary force for the obwhol-clause). According toGunthner,the availability of the two patternsandthe conditionfor the weil “be-cause”-clauseareanalogousto theobwhol-clause.

62

¿Fromthe examplesin Japanese(previous subsection)andGerman,we mayinfer thefollowing. First, subordinate(-like) clausescanexist on their own (with-out the main clause)or weakly connectedto the main clause. Second,this typeof clausescanactuallybe consideredasutterancesandthusthey may containISpartitions.

4.4 Although-clauseasa Coordinate Structure

Theuseof subordinate(-like) clausesasanindependentutterancecanbeobservedin Englishaswell (Quirk etal. 1985,p. 564).For example,thewordbecauseherecanbe analyzedasa discourseconnectorbecauseit connectthe proposition“hedid it” with theresponseasthereasonfor theproposition.

(23) Q: Why did hedo it?

A: Becausehewasangry.

Similarly, thefollowing exampleseemspossible.

(24) A: I heardthatyouwentto theparkyesterday.

B: Althoughit wasraining.

This examplecanbeconsideredcompletelyin parallelto (23). Suchanexamplemightbefoundin spokencorpora,but I havenotbeenableto checkthispossibility.Thewordalthoughis aconcessive,discourseconnector. For this typeof utterance,it is naturalto consideranIS thatis felicitousto thecontext. While thedistinctionbetweencoordinateandsubordinatestructuresis not necessarilyclearin English(e.g.,Quirk et al. 1985,p. 927), the analogousdistinctionis clearly seenin theGermanexamplesbecauseof the V2 phenomenon.While IS-markingtoo is notnecessarilyclearin English,thisaspectis observedin theJapaneseexample.

Let us now turn to the casewherethe subordinateclausefollows the mainclausewith particularplacementsof phonologicalprominencesuchasthefollow-ing (Quirk etal. 1985,p. 1077).

(25) a. Ravendidn’t leave thepartyearlybecauseCAROL wasthere.

b. Ravendidn’t leave thepartyEARLY, becauseCAROL wasthere.

In fact, thesamepatternseemspossiblewith althoughaswell (replacingbecausewith althoughabove). While the scopeof the negation includesthe subordinateclausein (a), it is not the casein (b). This suggeststhat thereis a differencebe-tween(a) and(b) with respectto thestrengthof theconnection.Analogousto theobservation of Gunthner(1996)in German,we may considerthe entiresentence(a) asasingleutterance,but thesentence(b) possiblyasacoordinatestructurecon-sistingof two utterances.Althoughthisanalysisdependsonthedegreeof connect-ednessbetweenthetwo clauses,suchananalysiswouldbepossibleasin Gunthner(1996).

63

In summary, thepresenceof independentIS in thesubordinate(-like) clauseinthepotentialcounterexamplesis actuallynot inconsistentwith thecurrentposition,andthusis not consideredascounterexamples.I suggestthattheanalysisof com-plex structuresproposedhereis not specificto thealthoughandbecause-clausesbut applicableto subordinateclausesheadedby varioussubordinators.In addition,wecanmakearelatedpredictionbasedon theIS-relateddistinctionbetweensince(only thematic)andbecause(Quirk etal. 1985;Lambrecht1994):thesince-clausewouldnotgive riseto anindependentutteranceasthebecause-clausedoes.

5 Conclusion

Thispapersupportsthetraditionalview of IS asanon-recursive, matrix-level phe-nomenon,andarguesthat thealternative views arenot sufficiently motivatedandthatpotentialcounterexamplesto thetraditionalview canbeanalyzedin asystem-aticmanner.

Structurallyspeaking,the domainof IS basedon the traditionalview is notfixedwith respectto theDS unit (i.e., clause).As a consequence,IS appearsen-tangledaroundclauses.Nevertheless,this typeof entanglementseemsinevitableto maintainthe simpleview of IS-DS relation (6) in the spirit of Danes (1974).Oneimplication of the currentpositionwith respectto NLP systemdesignis asfollows: dueto the entanglementof IS’s with clauses,we cannotarrangeIS andDS processingsequentially. In orderto dealwith complex sentences,the IS andDS processingmustproceedin parallelat somepoint.

In this paper, we discussexamplesin English, German,and Japanese.Al-thoughthe dataarefairly consistentin my view, examplesfrom otherlanguagesmayrevealmoreabouttheissuesunderinvestigation.Thus,it wouldbevery inter-estingto comparea largernumberof languagesalongthecurrentline.

Acknowledgements

I would like to thankBonnieWebberfor discussionandpotentialcounterexam-ples;Mark Steedmanfor discussionaboutrelevant topicsat variouspoints;ClaireGardent,Aravind Joshi,andMarthaPalmerfor commentson my dissertation,onwhichpartsof thispaperis based;andthereviewersof theworkshopfor stimulat-ing comments.

Bibliography

Collins, P. C. (1991). Cleft and pseudo-cleftconstructionsin English. Rout-ledge.

Danes, F. (1974).Functionalsentenceperspective andthe organizationof thetext. In F. Danes (Ed.),Papers on functionalsentenceperspective, pp.106–128.Prague:AcademiaandTheHague:Mouton.

64

Delin, J. (1995).Presuppositionandsharedknowledgein it-clefts. LanguageandCognitiveProcess10(2), 97–120.

Firbas,J.(1964).Ondefiningthethemein functionalsentenceanalysis.TravauxLinguistiquesdePrague1, 267–280.

Grosz,B. J.andC. L. Sidner(1986).Attention,intentions,andthestructureofdiscourse.ComputationalLinguistics12(3), 175–204.

Gunthner, S.(1996).Fromsubordinationto coordination?verb-secondpositionin Germancausalandconcessive constructions.Pragmatics6(3), 323–356.

Hajicova, E. (1984).Presuppositionandallegationrevisited.Journal of Prag-matics8, 155–167.

Hajicova,E.,B. H. Partee,andP. Sgall(1998).Topic-FocusArticulation,TripartStructures,andSemanticContent. Kluwer.

Hajicova, E., H. Skoumalova, andP. Sgall (1995).An automaticprocedurefortopic-focusidentification.ComputationalLinguistics21(1), 81–94.

Halliday, M. A. K. (1967).Noteson transitivity andthemein English(PartII).Journal of Linguistics3, 199–244.

Heim, I. (1982).TheSemanticsof DefiniteandIndefiniteNounPhrases. Ph.D.thesis,Universityof MassachusettsatAmherst.

Hoffman,B. (1995).TheComputationalAnalysisof theSyntaxandInterpreta-tion of “F ree” Word Order in Turkish. Ph.D. thesis,Universityof Pennsyl-vania.

Komagata,N. N. (1999).A ComputationalAnalysisof Information StructureUsingParallel ExpositoryTextsin EnglishandJapanese. Ph.D. thesis,Uni-versityof Pennsylvania.

Kruijf f-Korbayova, I. andB. L. Webber(2001).Concession,implicatureandalternative sets.In Proceedingsof theInternationalWorkshopon Computa-tional Semantics(IWCS-4),Tilburg, Jan.,2001.

Kuno,S.(1973).TheStructure of theJapaneseLanguage. MIT Press.

Lambrecht,K. (1994).InformationStructure andSentenceForm: Topic, focus,andthementalrepresentationsof discoursereferents. CambridgeUniversityPress.

Mathesius,V. (1975).A FunctionalAnalysisof PresentDay Englishon a Gen-eral LinguisticBasis,editedbyJosefVachek. TheHague:Mouton.

Noda,H. (1996).WA-toGA (WA andGA). Kuroshio.

Partee,B. H. (1996).Allegationandlocal accommodation.In B. H. ParteeandP. Sgall (Eds.),Discourseandmeaning:papers in honorof EvaHajicova,pp.65–86.JohnBenjamins.

65

Prevost,S.(1995).A Semanticsof ContrastandInformationStructure for Spec-ifying Intonationin SpokenLanguage Generation. Ph.D. thesis,Universityof Pennsylvania.

Prince,E. F. (1978).A comparisonof wh-cleftsandit-clefts in discourse.Lan-guage 54(4), 883–906.

Prince,E. F. (1984).Topicalizationandleft-dislocation:A functionalanalysis.In S.J.WhiteandV. Teller(Eds.),Annalsof New York Academyof Sciences,Vol. 433: Discoursesin ReadingandLinguistics, pp. 213–225.New York:TheNew York Academyof Sciences.

Quirk, R., S. Greenbaum,G. Leech,andJ. Svartvik (1985).A ComprehensiveGrammarof theEnglishLanguage. Longman.

Rooth,M. E. (1985).Associationwith Focus. Ph.D. thesis,Universityof Mas-sachusettsatAmherst.

Sgall,P., E. Hajicova, andJ. Panevova (1986).Themeaningof thesentenceinits semanticandpragmaticaspects. D. Reidel.

Steedman,M. (2000).Informationstructureandthesyntax-phonologyinterface.LinguisticInquiry 31(4), 649–689.

Vallduvı, E. (1990).Theinformationalcomponent. Ph.D. thesis,UniversityofPennsylvania.

66

Inf ormation Structur eand the Inter pretation of“otherwise”

IVANA KRUIJFF-KORBAYOVA

Computerlinguistik,Universitat desSaarlandes,Saarbrucken,Germany

[email protected]

BONNIE LYNN WEBBER

Informatics,Universityof Edinburgh,80SouthBridge,Edinburgh EH11HW, U.K.

[email protected]

ABSTRACT. Wehavebeeninvestigatingwhetherandhow theinterpretationof discourseconnec-tivesis sensitive to theInformationStructure(IS) of theclausesor sentencesthey relate.Herewefocusontheanaphoricconnective “otherwise”andshow how theIS of its antecedentaffectswhatconditionit canbe“otherwise”to. Thiswork is partof a largerenterpriseaimedatunderstandingwhatrole(s)sentence-level IS playsin theinterpretationof largerunitsof discourse.

1 Intr oduction

It is well-known thatInformationStructure (IS) influencestheinterpretationof in-dividualsentences.Of thefamoussignin theLondonUnderground,“Dogsmustbecarried”,Halliday (1970)observesthat this text canbepronouncedwith differentintonationpatterns,e.g., (1) vs. (2) reflectingdifferent IS. Thereby, different in-structions(here,paraphrasedin italics) areconveyedto passengers.Onesupposesthat(2) wasnot theintentionof theLondonTransportAuthority.

(1) DogsmustbeCARRIED.H* LL%

If there is a dog, carry it.

(2) DOGS mustbecarried.H* LL%

Carry a dog.

In English,IS is mostoften conveyed by intonation. In languageswith freerword order, differencesin IS aremostoftenconveyedby differentword ordering.Forexample,theCzechcounterpartsof (1)and(2),conveying thesameinstructionsto thehearer, are(3) and(4), respectively:

67


(3) PsiDogsnom

serefl

musımust3pl

NEST.carryinf

(4) MusıMust3pl

serefl

nestcarryinf

PSI.dogsnom

Over the pastdecade,the understandingof IS within the sentencehasbeenenrichedby intensive researchin formal semantics. It is now widely acceptedthat IS affectsbothinterpretationandrealization,eventhoughthereis no uniformaccount.However, muchlessis known aboutwhat,if any, useis madeof IS beyondclauseandsentenceboundariesandhow IS interactswith otheraspectsof discoursestructureandsemantics.Ourwork extendstherepertoireof IS-sensitiveaccountsinthisdirection.In thispaper, weconcentrateonhow theIS of aprevioussentenceorclausecanaffect themeaningprojectedthroughthesubsequentadverbialdiscourseconnective “otherwise” (“jinak”, in Czech). We show that an IS-basedaccountof its meaningprovidesaccessto contextually appropriateinterpretationsthatareunavailableto accountsthatignoreIS.

Webberet al. (1999)have arguedthat “otherwise”contributesmeaningto thediscoursein partthroughstructure,in partthroughanaphora:roughly, they saythatthecomplementof theanaphorically-derived argumentof “otherwise”servesasaconditionunderwhich the interpretationof its structuralmatrix holds.1 As mightbeexpected,differentwaysof resolvingtheanaphoricargumentleadto differentinterpretations,asin (5a)vs. (5b):

(5) If youhave broughtadog,youmustpay50p.

a. Otherwiseyouwill notbeallowedto enter.

b. Otherwiseyoucancomein for free.

whichcanbeparaphrasedby resolvingtheanaphorandmakingtheanaphorically-derivedconditionexplicit:

(6) a. If you havebroughta dog and you do not pay 50p, you will not beallowedto enter.

b. If youhavenotbroughta dog, youcancomein for free.

Here,theantecedentusedin (6a) is theprecedingmainclause,while that for (6b)is thepreceding“if ”-clause.2

1(Webberet al. 2001) presentasevidencefor this, inter alia, the fact that the first argumentof “otherwise” may not be explicit, but ratherhave to be derived by inferencefrom the previousdiscourse,andthefactthat it canbehave like a ‘donkey’ pronoun,deriving its first argumentfrom arelative clause— e.g.,“Farmerswho beattheirdonkeys would otherwisebebeatingtheir wives.”

2As with anaphoricpronouns,anautomatedprocedurefor resolvinganaphoric“otherwise”mustbeableto rejectcontextually inappropriatewaysof resolvingit asin

(6a& ) If youhavenotbroughta dog, youwill notbeallowedto enter.

(6b& ) If youhavebroughta dog andyoudo notpay50p,youcancomein for free.

But wedo notconsiderthisaspectof theproblemany furtherin thispaper.

68

That IS canaffect whatconditionscanbederivedcanbeseenby considering“otherwise” in the context of two differentsingle-clauseutterances,which differonly in their IS — here(7) vs. (8) in bothEnglishandCzech.3

(7) YoumustCARRY adog. Otherwiseyoumight getHURT.H* LL% H*LL%

PsaDogacc

musıtemust2pl

NEST.carry.

JinakOtherwise

bystebe2pl

mohlicouldpl

prij ıtcome

kto

urazu.injurydat.

(8) Youmustcarrya DOG. Otherwiseyoumight getHURT.H*LL% H*LL%

MusıteMust2pl

nestcarry

PSA.dogacc.

JinakOtherwise

bystebe2pl

mohlicouldpl

prij ıtcome

kto

urazu.injurydat.

The“otherwise”clausein (7) will be interpretedaswarningthehearer(H) thatHmightgethurt if s/hehasadogbut isn’t carryingit (e.g.,H mightgettangledup inthedog’s lead).On theotherhand,the“otherwise”clausein (8) warnsH thats/hemightgethurt if notcarryingadog,period(e.g.,H mightbewalkingpastfanaticalmembersof theRoyal KennelClub).

If the IS of onesentenceor clausecanaffect how anotheris interpreted,thenIS mustbe incorporatedinto anaccountof discourseinterpretationanddiscourseupdating.Thiswedo in termsof Rooth’snotionof analternativeset(Rooth1985;Rooth1992)andthealternative-setsemanticsof informationstructureworkedoutin (Steedman2000a;Steedman2000b),andrefining our earlierpresentationsin(Kruijf f-Korbayova andWebber2000a;Kruijf f-Korbayova andWebber2000b).

Thepaperis organizedasfollows: In Section2 we presenttheapproachto ISandIS-sensitive context updatingwe areemploying. In Section3 we describeourIS-sensitiveanalysisof α otherwiseβ whereα isasimplesentence.In Section4wedescribeour IS-sensitive analysisof α otherwiseβ whereα is acomplex sentence,makingmoreoptionsavailable. Section5 concludesthepaperanddelineatesthefuturedirectionsof this work.

2 Inf ormation Structure and Context Updating

The notion of IS we areemploying originatesin the work of Mathesius(1975),andhasbeenelaboratedin subsequentwork within thePragueSchool(Sgallet al.1986)andby others,e.g.,(Firbas1992,Halliday1985,Steedman2000b).Specifi-cally, weadopttheformalaccountpresentedin (Steedman1996;Steedman2000a;Steedman2000b)which (1) providesa well worked out compositionalsemanticsof Englishintonationin IS terms;(2) interpretstheelementsof IS in termsof al-ternative sets,and(3) assumesa generalIS-sentencenotion of discoursecontextupdate. Leaving terminologicaldifferencesaside,Steedman’s accountis by and

3Throughoutthe paper, SMALL CAPITALS indicateintonationcenters(pitch accents),therebydistinguishingFocusis from Backgroundis within bothThemeis andRhemeis.

69

large straight-forwardly compatiblewith the PragueSchoolapproach,and thuswhenanalyzingCzechexamples,we cancombineSteedman’s accountwith SgallandHajicova’s ideasrelatingIS andword order(Hajicova andSgall 1987;Sgalletal. 1986).

Building onthefindingsoriginatingin thePragueSchool(Firbas1992;Mathe-sius1975;Sgalletal. 1986),Steedmanrecognizestwo dimensionsof IS: Thefirstdefinesapartitioningat thesentence-level into Themeis andRhemeis; thesecondisa furtherpartitioningof both into Backgroundis andFocusis.4 Thelatterpartition-ing is relatedto Halliday’s Given-New dichotomy(Halliday 1970;Halliday 1985)andconcernsdistinguishingthe Themeis andtheRhemeis from otheralternativesthatthecontext makesavailable.

In English,Czechandmany other languages,IS is establishedasa resultofan interplayof intonation,word orderandgrammaticalstructure.Below we givethreeof thepossibleIS partitionsinto Themeis-Rhemeis thatSteedman’s approachprovidesfor thestring“You shouldcarrythedog”.5 Thesituationis onein whichthefirst author(IKK) is transportingadog,a largebaganda trolley by theUnder-ground,andasksthesecondauthor(BW) a question,which helpsto fix the IS ofthereply.

(9) Q: How shouldI transporttheDOG?

A: Youshould' (�) *Theme

CARRY

H*' (�) *Rheme

theDOG.L+H*' (�) *

Theme

i. θ(9): λQ + Q � h �-, dog1 � ρ(9): λx + λy+�, carry � x � y�ii. θ-AS(9): ./� Q + Q � h � dog1 � , � Q + Q � h � bag3 � , � Q + Q � h � trolley4 �10

ρ-AS(9): . lead� h � dog1 � , carry � h � dog1 � , wheel� h � dog1 �10(10) Q: Who shouldcarrytheDOG?

A: YOU

H* L' (�) *Rheme

shouldcarrytheDOG.LH%' (�) *

Theme

i. θ(10): λx + carry � x � , dog1 � ρ(10): λQ + Q � , h�ii. θ-AS(10): ./� Q + Q � dog1 � , � Q + Q � bag3 � , � Q + Q � trolley4 �10

ρ-AS(10): . carry � h � dog1 � , carry � s� dog1 � , carry � of f icer5 � dog1 �10

4Alternative termsusedfor similar (but not identical)IS partitionsin otherworksare,e.g.,Topic-Focus(Sgall et al. 1986),Background(=Link+Tail)-Focus(Vallduvı 1992). We adoptSteedman’sterms,but addthesubscriptsin Themeis, Rhemeis andBackgroundis, Focusis in orderto avoid con-fusionwith someotherusesof thesameterms.

5For thetimebeing,weignorethemodalityintroducedby “should”andany aspectsof thespeech-actbeyondsimpleassertion.

70

(11) Q: WhatshouldI CARRY?

A: YoushouldCARRY

L+H* LH%' (�) *Theme

theDOG.H*LL%' (�) *Rheme

i. θ(11): λx + , carry � h � x� ρ(11): λQ + Q � , dog1 �ii. θ-AS(11): ./� x + wheel� h � x� , � x + push� h � x� , � x + carry � h � x�10

ρ-AS(11): . carry � h � bag3 �2� carry � h � dog1 �10For eachsentence,(i) provides a simplified IS-partitionedlogical form, whereθ andρ areoperatorswhich ‘wrap’ Themeis andRhemeis, respectively. WithinThemeis andRhemeis, asteriskson terms(e.g., 3 carry) indicateelementsthatbe-long to the respective Focusis. TheseIS-partitionedlogical forms representthelinguisticmeaningof thesentences,andserveasinput for adiscourse(context) up-datefunctiondescribedbelow. (ii) indicatestheThemeis alternative set(θ-AS) andRhemeis alternative set(ρ-AS), which areexplainedbelow. Becauseeachexam-ple containsFocusis within Themeis (indicatedby a , -term),whichentailscontrastwith a previousThemeis (andhencealternativesto contrastwith), eachθ-AS con-tainsmorethanoneelement.(Without pitch accentsin Themeis, andthuswithoutcontrast,theθ-AS wouldbeasingletonset.)

2.1 Alter nativeSetSemanticsfor IS

Elaboratingon Rooth’s alternative semantics(Rooth1992),Steedmanassignsthefollowing semanticsto IS (cf. Steedman2000a):4 Themeis presupposesaRhemeis-alternativeset(ρ-AS).4 Focusis within Rhemeis restrictstheρ-AS to thesingletonsetcorresponding

to theassertedproposition.4 Themeis alsopresupposesaTheme-alternativeset(θ-AS).4 Focusis within Themeis restrictstheθ-AS to thesingletonsetcorrespondingto Themeis.

ρ-AS correspondsto whatRoothcalls thecontextual alternativeset(Rooth1985;Rooth1992).θ-AS is asetof alternative themeswith respectto thecontext, corre-spondingto whatRoothcallsthequestionalternativeset. Thenotionof alternativeset is also closely relatedto the notion of secondarydenotation(KarttunenandPeters1979).

Following (Steedman2000a),we take ρ-AS to bea subsetof thepropositionssupportedby thecontext, whosecharacteristicfunctionis obtainedsystematicallyfrom theIS-partitionedlogical form. As notedin (Steedman2000a,p.10),alterna-tivesetsmaynotbeexhaustively known to hearers,andin practiceonewouldwantto computewith amoreabstractform.

71

c1 c2 c3θ 5 ψ 6 ρ 5 ψ 6

Figure10.1: IS-sensitive updateof context c1 with ψ: c1 7 θ � ψ �98 c2 7 ρ � ψ �98 c3

2.2 IS-sensitiveContext Updating

Wefollow (Krifka 1993;Kruijf f-Korbayova1998;Steedman2000a)in definingtheupdatingof an input context c1 with an IS-partitionedlogical form p ascompris-ing two phases,a Themeis updatephase(c1 7 θ � ψ �98 c2) anda Rhemeis updatephase(c2 7 ρ � ψ �98 c3), wherec2 andc3 areresultingcontexts. (SeeFigure2.2).

In the Themeis updatephase, the input context c1 is checked asto whetheritsupportsor canaccommodatethepresuppositionsof thethemeθ � ψ � – namely, theThemeis-alternative setθ-AS andthe Rhemeis-alternative setρ-AS. This yieldsarestrictedcontext c2 whereθ � ψ � holds. In theRhemeis updatephase, onealterna-tive accordingto theρ-AS is selected,which yieldsthefinal context c3. Updatingfails if eitherupdatephasedoes.

3 IS and “otherwise”: single-clauseantecedents

As notedearlier, Webberet al. (1999)have arguedthat“otherwise”hasoneargu-mentestablishedanaphorically, andoneprovided structurally. It is theanaphoricargumentthat provides the condition that “otherwise” appealsto andwhoseIS,we arearguing, the interpretationof “otherwise” mustbe sensitive to. This doesnot mean,however, that theantecedentof “otherwise” is limited to IS-partitionedutterances:just thatIS-partitioningprovidesrelevantpossibilities.

Becauseexample(5), given earlier, containstwo differentclauses(main andsubordinate)thatcanserveasantecedentsfor “otherwise”,thereareat leastthetwopossibleconditions— shown in (6a)and(6b) — thatotherwiseβ canderive andapplyto theinterpretationof β. Examplessuchasthis arediscussedin Section4.Herewe focuson caseswherethe condition that “otherwise” appealsto derivesfrom asingleclauseantecedent.

Evenhere,theanalysisin (Webberet al. 1999)mustberefinedin two waystotake accountof IS:

1. The antecedent(A) of “otherwise” shouldnot be treatedasan atomicunit:rather, “otherwise” canappealto a condition“C” deriving from eitherA’sThemeis or its Rhemeis.

2. The context that β is assertedwith respectto is not strictly worlds consis-tentwith therealworld or thecurrentdiscoursecontext otherthanthose“C”worlds: rather, it may or may not be consistentwith the Themeis of its an-tecedentA aswell.

72

(In thefollowing examples,“Otherwiseβ” itself hasanIS-partitioning.However,we do not explicitly indicate it, becauseit is not relevant to the points we areadvancing. We will make a point aboutthe IS-statusof “otherwise” itself at theendof this section.)

The examplesbelow addressthe first point, showing that the condition that“otherwise”appealsto mayderiveeitherfrom theThemeis of its antecedent— wecall this thefull Themeis-complementcondition– asin (12i), or from its Rhemeis —we call this the full Rhemeis-complementcondition– asin (12ii). (Thecorrespond-ing paraphrasesof the“otherwiseβ” areshown in italics.)

(12) Q. WhatshouldI doat a RED LIGHT?

i. At a redlight,' (�) *Theme

STOP.H*LL%' (�) *Rheme

Otherwiseyoucango straighton.

NaAt

cervenouredacc

zastavte.stopimp2pl

JinakOtherwise

muzetecan2pl

jetgoinf

rovne.straight.

If thelight is not red,youcango straight on.

ii. At a redlight,' (�) *Theme


Otherwiseyouwill geta ticket.

NaAt

cervenouredacc

zastavte.stopimp2pl

JinakOtherwise

dostaneteget2pl

pokutu.fineacc

If (thelight is redand)youdo not stop,youwill get a ticket.

Thefull Themeis-complementconditionin (12i) correspondsto “not beingata redlight” (or, “the light not beingred”). The full Rhemeis-complementconditionin(12ii) correspondsto “not stopping”.

As to our secondpoint, the context in which β is asserted,at issueis theThemeis of the antecedentof “otherwise”. When“otherwise” appealsto the fullThemeis-complementcondition,thereis only onecontext with respectto which βcanbeasserted,namelytheinitial context beforeassertingα. Thecontext updatedwith α’s Themeis is irrelevant, becauseit is incompatiblewith the full Themeis-complementcondition: example(12i) cannotbe interpretedas If the light is redandif thelight is not red,youcango straight on.

On theotherhand,when“otherwise”appealsto thefull Rhemeis-complementcondition, it appearsthat the IS-partitioningmakes two contexts available withrespectto whichβ canbeasserted:theinitial context beforeassertingα’sThemeis(asin (13i)), andthecontext updatedwith α’s Themeis (asin (13ii)).

(13) Q. WhenshouldI STOP?

i. Stop' (�) *Theme

ata redLIGHT.H* LL%' (�) *

Rheme


73

ZastavteStopimp2pl

naat

cervenou.redacc.

JinakOtherwise

muzetecan2pl

jetgoinf

rovne.straight.

If thelight is notred(i.e., in otherconditionsthanbeingat a redlight),youcango straight on.

ii. Stop' (�) *Theme


Rheme

Otherwiseyoumight getrear-ended.

ZastavteStopimp2pl

naat

cervenou.redacc.

JinakOtherwise

bybe-aux3sg

dointo

vasyougen

nekdosomebodynom

mohlmightsg

narazit.bumpinf

If youstopandthelight is not red,youmightget rear-ended.

Thesensethat(13ii) conveysbothin Englishandin Czechthatoneshouldonlystopat a red light, comesfrom this interpretationof “otherwise” in termsof stoppingunderall conditionsotherthanthelight beingred.

Thereis onefurtherpoint to make beforepresentingour analysisof this phe-nomenonandreview of theseexamplesin moredetail. That is that “otherwise”itself is a contrastive (part of the) Themeis, andwhat we have seenherearedif-ferentwaysin which it relatesto theinput context: in example(12i), “otherwise”contrastswith the precedingThemeis (and thereforepicks up the full Themeis-complementcondition),while in examples(12ii), (13i) and(13ii), it contrastswiththeprecedingRhemeis (andthereforepicksup thefull Rhemeis-complementcon-dition). Example(12i: ) below illustratesthis Themeis contrastevenmorevividly,in thatthepitchaccentson“red” asFocusis within theThemeis of thefirst sentenceindicatesthespeaker’s awarenessof alternativesthatthe“otherwise”sentencethenexplicates.

(12i: ) At a RED light,L+H* LH%' (�) *

Theme


Otherwiseyoucancontinue.

3.1 Analysis

Weproposethefollowing IS-sensitive refinementof theanalysisof “otherwise”in(Webberetal. 1999):Let usassumethatα is theantecedentof otherwiseβ, andc0

is thecontext prior to updatingwith α (ratherthantherealworld). TheIS-sensitiveupdateenablesusto distinguishbetweenthefollowing subsetsof c0:4 thesubsetwhereα’s Themeis andalternativesto α’s Rhemeis hold (i.e., ex-

cludingα itself);4 thesubsetwherealternativesto α’s Themeis hold;4 thesubsetwherealternativesto α’s Rhemeis hold (irrespective of Themeis).

74

................

............................. ...........

.................

....................................................................

θ 5 β 6 ρ 5 β 6c0 c1 c2

c3 c4 c5

θ 5 α 6 ρ 5 α 6

c;3 c;4 c;5θ 5 β 6 ρ 5 β 6c; ;4 c; ;5c; ;3 θ 5 β 6 ρ 5 β 6

θ 5 α 6 ρ 5 α 6ρ 5 α 6

Figure10.2: IS-sensitive updatingwith “α + Otherwiseβ”.The dottedarcsindicatethe two possiblewaysof resolving“otherwise”with re-spectto thesimplestIS-partitioningof theantecedent,andthedashedarcsindicatethetransitionsto thecorrespondingcontexts.

In (Webberet al. 1999),β is assertedsolelywith respectto c0 � cα, thesubsetofc0 wherealternativesto α hold. Herewe refinethiswith respectto theabove threesubsets,defininghow aninputcontext is updatedwith thesequence“α. Otherwiseβ” for asingleclauseα:

1. c0 is updatedwith α asdescribedin Section2.2: c0 7 θ � α �98 c1 7 ρ � α �98 c2.

2. Updatingwith “otherwiseβ” involveseither:4 c1 beingupdatedwith “otherwiseβ”, which involvesconstructingcon-text c3 asthe Rhemeis-complementof c2 with respectto c1 andthenupdatingthiscontext with β: c1 7 ρ � α �<8 c3 7 θ � β �98 c4 7 ρ � β �98 c54 c0 beingupdatedwith “otherwiseβ” in oneof two ways:

– Context c3 : is constructedastheThemeis-complementof c1 withrespectto c0 andthenc3 : is updatedwith β:c0 7 θ � α �<8 c:3 7 θ � β �98 c:4 7 ρ � β �98 c:5

– Context c3 : : is constructedastheRhemeis-complementof c1 withrespectto c0 andthenc3 : : is updatedwith β:c0 7 ρ � α �=8 c: :3 7 θ � β �98 c:>:4 7 ρ � β �98 c:>:5

Thesecontext-updatingpossibilitiesareshown schematicallyin Figure10.2.

3.2 Examples

Wenow demonstratethisdetailedIS-sensitiveupdatinganalysisfor examplesfromthe introductionto this section. Example(12i) repreatedin (14) shows how theanalysisappliesto thecasewherea full Themeis-complementconditionis derivedfrom theThemeis of theantecedentof “otherwise”andβ is assertedwith respecttotheinitial context, c0. (Recallthatthis is theonlycontext-updatingpossibility.)

75

(14) At a redlight,' (�) *Theme



c0 7 λP+ at � h � red l ight �@? P8 c1 7 stop � h�98 c2

c0 7 at � h � red l ight �<8 c:3 7 λQ +Q � h�98 c:4 7 go straight � h�98 c:5If youare notat a redlight, youcangostraight on.

Example(13i) repeatedin (15) shows how the analysisappliesto the casewherea full Rhemeis-complementcondition is derived from the Rhemeis of theantecedent.As shown above, therearetwo possiblecontexts againstwhich β canbeasserted.In (15),β is assertedwith respectto theinitial context, i.e. c0.

(15) Stop' (�) *Theme


Rheme

Otherwiseyoucangostraighton.

c0 7 λP+ P ? stop � h�98 c1 7 at � h � red l ight �98 c2

c0 7 at � h � red l ight �<8 c3 7 λQ +Q � h�98 c4 7 go straight � h�98 c5

If thelight is notred(in otherconditionsthanbeingat a redlight), youcangostraight on.

In contrastwith (13i) is example(13ii), repeatedin (16). While it appealstothe full Rhemeis-complementcondition,β is assertedwith respectto the contextupdatedwith α’s Themeis, i.e. context c1.

(16) Stop' (�) *Theme

at a redLIGHT.H* LL%' (�) *

Rheme

Otherwiseyoumight getrear-ended.

c0 7 λP+ P ? stop � h�98 c1 7 at � h � red l ight �98 c2

c1 7 at � h � red l ight �<8 c3 7 λQ +Q � h�98 c4 7 get rear ended� h�98 c5

If youstopandthelight is not red,youmightget rear-ended.

The examplesin this sectiondemonstratea rangeof possibleantecedentsfor“otherwise”thatarenotavailablewithout takingIS into account.

4 IS and “otherwise”: complex-clauseantecedents

We now turn to examplesof the form consideredin (Webberet al. 1999),wheretheconditionusedfor interpreting“otherwise”comesfrom acomplex sentenceofthe form If φ, thenψ. Herewe show that thesameanalysisholdsasbefore,withoneaddition:

76

4 Whenboththesubordinateclauseandsomeelement(s)from themainclauseareincludedin theThemeis, a third possibility for deriving theconditiontowhich “otherwise” appealsis madeavailable: the conditioncanbe derivedfrom apartof theThemeis of theantecedent.

Therearesomewhatmoreexamplesto review, becausein bothEnglishandCzech,the main clausecan belongentirely to the Themeis or to the Rhemeis (i.e., theboundarybetweenThemeis andRhemeis cancoincidewith the clauseboundarybetweenφ and ψ), or the main clausecanbe divided over the Themeis and theRhemeis (i.e., theboundarybetweenThemeis andRhemeis splits ψ). Thefirst ofthesepossibilitiesis discussedin Section4.1,thesecondin Section4.2.

4.1 IS-boundary coinciding with clauseboundary

Whenthe IS-boundarybetweenThemeis andRhemeis coincideswith the clauseboundarybetweenφ andψ, theThemeis (Rhemeis) consistsof φ, andtheRhemeis(Themeis) of ψ. Theexamplesbelow show that,aswith simpleclauseantecedents,the conditionthat “otherwise” appealsto may derive either from the Themeis ofits antecedent(the full Themeis-complementcondition, as in (17i)), or from itsRhemeis (thefull Rhemeis-complementcondition, asin (17ii)).

(17) Q. WhatshouldI do if thelight is RED?

A. If thelight is RED,L+H*LH%' (�) *

Theme

stopat theCORNER.H*LL%' (�) *

Rheme

A : . Stopat theCORNER

H*LL%' (�) *Rheme

if thelight is RED.L+H*LH%' (�) *

Theme

i. Otherwiseyoucangostraighton.

If thelight is not red,go straight on.

ii. Otherwiseyouwill geta ticket.

If thelight is redandyoudonotstopat thecorner, youwill geta ticket.

Thefull Themeis-complementconditionin (17i) correspondsto “the light not be-ing red”, andthefull Rhemeis-complementconditionin (17ii) correspondsto “notstopping”(aswith thesimpleantecedentexamplesin (12i) and(12ii)).

When“otherwise”appealsto thefull Themeis-complementcondition,thereisonly onecontext with respectto whichβ canbeasserted,namelytheinitial contextbeforeassertingα. Thecontext updatedwith α’s Themeis is irrelevant,becauseitis incompatiblewith thefull Themeis-complementcondition:example(17i) cannotbeinterpretedasIf thelight is redandif thelight is notred,youcangostraighton.

On theotherhand,when“otherwise”appealsto thefull Rhemeis-complementcondition, it appearsthat the IS-partitioningmakes two contexts available with

77

respectto whichβ canbeasserted:theinitial context beforeassertingα’sThemeis(asin (18i)), andthecontext updatedwith α’s Themeis (asin (18ii)).

(18) Q. When(i.e.,underwhatconditions)shouldI stopat theCORNER?

A. Stopat theCORNER

L+H*LH%' (�) *Theme

if thelight is RED.H*LL%' (�) *

Rheme

A : . If thelight is RED

H*LL%' (�) *Rheme

stopat theCORNER.L+H*LH%' (�) *

Theme

i. Otherwiseyoucangostraighton.

If thelight is not red,(youneedn’t stop)andyoucango straight

ii. Otherwiseyoumight getrear-ended.

If you stopat thecorner and the light is not red, you might get rear-ended.

Again, thesensethat(18ii) conveys thatoneshouldonly stopat a redlight, comesfrom this interpretationof “otherwise” in termsof stoppingunderall conditionsotherthanthelight beingred. It appearsverydifficult to getthevariantof (18)withthepreposedrhematic“if ”-clause.Wethink thatthisis becausethisIS-partitioningrequiresa markedintonationpatternthatmaybedifficult in English.

4.2 IS-boundary splitting the main clause

When the IS-boundarybetweenThemeis and Rhemeis “splits” ψ, the Themeis(Rhemeis) consistsof φ andapartof ψ, while therestof ψ belongsto theRhemeis(Themeis). As before,“otherwise”canappealto the full Themeis-complementcon-dition andthe full Rhemeis-complementcondition, but anotherpossibility is thatthe conditionderives from just that part of the Themeis in the matrix clause,asillustratedbelow. This wecall thepartial Themeis-complementcondition.

(19) Q. Wheredoyou buy wine if it’s SUNDAY?

A. If it’s SUNDAY , webuy wineL+H*LH%' (�) *

Theme

over theSTATE LINE.H*LL%' (�) *

RhemeOtherwisewe justbuy beer.

a. If wedon’t buy wine, webuy beer.

b. If it is Sundayandwedon’t buy wine, webuy beer.

ThepartialThemeis-complementconditionin (19) correspondsto “we do not buywine”. The reasonwe give two possibleparaphrasesof “otherwisewe just buybeer” (otherwise β) is that the initial contex c0 can be updatedwith this par-tial Themeis-complement(γ) in eitherof two ways(shown schematicallyin Fig-ure10.3):

78

θ 5 β 6 ρ 5 β 6c5; ; c6; ;

........

.........

............

.................

........................................................................................................

ρ 5 β 6θ 5 β 6

c4; ;

c5;

...........

.....................

....................................................

c4; c6;

c0 c1ρ 5 α 6 : ψρθ 5 α 6 : ψθθ 5 α 6 : φ

ψθ

ψθ

c2cφ

(i)

(ii)

Figure 10.3: IS-sensitive updating with “α + Otherwiseβ” involving a partialThemeis-complementcondition.

4 It canbeupdatedjustwith γ, assertingβ with respectto theresult,asin (19a);4 It canbejustupdatedwith thatpartof α’sThemeis in the“if-clause” andthenγ, assertingβ with respectto theresult,asin (19b).

c0 cannotbeupdatedwith respectto theentireThemeis andthenγ becausethetwoareincompatible.

Finally, we considerthecasewherethe“if-clause” belongsto theRhemeis ofα, asin example(20a).Of concernis thepossibilitythat“otherwise”appealsto aconditionderivedfrom thatpartof theRhemeis in thematrix clause,whatwe callthepartial Rhemeis-complementcondition.

(20) Q. WhatshouldI do AFTER 5PM?

A. After 5pmLH%' (�) *

Theme

take a BREAK, if youareTIRED.H* H*LL%' (�) *

Rheme

a. Otherwise,you’ll startmakingmistakes.

If it is after 5pm,and if you are tired, and you do not take a break,you’ll start makingmistakes.

b. Otherwise,carryon until thejob is done.If it is after5pm,andif youarenot tired(andyoudonot takea break),carry on until thejob is done.

In example(20a),thepartialRhemeis-complementconditioncorrespondsto “youdo not take a break”. Here, the only context in which it makes senseto assert

79

........

...........

...............

.................................................................

ρ 5 β 6θ 5 β 6c5;c4; c6;

c0 cψρ 5 α 6 : φρ 5 α 6 : ψρθ 5 α 6 : ψθ

φ A ψρ

c2c1

Figure 10.4: IS-sensitive updating with “α + Otherwiseβ” involving a partialRhemeis-complementcondition.

“Otherwiseβ” (with “otherwise”appealingto this condition)is thecontext result-ing from updatingtheinitial context c0 with α’s Themeis andwith thatpartof α’sRhemeis constitutedby the“if-clause”. This is shown schematicallyin Figure10.4.

Wenoteexample(20b)becauseweareuncertainwhether“otherwise”appealsto just the complementof the “if-clause” (i.e., part of α’s Rhemeis) or the com-plementof the entireRhemeis of α. Herewe feel that more researchis neededconcerningthe statusof (postposed)subordinatedclauseswith respectto the IS-partitioning,in particular, whetherthey shouldbetreatedwithin thematrix clause,or asseparateutterances(with their own IS-partitioning)(cf. (Gunthner1996)foradiscussionbasedon spokendata;cf. alsoKomagata’s paperat thisworkshop).

5 Conclusionsand Further Research

While we must still completeour discussionof “otherwise” with complex an-tecedents,we hopethat we have convinced the readerthat that IS is crucial toany accountof thesemanticsof “otherwise”. We recognizethatseveralproblemsremainunaddressed:

4 As alreadynoted,we have not identifiedthe rangeof thingsthat canserveasantecedents(i.e., provide conditions)for “otherwise”nor identifiedfromwherein the discoursethey cancome,otherthenthe previous clauseor anembeddedrelative clause. More importantly, we have not said why theyprovide conditionsfor “otherwise”: That is, we haven’t addressedthe ba-sic problemof what(alternative) conditionsaspeaker mayhave in mindandwhatfeaturesof languagegive evidencefor them.

Herewe have claimedthatthealternativesetsof InformationStructuregivesuchevidence.But they areclearlynot theonly evidence(e.g.,multi-clause

80

antecedentsseempossiblefor “otherwise”,asdoaccommodatedantecedents,bothof which would beoutsidetherealmof IS). And a moreparsimoniousanalysisof thedatawe have presentedmaynot involve IS at all: For exam-ple, Matthew Stone(personalcommunication)haspointedout that all ourexamplesinvolve generics,whichcanbeanalysedasinvolving asetof casesunderdiscussion. We must understandwhetherand if so, how, thesetwoconceptsarerelated.4 Therearecasesof postposed“if ”-clausesthat arebestanalysedashavingtheirown IS – asin:

(21) Q. WhenshouldI stop?

A1. StopB CED FTheme

ata redlightH*LL%' (�) *

Rheme

— thatis,' (�) *Theme

if it’s aweekend.H*LL%' (�) *

Rheme

A2. Stop' (�) *Theme

at a redlightH*LL%' (�) *

Rheme

— thatis,' (�) *Theme

if you' (�) *Theme

don’t wantto geta ticket.H*LL%' (�) *

Rheme

In thecaseof (21:A2), the“if ” clauseis playinga role similar to an“other-wise” clause,sothataddingan“otherwise”clauseappearsredundant.In thecaseof (21:A1), it may be that “otherwise” caneithercombinethe rhemesinto asingleconditionor considerthelateroneasaconditionof its own.4 Therole that the“otherwise”clauseplayswith respectto theprecedingdis-courseis clearly tied, at leastpartially, with the condition it is taken to beotherwiseto: In thecomplex “if ”-clauseantecedentswe have discussed,be-ing otherwiseto the Rhemeis (in main or subordinateclause)provides anexplanation, while beingotherwiseto theThemeis providesa elaboration ofwhatholdsin othercircumstances.While thismaycall into questiontheno-tion in RhetoricalStructureTheory(MannandThompson1987)thatthereisan“otherwise”rhetoricalrelationsignalledby theuseof “otherwise”,it stillgoesonly asmallway towardscharacterizingwhatis happening.4 Finally, we alludedearlierto waysin which thethemesof subsequentutter-ancesmay be relatedandhow “otherwise” wasa prime exampleof a con-trastive relationbetweenthemesor betweenthemeandpreviousrheme.Dis-covering andenumeratingthesepossibilitieswould do much to clarify therelationshipsbetweendiscoursestructureandInformationStructure.

81

Acknowledgements

I. Kruijf f-Korbayova’s work wasfundedby a Visiting Fellowshipfrom theBritishAcademyPostdoctoralFellowship,November1999— April 2000,andaPostdoc-toral Fellowship from theRoyal Society/ NATO, April 2000— March2001.Wewould like to thankMark Steedmanfor adviceandcommentson earlierdrafts.

Bibliography

Firbas,J. (1992).FunctionalSentencePerspectivein WrittenandSpokenCom-munication. Studiesin EnglishLanguage.Cambridge:CambridgeUniver-sity Press.

Gunthner, S.(1996).Fromsubordinationto coordination?verb-secondpositionin Germancausalandconcessive constructions.Pragmatics6(3), 323–356.

Hajicova, E. andP. Sgall (1987).The orderingprinciple.Journal of Pragmat-ics 11(4), 435–454.

Halliday, M. A. (1970).A Coursein Spoken English: Intonation. Oxford: Ox-ford UniersityPress.

Halliday, M. A. (1985). Introductionto FunctionalGrammar. London,U.K.:EdwardArnold.

Karttunen,L. andS. Peters(1979).Conventionalimplicature.In C.-K. Oh andD. A. Dinneen(Eds.),Syntaxand Semantics:Presupposition, Volume11,pp.1–56.AcademicPress.

Krifka, M. (1993).Focusandpresuppositionin dynamicsemantics.TheoreticalLinguistics19, 269–300.

Kruijf f-Korbayova, I. (1998).The DynamicPotential of Topic and Focus: APraguianDiscourseRepresentationTheory. unpublishedPh.D.thesis,Fac-ulty of MathematicsandPhysics,CharlesUniversity, Prague,CzechRepub-lic.

Kruijf f-Korbayova, I. andB. Webber(2000a,January10-12).Concession,im-plicature,andalternative sets.In H. Bunt (Ed.),Proceedingsof theInterna-tional Workshopon ComputationalSemanticsIWCS-4, Tilburg, theNether-lands,pp.227–248.

Kruijf f-Korbayova, I. andB. Webber(2000b,July). Discourseconnectives,in-ferenceandinformationstructure.In J. Bos andM. Kohlhase(Eds.),Pro-ceedingsof theICoS-2WorkshoponInferencein ComputationalSemantics,Dagstuhl,Germany.

Mann,W. C. andS.A. Thompson(1987).Rhetoricalstructuretheory:A theoryof text organization.ISI Reprint SeriesISI/RS-87-190,USC InformationSciencesInstitute,Universityof SouthernCarolina.

82

Mathesius,V. (1975). On information bearingstructureof the sentence.InS. Kuno (Ed.), Harvard studiesin syntaxand semantics. Harvard Univer-sity Press.

Rooth, M. (1985). A Theoryof Focus Interpretation. Ph.D. thesis,GraduateSchoolof theUniversityof Massachusetts,Amherst,Massachusetts.

Rooth,M. (1992).A theoryof focusinterpretation.Natural Language Seman-tics 1, 75–116.

Sgall,P., E. Hajicova, andJ. Panevova (1986).Themeaningof thesentenceinits semanticandpragmaticaspects. Dordrecht,TheNetherlands:Reidel.

Steedman,M. (1996).SurfaceStructure and Interpretation. Cambridge,MA:M.I.T. Press.

Steedman,M. (2000a).Informationstructureandthe syntax-phonologyinter-face.LinguisticInquiry 31(4), 649–689.

Steedman,M. (2000b).TheSyntacticProcess. Cambridge,MA: M.I.T. Press.

Vallduvı, E. (1992).TheInformationalComponent. New York: Garland.

Webber, B., A. Joshi,A. Knott, andM. Stone(2001).Anaphoraanddiscoursesemantics.Submittedto theJournalof ComputationalLinguistics.

Webber, B., A. Knott, M. Stone,andA. Joshi(1999).Discourserelations: AstructuralandpresuppositionalaccountusinglexicalisedTAG. In Proceed-ingsof the36th AnnualMeetingof theAssociationfor ComputationalLin-guistics, CollegeParkMD, pp.41–48.

83

ConnectingInf ormation and DiscourseStructur eLevelsthr ough“K ontrast:”Evidencefr om Colloquial RussianParticles –TO,ZE, and VED’

SVETLANA MCCOYBostonUniversity

[email protected]

ABSTRACT. Thenotionof kontrast,or theability of certainlinguistic expressionsto generateasetof alternatives,originally proposedby Vallduvı andVilkuna (1998)asaclause-level concept,is re-analyzedhereasconnectingthelevel of informationpackagingin theclauseandthelevel ofdiscoursestructurein thefollowing way: kontrastis encodedat theclausallevel but hasrepercus-sionsfor discoursestructure.Thisclaimis supportedby evidencefrom thedistributionpropertiesof threecolloquial Russianparticles–TO, ZE, andVED’ which areanalyzedas unambiguousmarkersof kontrast. Both the placementof theseparticlesat the clausallevel andtheir role indiscourseareviewed asconsequencesof the typeof thekontrastsetandthecognitive statusofinformationmarkedby eachparticle.1

1 Intr oduction

Thenotionof kontrast, introducedin Vallduvı andVilkuna(1998,V&V hereafter),establishesa conceptualdistinctionbetweentwo notionsthathave beenconflatedin the literatureby the term of focus: one is rheme, a conceptwhich, in oppo-sition with theme, belongsto the domainof informationpackaging(Chafe1976inter alia) andtheotheris what theauthorslabel askontrast, thenotioncoveringquantificationalphenomenaof a moreformal semanticnature(Rooth1985,1992,Krifka 1991-92,etc.).V&V usethetermkontrast

asa cover term for several operator-like interpretationsof focus thatone finds in the literature: identificationalfoci, exhaustivenessfoci,contrastive foci, contrastive topics, and also interrogative wh-words,

1This paperis basedon chapter8 of my Ph. D. dissertation.I would like to expressgratitudeto my committeemembersfor all thehelp I receivedfrom them: Mary CatherineO’Connor, BruceFraser, EnricVallduvı, CatherineV. Chvany, andShanley Allen. Specialthanksgo to my first readerPaulHagstrom.

85


whichhavebeenseenasparadigmatic“foci” by many semanticistsandsyntacticians...

Thebasicideabehindthenotionof kontrastis thefollowing: if anex-pressiona is kontrastive, a membershipsetM= . ...,a,...0 is generatedandbecomesavailableto semanticcomputationassomesortof quan-tificationaldomain.(V&V : 83)

V&V provide a convincing setof datafrom Finnish,Hungarian,andCatalan(with referencesto Englishandotherlanguages)to demonstratethat kontrastive-nessandrhematicitymustbetreatedseparately. They assignformal features[K:+]for kontrastive elementsand[K:-] for non-kontrastive ones;similarly, [Rh:+] forrhemesand[Rh:-] for themes.Thus,therearefour possiblecombinationsof thesefeaturesfor any givenelement:

1. [K:+; Rh:+] – this combinationhasbeendiscussedin literatureunderthelabelof contrastivefocus(also,identificationalfocusandexhaustivenessfo-cus);2

2. [K:+; Rh:-] – thiscombinationhasbeenknown ascontrastivetopics;3

3. [K:-; Rh:+] – regular, non-kontrastive rhemes,or foci;

4. [K:-; Rh:-] – regular, non-kontrastive themes,or topics.

As pointedout by V&V, in different languageskontrast is expressedby dif-ferentlinguistic resources.For example,in Englishkontrast is signaledmainly byprosodywhile in Finnish it is encodedby syntacticposition. Moreover, thereisa conflict betweena limited setof structuralresourcesin a languageanda setofinterpretive categoriesthatneedto beexpressed.

Thegoalof thispaperis todemonstratethattheuseof thenotionof kontrastcanbeextendedfrom aclause-level notionto oneconnectingtheclause-level structureandthestructureof discourse.Eventhoughthisset-generatingpowerisencodedontheclausallevel, it is oneof thelinguistic meansthatholdsthediscoursetogetherandreflectsthe speaker’s assumptionsaboutthe hearer’s stateof knowledgeandattention.

The linguistic evidencefor this claim comesfrom discourseparticlesof col-loquial Russian–to, ze, andved’, which areanalyzedaslexemeswhoseprimaryfunctionis to signal,or unambiguouslymark,kontrastandwhicharelabeled“kon-trastive markers,” or “k-markers” (McCoy 2001).

2As pointedout by V&V, thereis an additionalvariablepresentin this combination,i.e., thescope:[K:+] hasa scopeand[Rh:+] hasascope,andthey neednotbut maycoincide.I will indicatethescopeof thekontrastive elementwithin therhemeby [ G K H RhemeII�I ].

3V&V (107,fn. 4) make animportantpoint thatnotall thematicconstituentscanbekontrastive:only links (i.e., pointersto a specificfile card for the entry of rheme) canwhile tails (i.e., the re-mainingpartof theme) cannot(seeVallduvı 1992for terminology).Thescopeof the[K:+] elementwithin the link will bemarkedhereby [ G K H Link II�I ].

86

Until recently, only descriptive analyseswereavailablefor this groupof parti-clesandparticles–to, ze, andved’werelabeled“emphatic,” “contrastive,” “intensi-fying,” “expressive,” “emotional,” “strengthening,” “enunciative,” etc.,with furtherclassificationof context-dependentmultiple meaningsor functionsfor eachparti-cle (seeVasilyeva 1972,collectionof articlesin threevolumesof LesParticulesEnonciativesen RusseContemporain, etc.) However, descriptive approachestotheseparticleshave proved inadequateanda searchfor anunderlyingmeaningof(someof) theseparticlehasbeeninitiatedin thefollowing works: Bitextin (1994),Parrott (1997),BonnotandKodzasov (1998),Feldman(forthcoming). However,theseearlystudiesin the“unifying” directionhavetheir limitations: they areeitherbasedonasingleframework and/orchooseto dealwith unifying asingleaspectoftheparticle(s),suchasdiscourseroleor cognitive statusof informationmarked.

In the previous analyses,the placementrulesof theseparticlesat the clausallevel madereferenceto eitherthepositionof theparticlewith respectto a prosod-ically prominentelementor to someinformation structureconstructas ‘theme-rheme’,‘topic-focus’, etc. However, no comprehensive semanticexplanationhasbeenoffered for explaining what the prosodicmarking is usedfor andwhat therelationshipbetweenprosodyand information packagingat the clausallevel is.Similarly, thediscourserole of theseparticleshasbeenanalyzedasgiving rise tomultiple implicaturesandinferencepatternsandthuscontributing to thecohesionandcoherenceof discourse.

In this paperit will be shown that suchpropertiesof theseparticlesas theirposition in the clauseand their role in discourseare consequencesof their twoessentialpropertiesask-markers:

1. thetypeof setit marksandthetypeof membershipwithin thisset;

2. thecognitive statusof thereferentsmarkedby it.

Below, these3 particleswill begivena unifying analysiswhich integratesthefollowing currentframeworks: thetheoryof “kontrast”(V&V); cognitive statusesof referentsin discourse(Yokoyama1986,Gundel,Hedberg, andZacharski1993);informationpackagingon theclausallevel (Vallduvı 1992);andhierarchicalorga-nizationof discourse(Buring 20004).

Thedatacomefrom avarietyof colloquialRussiantexts,suchasProtassova’scorpusof the CHILDES database(MacWhinney 2000) and Zemskajaand Ka-panadze(1978). The structureof the paperis the following: in the next threesections,the kontrastive propertiesand their consequenceswill be discussedforeachparticle individually, with particle –to given a more detailedanalysisthanparticleszeandved’. The final sectionsummarizesthe findingswith respecttowhat the analysisof thesethreeparticlesask-markers revealsaboutthe role ofkontrastat theclausalanddiscourselevels.

4For similar proposalsseeKanerva andGabriele(1995), van Kuppevelt (1995,1996a,b),andRoberts(1996).

87

2 K-Mark er –TO

I will first discusstwo essentialpropertiesof –to: thetypeof setandthecognitivestatusof informationmarked by it. ThenI will demonstratehow theseessentialpropertiesdeterminethepositionof –to in theclauseandits role in discourse.

2.1 Typeof SetMark ed by -TO

Particle–to marksa setof setsof relatedpropositions(equivalently, a setof ques-tions)whichisgeneratedby introducingalternativestoakontrastiveelementwithinthe link anda kontrastive elementwithin the rheme. Consideran examplefromCHILDES:

(1) [Varjaandhermotherarelookingatapictureof adogwhoputherpaw ona bear. They first discussthedog, thenstarttalking aboutthebear. Varjagetsdistractedby takingascoopinto herhand.Mothersays:]

U tebja-TOAt you-TO

sovok,scoop

abut

chtowhat

uat

medvedjabear

v lape?in paw

(CHILDES,seance2)

YOU(-TO) have aSCOOP, but whatdoestheBEAR have in his paw?

The informationstructureof the(Englishversionof the)utterancecontaining–to is shown in (2). Below it, in (3), a generalizedstructureof theutterancewith–to is given:

(2) [+K/Link You] (-TO) have [+K/Rhemeascoop] +J+J+(3) [+K/Link A]-TO hasproperty[+K/Rhemex]

Thedefault casefor (3) is whenthekontrastive (elementwithin the)link A belongsto a setof entities: M1= . A, B, C, +J+J+K0 andthe kontrastive (elementwithin the)rhemex belongsto a setof properties:M2= . x, y, z, +J+J+L0 .5 Thus,thepropositionx(A) markedwith –TO makesthehearergenerateasetM3, shown in threealterna-tivewaysin (4/5/6). In (4), thesetM3 is representedasasetof setsof propositions,in (5) thesamesetis shown asa setof questions(following Hamblin1958/1973andKarttunen1977),andin (6) thesetM3 markedby –to is shown graphically:

5However, amoremarked(i.e., lessfrequent)scenariois alsopossible:i.e.,whenthekontrastsetfor links is a setof properties,while thesetof alternativesto therhemeconsistsof entities,asin (i):

(i) [Varjacan’t stoprunning.Motheris asking:who is theonerunning?]

NuWell

[+K/Rhemekto]who

[+K/Link begaet]-TO?is-running-TO

(CHILDES,seance2)

As for somebodyrunning(-TO), who is doingthis?

88

(4) M3= MN. x(A), y(A), z(A), +J+J+O0 ;. x(B), y(B), z(B), +J+J+O0 ;. x(C), y(C), z(C), +J+J+O0 ; +J+J+QP(5) M3= M What is true of A?; Whatis trueof B?;Whatis trueof C?; R9R9RSP(6) Thesetof setsof propositions(or, a setof questions)M3 markedby –TO:

QuestionUnderDiscussion:Which entity possesseswhatproperty?T ! U

subq1 subq2 +J+J+ subqnWhatpropertydoesA have? WhatpropertydoesB have? WhatpropertydoesC have?V W X V W X V W Xx(A)-TO y(A) z(A) x(B) y(B) z(B) x(C) y(C) z(C)

So,theprimaryfunctionof –to is amarkerof asetof setsof propositionswhichis generatedby introducingalternativesto thekontrastive link andthekontrastiverheme.While for thepropositioncontaining–to thetruthvalueis asserted,it is notthecasewith thealternative propositions:they areonly madesalientwith thehelpof –to.

2.2 CognitiveStatusof ReferentsMark ed by -TO

Particle-to marksinformation(estimatedby thespeaker to be)known to thehearerbut notactivatedin thehearer’s mind at thetimeof theutterance.6

How canthespeaker assumethattheinformationis alsoknown to thehearer?The speaker cansafelyassumethat the heareralsoknows the informationif thisinformationis sharedthrough:code(language,culture),encyclopedicknowledge,deixis,or commonexperience(for moredetail,seeYokoyama1986).Theexamplein (7) illustratesa speechsituationwherethe sourceof the speaker’s assumptionaboutthepropositionmarkedby –to beinglocatedin thehearer’s knowledgesetisdeixis,while in (8) it comesfrom commonexperience:

(7) [A to B, aftera long silence,on thetopof amountainbeforedawn]

Tisina-TOQuietness-TO

kakaja!what

(Vasilyeva1972:68)

‘How quietit is!’ or:‘As for thestateof quietness(-TO), how quietit is!’(Vasilyeva’s translation:How wonderfullyquiet!)

6Sinceall threeparticlesarecapableof markinginformationwhich is, in speaker’s estimation,currentlyactivatedin thehearer’s mind or inferablefrom informationlocatedthere,I amprimarilyconcernedwith cognitive statusesthatarelocatedfurtheraway from theinformationwhich is at thecenterof thespeaker’s andthehearer’s attentionat themomentof theutterance.

89

(8) [Varjapracticespulling lacesthroughholes.After severalotheractivities,Mothersays:]

Oj,Oh

slushaj,listen

abut

mywe

swith

tobojyou

[ +K/Link chernen’kij]-TOblack-TO

shnurocheklace-DIM

[+K/Rhemetakso

ninot

razuonce

eshcheyet

ieven

neNEG

prodevali ],pulled-through,

da?yes

(CHILDES,seance5)

‘Oh, listen,asfor theBLACK(-TO) lace,youandI haven’t pulledit throughyet evenonce,right?’

To summarizetheessentialpropertiesof –to asak-marker: with respectto thetypeof setmarked, -to marksa setof setsof propositionswhich differ from eachotherin thevaluesof thekontrastive link andthekontrastive rheme.With respecttocognitive statusof informationmarkedby –to, this informationis assumedby thespeaker to beknown to thehearerbut notactivatedin thehearer’s mindat thetimeof theutterance.In thenext subsectionI will demonstratethat thesetwo essentialpropertiesof -to determineits positionin theclauseandits role in discourse.

2.3 Consequences:Position in the Clauseand Role in Discourse

Thepositionof –to in theclauseisbestanalyzedwith respectto theelementmarked[+kontrast].Thus,-to is anenclitic to thekontrastive elementwithin the link. Thisis a directconsequenceof its kontrastive properties:first, thekontraston the linkvalueresultsin a setof setsof propositions(cf. similar treatmentof “contrastivetopics” in Buring 2000,etc.); second,thecognitive statusof informationmarkedby –to — known to the hearerbut not activatedin the hearer’s mind — is morecompatiblewith thefunctionof the link thantherheme.

Theroleof –to at thediscourselevel is alsoaconsequenceof beingamarkerofa setof questions(equivalently, a setof setsof propositions).7 In a discoursetree,asin (9), any questionunderdiscussion(QUD) canbeconstrainedby k-marker -toin thewayasshown in (6):

(9) DiscourseSituationV W XQUD1 QUD2 QUDn

7Anotherconsequenceof the kontrastive natureof –to is its multiple implicaturesproposedinthe literature. Since–to is a marker of a set of setsof propositions,it also markscontrastandemphasis,functionsas a theme/topicmarker or a marker of contrastive topics, implies plurality,etc. The following implicaturesare consequencesof its secondessentialproperty—amarker ofinformation(assumedby speaker to be)known to hearerbut notactivatedin discourseyet: a markerof unexpectednessin addressinga topic; the fact that it addsa toneof familiarity, conversationalspontaneity, intimacy, etc.

90

To summarize,thenotionof kontrastis thecoresemanticmeaningof particle–to: by analyzingit asa marker of a setof setsof propositions(equivalently, asetof questions)andasmarkingreferentsknown to the hearerbut not currentlyactivated,its positionin theclauseandits role in discourseareaccountedfor.

3 K-Mark er ZE

K-marker ze differs from –to in both the type of set and cognitive statusof in-formationmarkedand,accordingly, its otherimportantproperties(positionin theclause,role in discourse,etc.)arealsodifferentfrom thoseexhibitedby –to. How-ever, aswith –to, kontrastivenessis thecoresemanticmeaningof this particleanddeterminesits distribution.

3.1 Typeof SetMark ed by ZE

Particle zemarksa setof propositionswhich differ from eachotherin the valueof (at least)oneterm. Thekontrastsetfor zecontainsmemberswhich aremutu-ally exclusive: if onepropositionis true, the otherone(s)is/arefalse. Thus, therelationshipamongthe membersof the zesetoften involves(binary) opposition,contradiction,or negation.Thekontrastsetmarkedby zeis generalizedin (10)andillustratedwith asampleof naturallyoccurringdatain (11):

(10) The kontrast setmarked by ZE:M= Y X, X’ Z , whereX= [ X’(X is trueif andonly if X’ is false)

(11) [Varjanoticesafly on thewindowsill andaskshermotherto kill it]VAR: Onamuxa,muxa.MOT: Muxa,muxa,da.VAR: Ubit’, ubit’ ee!MOT: Onazheuzheubita. (CHILDES,seance4)

VAR: It’sa fly, afly.MOT: A fly, a fly, yes.VAR: Kill, kill it!MOT: It (ze) is alreadykilled.

Gloss:

Onashe

zheze

[ \ K uzhealready

ubita].killed (participle).

‘(But) it (ze) is alreadykilled.’

In (11),mother’s utterancewith zecorrectsthepresuppositionof Varja’sprevi-ousutteranceKill it! , i.e.,Thefly is alive. Themembersof thekontrastsetmarkedby zearemutuallyexclusive: it is impossiblefor thefly to bealive (presupposition

91

of Varja’s utterance)andto bealreadykilled (mother’s utterancecontainingze) atthesametime.8

3.2 CognitiveStatusof ReferentsMark ed by ZE

Particle ze marksthe membershipset, onememberof which is activatedin thehearer’s mind at thetime of theutterance,while theother(incompatible)memberis viewedby thespeaker asthoughit is (or shouldhave been)known to thehearerandshouldhave beenactivatedat this time.

For example,in (11),theactivatedmemberof thesetis Thefly is alive (presup-positionof Kill, kill it! ). Theothermemberof theset(is treatedby thespeaker asthoughit) shouldbeknown to thehearerand,therefore,shouldbeactivated— i.e.,in thespeaker’s estimation,thereareenoughvisualcuesfor thehearerto make thecorrectconclusion(Thefly is alreadydead).

Now let us considersomeconsequencesof the essentialpropertiesof zeasak-marker.


Theplacementof particlezecanbedefinedwith respectto thekontrastiveelement,which is a propositionalterm(usuallywithin therheme). Theconditionof mutualexclusiveness(or binary opposition)on the setmarked by zemakesthis particlea good candidateof marking rhematickontrasts,especiallycasesof kontrastive

8An interestingcaseis the useof zewith wh-words. This caseis somewhat cumbersomeforthe presentanalysis:how in the world canthe kontrastsetbe mutually exclusive with wh-words?It seemsthat thereis a way out of this paradox: zewith wh-words is usedto indicatethat fromthe speaker’s point of view, the setof possibleanswersis emptyof reasonablealternatives(andisnow opento unreasonablealternatives). While the setof possibleanswersis emptyof reasonablealternatives for the speaker, it might not be so for the hearer: the speaker expectsthe hearertoprovide an answer(which is probablynot so unreasonablefrom the hearer’s point of view). Anexample,recordedin a homesetting,is providedin (i):

(i) [Fatherputsaway his son’s library booksandtells him that mothertook thembackto thelibrary (which shedid not do). Theboy asksmotherif sheindeedtook themback.Shesays‘no’. He thensays(in thepresenceof bothparents):]

U kogo zemnetogdasprosit’?‘Who (ze) (in theworld/thehell ]]] ) shouldI askthen?’

In (i), from thespeaker’spointof view, thesetof reasonableanswers(individualswhocanprovidean answer)is exhaustedor empty; the child now throws the ball to the parentswho might be ableto provide an alternative explanation,which at this point in discourseis viewed by the child asanunexpected/non-predictable/unreasonablealternative. However, the answerto the child’s questionis indeedknown to the parent(s)who optednot to reveal it to him. It seemsthat an explanationalongtheseline would bring usascloseto themutuallyexclusive setaswe couldpossiblygetwithwh-words.

92

verumfoci9 — i.e., kontrastnot on the lexical meaningof the verb but on someverbal inflectionalcategory, suchas tense,aspect,etc. It is possibleto translateutterancescontainingzeascasesof rhematicpolarityof verumfocus;for example,alternative translationsfor theutterancewith zein (11) areIt [=the fly] IS alreadykilled or It is ALREADYkilled.

Theroleof zein discourseis alsoaconsequenceof its kontrastiveproperties:zerefersbackto asalientelementor someunresolved(from thespeaker’s viewpoint)questionin thediscourseor discoursesituation.10

To summarize,by analyzingparticlezeasa marker of asetof mutuallyexclu-sive propositionsandasa marker of informationwhich is relatedthroughthesetrelationshipto informationthat is activatedin discourse,its distributional proper-tiesat theclausalanddiscourselevelsarelargely accountedfor.

4 K-Mark er VED’

Etymologically, particleved’ is a form of theverbvedat’ ‘to know’, from which ithasinheritedfactivity.

4.1 Typeof SetMark ed by VED’

Similarly to ze, particleved’ marksa setof propositionswhich differ from eachotherin the valueof at leastoneterm. However, thereareimportantdifferencesbetweenthem:for ved’ thesetmembershipis restrictedto propositionswhichhaveillocutionaryforceof assertions(while zedoesnothave this restriction).Also, themembersof the set are not mutually exclusive (as in the caseof ze) but rathersupplementary;whenved’ doesexpresscontradiction,it is moreof the ‘yes,but^J^J^ ’ type.

The examplesin (12-13) illustrate that the differencebetweenze and ved’comes(partially) from the differencebetweena mutually exclusive set( ze) andasupplementaryset(ved’):

(12) [Varja andGrandmotherarelooking at a pictureof birds standingon theground.Varjabelievesthat thebirdshave fallendown. Grandmothercor-rectsVarja:sincethey arestandingandnotlying down, they havenotfallendown:]*VAR: Etapit’ki.

9Thanksto EnricVallduvı for pointingthis out.10Multiple implicaturesassociatedwith zeat the discourselevel canalsobe viewed as follow-

ing from its two essentialproperties.Sinceze is a marker of a restrictedset,containingmutuallyexclusive propositions,it is subsequentlya marker of contrast,a marker of emphasis,a marker ofcontrastive focus,a marker of contrastive topic, it addsa toneof an indisputableargumentandisperceived asa verbalattackon hearer, etc. Frombeinga marker of activatedinformationwhich isto a largeextentknown to thehearer, it follows that zecanalsobeanalyzedasa marker of a refer-encepoint in theactivateddomainof reference,asbeingperceived to bea verbalattackon hearer,asaddinga toneof an indisputableandirrefutableargument,andconveying emotionsof irritation,annoyance,andimpatience,etc.

93

%eng:Thesearebirds.*VAR: Upai.%eng:Fell down.*GPP:Oni neupali.%eng:They did not fall down.*GPP:Pochemuzhety dumaesh’,chtooni upali.%eng:Why do you think, thatthey fell down.*GPP: Odnastoit na penechke, drugajastoit na zemle,a tret’ja stoit nazemleszadipenechka.%eng: One is standingon a little stump, the secondis standingon theground,andthethird is standingon thegroundbehindthestump.*GPP:Vot tak,szadipenechka.%eng:Thatway, behindthestump.*GPP:A pochemuty dumaesh’,chtooni upali,oni zhnelezhat,oni stojat.%eng:And why do you think, that they fell down, they arenot lying, butstanding.*GPP:Kogdakto-nibud’ upadet,on lezhit.%eng:Whensomebodyfalls,heliesdown. (CHILDES,seance3)

GPP:Pochemuzhety dumaesh’,chtooni upali.‘Why (ze) in theworld do you think thatthey fell down.’

GPP:A pochemuty dumaesh’,chtooni upali,oni zhnelezhat,oni stojat.‘And why do you think thatthey fell down, they (ze) arenot lying down, they arestanding.’

In (12), zemarksasetcontainingmutuallyexclusive members(thusit conveysemotionsof contradiction,correction,andis perceived asa ‘verbalattackon thehearer’): Varja believes that Thebirds fell down (sincethey arestandingon thegroundand not flying), while the Grandmotherrefutesthat by saying that Thebirds did not fall down (sincethey arestandingandnot lying down). Later, thisargumentis broughtto amediator, Varja’smother(Katja). In presentingthegistoftheargumentto Katja, theGrandmotherusesved’ andnot ze(sinceno oppositionis expectedfrom anadultwho knows thelexical meaningof upali ‘ fell down’):

(13) [Grandmotherre-tellstheargumentaboutthe“f allen” birdsto themother]

VotPART

onashe

govorit,says,

chtothat

etithese

ptichkibirds-DIM

upali,fell-down,

aand

jaI

schitaju,think

chtothat

onithey

[ _ K neNEG

upali ],fell-down,

onithey

[ _ K stojatare-standing

ved’ved’

naon

nozhkaxfeet-DIM

].

(CHILDES,seance3)

‘So shesaysthat theselittle birds fell down andI think that they did notfall down, they arestanding,(ved’) youknow, on their feet.’

So, in (13) ved’ marksa supplementaryset, the membersof which aresyn-onymsratherthanantonyms(thus,it constitutesan‘extensionin logic’ andnot ‘averbalattack’): Y Thebirdsdid not fall down;Thebirdsarestandingontheir feetZ .

94

4.2 CognitiveStatusof ReferentsMark ed by VED’

Similarly to –to, particleved’ marksinformationknown to thehearerbut not acti-vatedin thehearer’s mind at thetime of theutterance.Thedifferencebetweenzeandved’ comes(partially) from thedifferencein thecognitive statusof informa-tion marked.Considerexample(14) from ZemskajaandKapanadze(1978:52-53;glossesandtranslationadded):

(14) [Two retiredballetdancershadpreviouslytalkedaboutwhatgreatactorstheyusedto work with in their youth.Later, oneof themrealizesthatthey forgotto mentiontwo of themostfamousactors(Cirkov andCerkasov) andremindsherfriendaboutthat]

A: Kogdamy s tobojgovorili o mjuzik-holle,my govorili: axkakiebyliaktery!‘When you andI talked aboutthe musichall, we weresaying“Ohwhatactorswerethere!”’

MyWe

swith

tobojyou

ved’ved’

neNEG

skazali,said

ctothat

tamthere

bylwas

Cirkov,Cirkov,

tamthere

bylwas

Cerkasov!...Cerkasov‘You know, we did not say that therewas Cirkov there,therewasCerkasov there!..’

B: [overlappingwith theendof theprecedingutterance]Dagospodi!Cerka-a-sov, Cirko-o-v...‘Oh, my god! Cerka-a-sov, Cirko-o-v...’

In (14),ved’marksinformationknown to boththespeakerandthehearerwhichis notactivatedin thehearer’s mindat themoment:talkingaboutwhatgreatactorswerein the theaterat the time whenboth of the interlocutorsworked there,theyforgot to mentiontwo mostfamousactors.


Similarly to theothertwo particles,thepositionof ved’ in theclauseis bestdefinedwith respectto thekontastive element:it is eithera proclitic or anenclitic to suchelement.

The discourserole of ved’ is determinedby the type of set it marks: beinga marker of a supplementaryset, it relatesmembersof the set to eachother bybuilding asuper-questionabove them.11

11Also, multiple implicaturesof ved’ canbetreatedasconsequencesof its kontrastiveness:sinceved’ is a marker of a (supplementary)setof propositions(assertions),it is a marker of contrast,amarker of emphasis,anassertionmarker, etc.Sinceved’ (like –to) is a marker of informationthatis

95

-TO ZE VED’ESSENTIAL PROPERTIES:Type of setmarked

A set of sets ofpropositions (= aset of questions);requires two [+K]elements:in the linkandtherheme

A setof mutuallyex-clusive propositionswhich differ in thevalue of at leastone[+K] element

A supplementaryset of propositionswhich differ in thevalue of at leastone[+K] element

Cognitivestatus ofreferents

Familiar to thehearer but not ac-tivated in his/hermind

One member: infocus or activated;the other member:(treated as) familiarandactivated

Sameasfor -TO

CONSEQUENCES:Position in aclause

Enclitic to [+K] ele-ment,usuallylink

Enclitic to [+K] ele-ment

Procliticor encliticto[+K] element

Role in dis-course

Generatesa set ofsub-questions (sis-ters dominated bythesameQUD); par-tially answersQUDandmakessalienttheothersub-questions.

Refersbackto salientelementor someun-resolved questionin the previousdiscourse

Generatesa set bybuilding a super-question above themembersof theset

Table11.1:ComparingKontrastive Markers–TO, ZE, andVED’

To summarize:ved’ hasbeenobserved to sharecertainpropertieswith zeandotherpropertieswith –to. By addressingthequestionsof whattypeof setandwhatcognitive statusof informationeachof theseparticlesmarks,their differencesandsimilaritiesareexplained.

5 Conclusion

Thenotionof kontrasthasbeenfoundto beacoresemanticmeaningof suchlexi-cal itemsasparticles–to, ze, andved’: theselexemesareanalyzedasunambiguousmarkersof kontrast.Thecoresemanticmeaningof eachof theseparticleshasbeenfoundto encompasstwo issues:thetypeof setandthecognitive statusof referentsmarkedby eachof theseparticles.Amongimportantpropertiesof theseparticles,which areneverthelesstreatedasconsequencesof their essentialkontrastive prop-erties,arethe placementof the particlesat the clausallevel andtheir role in theorganizationof discourse. Their distribution propertiesand the choiceof parti-clesin a particularcontext is alsomotivatedby thedifferencesin their kontrastive

assumedby speaker to beknown to hearerbut not activatedyet, it is a marker of unexpectednessinaddressinga topic, a marker of encyclopedicknowledge,andit is perceived asa (polite) reminder,etc.

96

nature.Thefindingsaresummarizedin Table4.3.Theexampleof thesethreecolloquialRussianparticlessuggeststhatthenotion

of kontrasthasapplicationsbeyond the level of theclauseto serve asa constructwhichconnectsthelevelsof clausalanddiscoursestructure.

Data Sources

CHILDES: Other Languages- Russian– Protassova. In: MacWhinney, Brian.2000.TheCHILDESDatabase:Toolsfor AnalyzingTalk. 3d Edition. Vol. 2: TheDatabase. Mahwah,NJ: LawrenceErlbaumAssociates.

Zemskaja,E. A. and L. A. Kapanadze,eds. 1978. RusskajaRazgovornajaRec’: Teksty. Moskva: Nauka.

Bibliography

Bitextin, AleksandrBorisovic. 1994. Casticy -to, ze, ved’ i Vvodnye Kon-strukciiTipakakizvestnokakSredstvaUkazanijanaIzvestnost’Propozicional’nogoSoderzanijaPredlozenijaSlusajuscemu.AKD, Moskovskij Gosudarstvennyj Uni-versitet.

Bonnot, Christineand Kodzasov, C. B. 1998. Semanticeskoe Var’irovanieDiskursivnyx Slov i Ego Vlijanie na Linearizaciju i Intonirovanie (Na PrimereCastic ZE i VED’). In: Kiseleva, K. and Paillard, D., eds. DiskursivnyeSlovaRusskogoJazyka:OpytKontekstno-Semanticeskogo Opisanija.Moskva: Metatekst.

Buring,Daniel.2000.On D-Trees,Beans,andB-Accents.Ms., UCLA.Chafe,WallaceL. 1976. Givenness,Contrastiveness,Definiteness,Subjects,

Topics,andPointof View. In: Li, CharlesN., ed. SubjectandTopic. New York:AcademicPress,25-55.

Feldman,Anna. Forthcoming. DiscourseMarkers- Accessing‘Hearer-Old’Information:TheCaseof theRussianZe. In: Melby, Alan et al., eds.Proceedingsof the27thLACUSForum.

Gundel,JeanetteK., Nancy Hedberg, and Ron Zacharski. 1993. CognitiveStatusandtheFormof ReferringExpressionsin Discourse.Language 69(2): 274-307.

Hamblin, C. L. 1958. Questions.AustralasianJournal of Philosophy36(3):159-168.

Hamblin, C. L. 1973. Questionsin MontagueEnglish. Foundationsof Lan-guage 10: 41-53.

Kanerva,JonniM. andLeslieA. Gabriele.1995.IntonationandFocusLayers.In: Beckman,Jill N., ed.Proceedingsof theNortheastLinguisticSociety25(NELS25). Vol. 1: Papersfrom theMain Session.Universityof Pennsylvania,335-346.

Karttunen,Lauri. 1977. SyntaxandSemanticsof Questions.LinguisticsandPhilosophy1: 3-44.

97

Krifka, Manfred. 1991-2.A CompositionalSemanticsfor Multiple FocusCon-structions.LinguistischeBerichte, Suppl.4: 17-53.

McCoy, Svetlana. 2001. Colloquial RussianParticles–To, Ze, And Ved’ AsSet-Generating(“Kontrastive”) Markers:A Unifying Analysis.Ph.D.dissertation,BostonUniversity.

Les Particules Enonciativesen RusseContemporain 1 (1986), 2 (1987), 3(1988).Paris: Institutd’EtudesSlaves.

Parrott, Lillian. 1997. DiscourseOrganizationandInference:The UsageoftheRussianParticleszeandved’. Ph.D. dissertation,HarvardUniversity.

Roberts,Craige. 1996. InformationStructurein Discourse:Towardsan In-tegratedFormalTheoryof Pragmatics.In: Yoon,J. H. andAndreasKathol, eds.OSUWorkingPapers in Linguistics49: Papers in Semantics, 91-136.

Rooth,Mats. 1985.Associationwith Focus.Ph.D.dissertation,UniversityofMassachusetts,Amherst,MA.

Rooth, Mats. 1992. A Theory of FocusInterpretation. Natural LanguageSemantics1: 75-116.

Vallduvı, Enric. 1992.TheInformationalComponent.New York: Garland.Vallduvı, Enric, andMaria Vilkuna. 1998. On RhemeandKontrast. Syntax

andSemantics29 (TheLimits of Syntax),79-108.van Kuppevelt, Jan. 1995. DiscourseStructure,Topicality andQuestioning.

Linguistics31: 109-147.van Kuppevelt, Jan. 1996a. Directionality in Discourse:ProminenceDiffer-

encesin SubordinationRelations.Journal of Semantics13: 363-395.vanKuppevelt, Jan. 1996b. Inferring from Topics. Linguistics& Philosophy

19: 393-443.Vasilyeva, A. N. 1972. Particles in Colloquial Russian:Manual for English-

SpeakingStudentsof Russian. Moscow: ProgressPublishers.Yokoyama,OlgaTsuneko. 1986.DiscourseandWordOrder. Amsterdam/Philadelphia:

JohnBenjaminsPublishingCompany.

98

Contrast fr om a ContrastivePerspective

VALERIA MOLNAR

Lund University, Sweden

[email protected]

ABSTRACT. Themainconcernof thepaperis to addresssomeproblemsthatarisewhendefiningtheconceptof contrast andto clarify the relationbetweencontrastandotherconceptsof infor-mationstructure.As to thedefinitionof thenotionof contrast,we needto separatethesemanticandpragmaticaspectsof contrastivenessaswell asidentify differenttypeswithin thesetwo maincategories. Thereis abundantcross-linguisticevidencefor the hypothesisthat certaintypesofcontrastmotivatea morefine-grainedanalysisof topicality andfocusing.It will beargued,how-ever, that contrastivenessnot only is a relevant featurefor the parametrizationof topicality andfocusing,but thatcontrastindeedrepresentsanautonomousconceptof informationstructuring.

1 Intr oduction

Themain concernof thepaperis to clarify the relationbetweencontrast andthetwo mainconceptsof informationstructuretopic andfocus. This requiresadefini-tion of contrastandthespecificationof thecompatibilityof differentcontrasttypeswith focusingandtopicality.

I will arguefor thefollowing assumptions:

(i) thenotionof contrastis a linguistically relevant phenomenonanddoesnotonly arise“from particularinferenceswhich we draw on thebasisof givenconversationalcontexts” (Lambrecht1994:290): thefeaturecontrastivenesshassyntacticandphonological consequencesfor therealizationof topic andfocus;

(ii) thedistinction of different typesof contrast is necessaryin orderto explainits impacton linguistic forms,andabove all theseparationof thepragmaticlevel of contrastivenessfrom thesemantic, quantificationallybased,level isrequired;

(iii) thelinguistic statusof contrast mustbedifferentiated:contrastis notsimplya featureof topicality andfocusing,but it seemsplausibleto regardcontrast–in certainlanguagesandin certainconstructions–asa furtherautonomous“packagingphenomenon”. Onthebasisof cross-linguisticevidencecontrastmustbeincludedinto theinventoryof pragmaticcategories.

99


I would like to startmy discussionwith someintroductoryremarkson the basicnotionsof contrast,topic andfocusandtherelationsbetweenthem.

Accordingtoatheoryneutraldefinition(cf. Bußmann1990:419),thenotionofcontrast hastwo maindimensions:firstly, it is usedasasynonym for “opposition”– eitheron the paradigmaticor syntagmaticlevel – andsecondlyit alsoincludesalsoanotheraspect,namelyhighlightingby accent.

The notion of focus is relatedto highlighting in somesense,too. The viewthat focusis the “information centreof the sentence”andcontainsthe new, non-presupposedpart of the utteranceis widely acceptedin the literature. It is fore-groundedmostoftenby stresswhile therestof theutteranceremainsin theback-ground. Besidesthis syntagmatic(horizontal)type of highlighting, it is claimedthat highlighting in the paradigmatic(vertical) dimensionmay alsobe involved.Thebasicideais thata setof alternativesexists for thefocusedconstituentwhichstandsin oppositionto all of them.Obviously, thetwo mainpropertiesof contrast- oppositionin two directionsandhighlighting - arealsotypical of the notion offocus.Contrast andfocusarethusoftenregardedasvery closelyrelatedconceptsin linguistic research.

Accordingto anextremeview, focusingis alwayscontrastive - andasall utter-ancescontainafocus,all utterancesmustnecessarilybecontrastive - or asDretske(1952)putsit, “all contingentstatementscontrast[ ^J^J^ ] onestateof affairs withanother”. While Bolinger (1961: 87) defendstheview that “in a broadsenseev-ery semanticpeakis contrastive”, Dretske (1972: 412) arguesthat “contrastivestatements”mustberegardedasaspecialclass,becausethey “embodyadominantcontrast,a contrastive focus,a featuredexclusionof certainpossibilities”. In ex-amples(1) and(2) from Dretske, differentcontrastsarefeatured:in (1) the verbsold, but in (2) thenominalphrasemytypewriter standsout asthefocal point andembodiesthedominantcontrast:

(1) I soldmy typewriter to Clyde.

(2) I soldmytypewriter to Clyde.

The decisisonasto whetheror not focusingis inherentlycontrastive dependsonhow importantwe judge dominantcontrastto be. If not only highlighting, butfurther conditionssuchas “a featuredexclusionof certainpossibilities”needtobe met for contrast,thencontrastcanonly be an optionalandnot an obligatorypropertyof focusing.

As to topic, threedefinitionsdominatethe linguistic landscape.Firstly thosewhich definetopic asthe notion of aboutness(Reinhart1982)or asan “addresspointer”, alsocalled“link” by Vallduvı (1992). Accordingto anotherinfluentialview, the topic shouldbe regardedasthe notion of frame: “the topic setsa spa-tial, temporal,or individual framework within which themainpredicationholds”(Chafe1976:51). A third definitionof topic is givenwith recourseto old informa-tion: the topic is eitheridentifiedasgiven informationor in weaker versions,the“givennesscondition” is animportantpartof topicality.

100

In noneof thesethreedefinitionsdo thepropertiesof topicality have a directrelationto thenotionof contrast.Theseparationof the topic from therestof thesentencecanbe madein an explicit, emphasizedway, but this is not necessarilyso. In the analysisof the relation of contrast to topic, their compatibility is acontroversialissue,especiallyin theorieswhereemphasisonly canbeassignedtofocus.

2 The definition of the notion contrast

Themostimportantcriteria thatarediscussedin the literaturein connectionwiththedefinitionof contrastarelistedbelow with regardto theirhierarchy(in orderofimportance):`

highlighting`dominantcontrast`membershipin aset`limited setof candidates`explicit mentioningof alternatives

Therelevanceof thecriteriais judgedin differentways- leadingto differenttypesof distinctionsbetweenthecontrastive andnon-contrastive, regular foci. It is im-portant to note that the judgementsare partly dependenton the theoreticalap-proach,partly on the type or numberof languagestaken into considerationandthepresenceof obligatoryformalmarkingof thedistinctionin theselanguages.

Thebasicrequirementfor contrastivenessis theexistenceof highlighting.Notonly Bolinger links every semanticpeakto contrastiveness,but also Lambrecht(1994)regardshighlightingasa sufficient conditionof contrast.Accordingto theoverwhelmingmajorityof linguists,though,theexistenceof a“dominantcontrast”dividing the utteranceinto two parts- background(alsocalled“presupposition”)andfocusis a necessaryrequirementof contrast.Rochemont(1986: 52) distin-guishescontrastive focusfrom presentationalfocuson thebasisof this criterion.

A further condition of contrastis membershipin a set, namelythat we cangeneratea setof alternativesfor thefocusedconstituent- or asJackendoff (1972:243) puts it “a coherentclassof possiblecontrastswith the focus.” This condi-tion is regardedasa sufficient prerequisiteof contrastalsoby linguistsworkingin the framework of “Alternative semantics”(e.g. Rooth 1985). In many ap-proaches,however, “membershipin a set” is narrowed down to the requirementof aclosedset.Thedecisive criterionfor contrastivenessis thus,accordingto Hal-liday (1967),Chafe(1976)andRooth(1992),theavailability of a limited numberof candidates.Halliday(1967:206)defines“contrastive” “ascontraryto somepre-dictedor statedalternative” andChafe(1976:34)alsofavourstheview – opposed

101

to Bolinger’s suggestion– that “contrastive sentencesare qualitatively differentfrom thosewhich simply supplynew informationfrom anunlimitedsetof possi-bilities”. Whenthesetof possibilitiesis unlimited,thesentencesuppliesonly “newinformation”andfails to becontrastive.

While Chaferegards“the limited numberof candidates”asthe essentialdis-tinction betweencontrastive and non-contrastive readings,Jacobs(1988: 113)claimsthat the candidatesexcludedmustbe explicitly mentionedin the context.Accordingto Jacobs,thefocusof negationis thusinherentlycontrastive (3), whilethecontrastivenessof foci of theillocutionaryoperatorsis context-dependent, onlypresentin caseof explicit mentioningof alternatives(4):

(3) DiesesBuchhatmir nicht2 Kl ausempfohlen,sondernGerda.F2

(4) A: Ich furchte,daßwiederdie Schwedengewonnenhaben.

B: KeineAngst! DiesmalhabenwırF1

gewonnen.

(5) Wießtdu,wergewonnenhat?WırF1

(habengewonnen)!

Therearefurthercontroversialquestionsin connectionwith thedefinitionof con-trast, namelywhetherit shouldbe regardedas a gradientor distinct notion andwhetheror not contrastis anindependentphenomenonof informationstructure.

If onesharesBolinger’sandLambrecht’sopinion,accordingto whichfocusingis alwayscontrastive, thenutterancescanonly beusedcontrastively andit is onlypossibleto speakof “clearor lessclearinstancesof contrastiveness”.Evenfor thisgradientapproachthe “clearestinstancesof contrastiveness”arethose“in whicha focusdesignatumexplicitly contradictsa statedor predictedalternative” (Lam-brecht1994:290). Most approachesargue,however, for thenon-gradientcharac-terof contrastivenessandfor adistinctionbetweencontrastive andnon-contrastivecaseson thebasisof oneof theabove-mentionedcriteria. A relevant taskfor lin-guistictheoryis to takeastandin thiscontroversy, by expandingthedatato a largenumberof languagessothatcross-linguisticallyvalid argumentscanbemade.

Thesecondimportanttheoreticalquestionis the linguistic statusof contrast-basedon the impactof contraston linguistic realization. The standardview rep-resentedby linguistsworking within different theoreticalframeworks (seeaboveDretske,Bolinger, Halliday, Rochemont,Rooth,Jacobsetc)is thatcontrastcannotbe regardedasan independentphenomenonof informationstructure,but only asa featureof focusingandtopicality, which servesto furtherparametrizetheseno-tions. Analysingcontrastasa “packagingphenomenon”was,however, suggestedby Chafein 1976andalsoby recentlinguisticanalysesof Finnishdata.

102

3 The linguistic relevanceof contrast - within focusingandtopicality

In oppositionto the view accordingto which contrastive casescannotclearly beseparatedfrom non-contrastive ones,thereis abundantevidence(i) thatcontrastinmany languagescanbe optionally marked or even mustbe marked by grammat-ical means,and(ii) that certainformal phenomena(syntactic,morphologicalandphonological)canonly beexplainedwith recourseto thenotionof contrast.Con-sequently, contrastshouldbea linguistically relevantphenomenon,andnot only acognitively motivatedcategory.

3.1 Contrast vs. Focus

Contrast– or thepresenceof certainkindsof contrastiveness– motivatestheinter-nal differentiationof focusing.It is plausibleto assumethatthenotionof contrastappliesonly to thosefoci which operateon predictedor statedalternativesanddonot simply expressnew information. A further differentiationof contrastive fociseemsto be motivatedby thestatusof alternatives: explicitly mentionedalterna-tivestriggerothersyntacticpatternsthansimplypredictedalternatives.

A distinctionbetweentwo maintypesof focusingis suggestedon thebasisofempiricaldatafrom Hungarianby E.Kiss(1998).Thedichotomy “identificationalfocus”vs. “information focus” is basedondifferencesbothin syntacticrealizationandsemanticallyrelevantcontent.Themostimportantsyntacticdifferenceis thattheinformationfocusremainsin situ(6) while theoperator(“identificational”)typemustbemovedto a specialoperatorposition,in Hungariane.g. into thepreverbalposition(7) (cf. E. Kiss 1998:249):

(6) Mari ki nezettmaganakEGY KALAPOT.‘Mary pickedfor herselfA HAT.’

(7) MariMary

egya

kalapothat.ACC

nezettpicked

kiout

maganak.herself.DAT

‘It wasa hat thatMary pickedfor herself.’

Thesemanticdifferencecanbespecifiedasfollows: while thefocusoperator“rep-resentsasubsetof thesetof contextually or situationallygivenelementsfor whichthepredicatephrasecanpotentiallyhold” (E. Kiss1998:245),theinformationfo-cusonly conveys new, nonpresupposedinformation. Thesentence(7) containinga focus operatorexpressesthat it was only a hat and nothing elseMary pickedout for herself,whereastheinformationfocusin (6) merelypresentsa hat asnewinformation,without suggestingthat thehatwastheonly oneof a setof relevantthingsfor Mary.

From a cross-linguisticperspective, the different typesof contrastivenessde-mand,however, a further division of the focusfield - not only a dichotomy, buteven a trichotomy seemsplausible. This is the resultof a necessarydistinction

103

of two typesof contrastive foci, also representedsyntactically: The focusoper-ator operatingeitheron a contextually openor closedset requiresmovementtoa verb-relatedFocus-position.Movementinto the verb-adjacentFocusposition-whichaccordingto proposalsmadewithin generative frameworksis locatedin theleft peripheryof thesentence,i.e. in theFocusprojectionof theCP-domain(seebelow 5.2.) - is overt in so-calledFocuslanguages(e.g. Hungarian,Basque),butcanalsobe covert anddelayeduntil LF (as in English). On the otherhand,forthe focustypeby which alternativesareexcludedfrom a closedset,movementisnotnecessarilyverb-related,but in thiscasetheleft-peripheralpositionis decisive.Contrastive foci aremovedin syntaxinto a projectionappearingsentence-initiallyin surfacestructurein a greatnumberof languages(Italian,Spanish,Catalan,Ru-manian,Arabic,Greek,Russian,Finnishetc).

Thedifferencebetweenthetwo typesof FocusmovementcanbeillustratedbycomparingFinnishandHungariandata:in Finnishtheleftmostpositionof thesen-tenceis responsiblefor theexpressionof contrastiveness(8-B) andcanbeoccupiedonly whenthefocusedconstituentrefersto alternativesin acontextually limited setwherethealternativesareknown to theparticipantsof thediscourse.Thisexplainswhy (9-A) is notanappropriateanswerto ageneralinformative question(9-F1)orto aconstituentquestion(9-F2)if thealternativesarenotexplicitly mentionedor atleastpresentin themindsof thediscourseparticipants(cf. Jarventausta& Molnar,2001):

(8) A: PekkalensiTukholmaan.‘Pekkaflew to Stockholm.’

B: [KONTRAST Reykjavikiin ] Pekkalensi.‘To Reykjavik, Pekkaflew.’

(9) F1: Whatdid Pekkado?

F2: Wheredid Pekkafly?

A: * [KONTRAST Reykjavikiin ]Reykjavik-to

PekkaPekka

lensi.flew

In Hungarianthe oppositeis true: the focusoperatormustbe adjacentto the fi-niteverbin surfacestructure(10-B1),but is notnecessarilyrelatedto contextuallyspecifiedalternatives:

(10) A: HovaWhere

repultflew

Peter?Peter

‘Wheredid Peterfly?’

B1: PterPeter

[FOCUS Reykjavikba]to Reykjavik

repult.flew

‘Peterflew to Reykjavik.’

B2: * [FOCUS Reykjavikba] Peterreplt.

104

Basqueyieldsalsostrongevidencefor theassumptionof two contrast-relatedfocuspositions.In Basque,bothtypesof contrast-relatedfocusingarepossible:Etxepare(1997)distinguishesbetween“emphaticfocus”(11-a),theverb-relatedfocus(hereincidentallyin sentence-initialposition,but as(11-b)shows its verb-adjacentposi-tion is essentialfor grammaticality),and“contrastive focus” on the left peripheryof thesentence(12-b)excludingacontextually explicit alternative (12-a)- howeverwithout therestrictiononverb-adjacency:

(11) a. PATATAKpotatoes

maitelove

dituAUX

Jonek.Jon

‘It is potatoesthatJonloves.’

b. * PATATAK Jonekmaite ditu .

(12) a. MIKELI,for-Mikel

ardoawine

ekarribring

diote.AUX

‘It is for Mikel thatthey broughtwine.’

b. JONEK,Jon

ardoawine

ekarribring

du.AUX

‘It is Jonthatbroughtthewine.’

3.2 Contrast vs. Topic

Contrastivenessis not only marked in the field of focusing,but alsoin combina-tion with topicality, in thelattercaseoftenby theuseof specialsyntactic patternssuchastopicalizationandleft-dislocation.Left-dislocationandtopicalizationseemhighly restrictedin discourse:they serve to changethecurrenttopic of conversa-tion, andcontrastivenessin somesenseis importantfor theadequateuseof thesemarkedconstructions.Therelevanceof contrastivenessin thecaseof topicalizationin Englishis illustratedby Prince(1984:218)by thedifferencein appropriatenessof examples(13)and(14):

(13) A: You seeevery WoodyAllen movie assoonasit comesout.B: No - AnnieHall I saw (only) yesterday.

(14) A: Why areyou laughing?B: # AnnieHall I saw yesterday. I wasjust thinking aboutit.

Theproperuseof left-dislocationin (15) is alsomotivatedby thefactthattheleft-dislocatedconstituentgallstonesis contrastedwith nervousbreakdown(cf. Prince1984:221):

(15) “Everybodyhastheir little bundle, believe me. I’ ll bet shehada nervousbreakdown. That’s not a goodthing. Gallstones, you have themout andthey’re out. But a nervousbreakdown, it’s very bad ^J^J^ ”

105

Contrastivenessin combinationwith topicality is not only expressedby markedsyntacticconstructionsbut also by intonational patterns. Here the “I-contour”(fall-riseaccent)in Germanor the “B-accent” in Englisharethemostthoroughlyinvestigatedphenomena.Buring (1997: 69) claimsthat recourseto alternativesistypicalnotonly for focusingbut alsofor topicsrealizedby thefall-rise,in thelattercasegiving riseto the“residualtopic”:

(16) Q: Whatdid thepopstarswear?A: The[FEMALE] T popstarswore[CAFTANS]F.

ResidualTopic: Whatdid themalepopstarswear?

On the basisof the datapresentedwe can concludethat the notion of contrastmustbe regardedasa linguistically relevant phenomenon,becausecontrastis anecessaryconditionfor theuseof certainsyntacticandphonologicalmeans- bothin the field of focusingand topicality. The datafrom different languagesshow,however, that this notion is importantto a differentdegree- dependingon whichlanguageand which specificstructureis considered. In thoselanguageswhereformal markingof contrastis only optionalandnot unambiguous(as in Englishor German),the strict distinctionof the contrastive andnon-contrastive casesisempiricallynot sowell-founded- supportingthegradientview of contrast.Froma cross-linguisticperspective it seemswell-motivated,though,to assumethat thenotion of contrasthasa crucial impact on linguistic forms, in many languagesdemandinganobligatoryformal realization.

4 Contrast types

The datagive empiricalevidencealso for the hypothesisthat thereis a needfordifferentiatingthe notion of contrast,becauselanguagesrequirea division be-tweencontrastandnon-contrastwith recourseto differentcriteriapresentedabove.Firstly, thedistinctionbetweencontrastingwithin a closed,limited seton theonehand,andrecoursetoalternativeswithin anopenset,ontheother, seemsmotivated:in the formercase,contrastivenesshasa pragmatic characterandthe realizationof contrastis relevant,whereasin thelattercasecontrastis semanticallyanchoredandoperationon quantificationaldomainsseemsimportant.Secondly, we needtodistinguishdifferentsemanticoperationson quantificationaldomainsandassumedifferenttypesof exclusion.

4.1 Contrast: pragmatic vs. semanticlevel

Thedistinctionbetweenthepragmaticandsemanticlevel of contrastingin thecaseof focusing is motivatedby the syntacticdifferencein their realization. In bothcasesthe focustype involved is “narrow focus”, leaving onepart of thesentencebackgrounded.

106

Movementof anarrow focusinto theleft sentence-peripheralpositionis possi-ble in many languages(in Finnish,Spanish,Italian,Rumanian,Russian),but onlywhenthe relevant set is limited. This type of contrastI will call CONTRAST -usingaspecialnotationwith capitalletters.Thedifferencebetweenthegrammati-cality of theItalianexamples(17)and(18) (cf. Zubizzareta,1998:20)canonly beexplainedby thefact that theexplicit mentioningof therelevantalternative(s)(orat leasttheir saliencein thespeaker’s andhearer’s minds)is a necessaryconditionfor movementinto theleft periphery:

(17) Whoateanapple?

*Gianni hamangiatounamela.‘Gianni haseatenanapple.’

(18) GIANNI hamangiatounamela(nonPiero).‘Gianni haseatenanapple(notPiero).’

For thesemanticallyanchored“narrow focusing”,whererecourseto a closedsetis not absolutelynecessary, the target positionof the focusmovementis obliga-torily verb-related- without the restrictionon the locationon the left-periphery.In Hungarianthepreverbalposition(Spec-F-position)is involved (19) andFocusmovementtakesplacein surfacestructure:

(19) A: Kiwho

evettate

almatapple-Acc

tegnap?yesterday

‘Who ateappleyesterday?’

B1: Tegnapyesterday

[FOCUS Janos]John

evettate

almat.apple

B2: * [FOCUS Janos] tegnapevett almat.

B3: * Tegnapevett almat [FOCUS Janos].

A contextually anchoredcontrastis evenpossiblewithout thesemanticallybasedexclusion. This is typical for contrastivenessin thefield of topicality - expressedby meansof topicalizationandby usingdifferenttypesof left dislocatedstructureswhichareemployed“to markashift in attentionfrom onetoanotherof twoor morealreadyactivetopicreferents”and“oftenhavea‘contrastive‘ function” (Lambrecht1994:183).For this typeof contextually anchoredcontrastI would like to suggestthelabelLD-CONTRAST .

4.2 Semanticdistinctions

In thelinguisticliteraturetwo typesof “exclusion”of alternativesaredistinguished.Most attentionhasbeenpaid to the type of exclusion typical of the HungarianFocusOperator specifiedas“exhaustive identification“(E. Kiss 1998).

107

Theidentificationof onememberof asetperformedby anoperatorfocusin theHungariansentence(20-B)entailsthatall othermembersof thesetareexcluded:

(20) A: Kiwho

jaris going

LundbanLund-in

egyetemre?university-to

‘Who is studyingin Lund?’

B: [PETER]F

Peterjaris going

LundbanLund-in

egyetemre.university-to

‘It is PETERwho is studyingin Lund.’

A further type of exclusion is the operationtype which is performedwhen ele-mentsarerealizedwith the fall-riseaccent.Evenby highlightingwith theprenu-clearrising accentof an“I-contour”, a setof alternativesis inducedandexcluded.Nevertheless,with therisingaccenttheexclusionis “weakened”.Thisaccenttypesignalsthusnot theexclusionof all othermembersof theset(21),but thatthereisat leastonememberfor which thepredication(or apartof it) doesnothold:

(21) [ PETER]CT [ LUNDBAN jar egyetemre]F.‘As for PETER,heis studyingin LUND.’

Also Jacobs(1997)(commentingon Buring’s 1997proposal)emphasizesthat thefirst highlighting of “I-Topikalisierung”containsrecourseto alternatives,but thishasnot thesamestatusastheoneconnectedwith focusing.Thecrucialdifferencebetweenthem is that only after utteranceswith “I-Topics” canquestionsbe leftopen;this is notpossiblewith utterancesthatonly containfocus.

To expressthesimilarity betweentheoperatorfocusandthecontrastexpressedby thefall-rise,it wassuggestedin Molnar (1998)thatbothnotionsareassociatedwith the feature[+ exclusive]. This featureappears,though,in the two casesincombinationwith differentvaluesof anotherfeature[+/- exhaustive]: For thefocusoperatorthefeatures[+ exclusive] and[+ exhaustive] arecharacteristicexpressingthatall alternativesareexcluded,whereasthecontrastive topic containsthecom-bination[+ exclusive] and[- exhaustive] , indicatingtheexclusionof only oneorsomeof therelevantalternatives. Thedifferencebetweenthetwo exclusiontypescan thusbe describedby the oppositionof “strong exclusion” (“all-exclusion”)and“weak exclusion” (“some-exclusion”).

Theinformationstructuralvalueof theweakexclusionis differently judgedinliterature:it is subsumedeitheronly underfocus,called“Contrafocus”by Kenesei(1989), or only under topic - called “S-Topic” by Buring (1997), “I-Topik” byJacobs(1997).I wouldliketo suggestthelabel“ I-CONTRAST ” for thisexclusiontype,emphasizingby this termnot only theimportanceof intonationfor this typeof contrast,but alsothatit cannotberestrictedsimply to topicalityor focusing.

108

(22) CONTRAST TYPES:

focus? topic?closedset CONTRAST LD-CONTRASTopenset FOCUS OPERATOR I-CONTRAST

5 The linguistic statusof contrast

In the lastpartof thepaperI would like to arguethat the linguistic statusof con-trastmustbedifferentiated:contrastivenessis not only a featureof topicality andfocusing,but it seemsplausibleto regardcontrast- in certainlanguagesandcertainconstructions- asanautonomousphenomenonof informationstructure.

This proposalis, however, not uncontroversial.Lambrecht(1994:290)claimsthat “contrastiveness,unlike focus,is not a category of grammarbut the resultofthegeneralcognitiveprocessesreferredto as‘conversationalimplicatures”’.Quitedifferent is Chafe’s approach,accordingto whom “the statusfocusof contrastisdifferent from the statusnew” (1976: 38). Chafeseparatesthe notion of con-trastivenessfrom the notion of focus, on the one hand,claiming that the latterparticipatesonly in the given-new distinction. On the otherhand,he arguesforthe distinctionbetween“real topics” of the topic-prominentlanguagesandcon-trastivenessin theso-called“doublefoci of contrast”.Thelatteris foundin theleftdislocation(23) andtopicalizationstructuresof English(24),containing“possiblepairingsof theatricaleventswith certaintimes”:

(23) As for theplay, Johnsaw it yesterday.

(24) Yesterday, Johnsaw theplay.

Chafeemphasizesthatcontrastivenessin bothcases- in “single focusof contrast”aswell asin “doublefocusof contrast”- shouldberegardedasafurther“packagingphenomenon”besidestopic andfocus. Datafrom several languagesdemonstrate,however, thatcontrastcanoverlapwith topicality andfocusingin differentways.According to the main hypothesisof this paper, from a cross-linguisticperspec-tive, thenotionof contrastcannotbe reducedto a featureservingto differentiatetopicality andfocusing,nor canit be regardedasa notion standingbesidestopicandfocusin informationstructure.As contrast– in certainlanguages,in certainstructures– canbesuperimposedon andcombinedwith thenotionsof topic andfocus,thenotionof contrastshouldbeaddedto theinventoryof relevant informa-tion structuralcategories.Thetwo main instancesof contrastattestedempiricallywhichsupportthis claim will bediscussedbelow.

5.1 I-CONTRAST

As to thephonologicallyrealizedtypeof contrastiveness,the“ I-CONTRAST ”, itis easyto prove that this contrasttype doesmorethansimply identify topicality:theanalysisof theconstituentcarryingthe(fall-)riseaccentin (25) astopic seems

109

plausible,whereasthe contrastmarked by the fall-rise in (26) cannotbe relatedto topicality. In thesecaseonecould only claim a secondaryfocusing,withouttopicality, though:

(25) [Auf der/NEUNundfunfzigstenStraße]T habeich [dieSCHUHEa ]F gekauft.“On fifty NINTH StreetI boughttheSHOES.” (Buring 1997)

(26) Manone

a /MUSSmust

dasthe

Buchbook-Acc

a NICHTnot

mogenlike

(, aberbut

manone

KANN).can

(Jacobs1997)

In Englishtheassignmentof thefall-riseaccentis notrestrictedto intonationstruc-turescontainingtwo pitch accents,but is possiblewith a singlepitch accent.Het-land(2001)claimsthat in thesecases(27) theconstituentwith thefall-riseaccentmustberegardedasthenuclearfocusof thesentence:

(27) Q: Did you feedtheanimals?A: I fed the ˇcat.

[Inf.Focus,fall-rise]

Thefunctionof “I-Contrast” seemsidenticalin all analyzedcases- independentlyof its cooccurencewith topicor focus:it evokesalternativesandis connectedwithopenquestions,motivatingtheanalysisthatI-Contrastis asuperimposednotiononbothtopicalityandfocusing.

5.2 KONTRAST

The secondtype of contrastwhich can be claimedto be an autonomousnotionof informationstructureis KONTRAST in Finnish. In Finnish,the typeof pitchaccent(i.e. the fall-rise contour)is, however, not an obligatory, decisive meansfor expressingcontrastiveness.Here, it is primarily the syntacticpositionthat isinvolved in the formal realizationof contrastiveness. The Finnish datademon-strateclearlythat“the semanticfunctionof contrastmaycutacross,andsupersedethe functionsof topic and focus” (E. Kiss 1995: 6). The KONTRAST-Positioncanapparentlyhostnot only a contrastive focus,but alsoa contrastive topic andappearsin two differentsentencepatternscalledby Vilkuna(1995:249)“TOPpat-tern”and“FOCTOPpattern”.Whereasin theso-calledFOCTOP-sentences(29-b),theKONTRAST-positionis occupiedby thenuclearfocusof thesentence,in theTOP-sentences(30-b), thehighlightedconstituentin theKONTRAST-positionistopic, followedby a furtherhighlightedconstituent,namelythenuclearfocus.Cf.Jarventausta& Molnar (2001):

FOCTOP-sentence(28-b):

(28) A: PekkaPekka

lensiflew

Tukholmaan.to Stockholm

‘Pekkaflew to STOCKHOLM.’

110

B: [KONTRAST Reykjavikiin ] Pekkalensi.‘Pekkaflew to REYKJAVIK.”

TOP-sentence:

(29) a. [KONTRAST Tukholmaan ] Pekkalensi[ FOCUS Finnairilla ],‘To STOCKHOM, Pekkaflew by FINNAIR,’

b. [KONTRAST Reykjavikiin ] (Pekkalensi)[FOCUS Icelandairilla ].‘To REYKJAVIK, Pekkaflew by ICELANDAIR.”

As to the semanticoperationsperformedby KONTRAST in the caseof a con-trastive topicandacontrastive focus,E.Kiss(1998:270)notesthatthey arediffer-entin thetwo cases:while “the contrastive focusexhaustively identifiesthesubsetfor which the predicateholds”, “the identificationperformedby the contrastivetopic is nonexhaustive.” TheFinnishcontrastis, thus,asopposedto theHungarianfocusoperator, semanticallyfairly underspecified(Vallduvı & Vilkuna1998:83).

Thesyntacticanalysisof KONTRASTin Finnishrequiresthatseveralsyntacticandpragmaticfeaturesbeconsidered.Thepositionis left peripheral andcanonlyhostaconstituentif it refersto explicitly mentioned- or at leastcontextually salient- alternativeswithin a closedset. TheKONTRAST-positionis somewhatdifferentfrom thesentence-peripheral CONTRAST-positionof otherlanguages(like Italianor Russian),in that it cannotonly be occupiedby the contrastive focus,but evenby a contrastive topic. The KONTRAST-positionin Finnish– asopposedto theFocus-positionin Hungarian– is notverb-related, whichis alsotypicalof theTopicposition in Hungarian: the differenceis, though,that the Topic position can beiterated,whereasKONTRASTis unique.

The listed featuresrequirethemodificationof thestructuralrepresentationofclausesproposedby Rizzi (1997)within thegenerative framework. Rizzi dissolvestheleft-peripheralcomplementizerlayer- thecategoryof CP- intoanumberof dif-ferentfunctionalprojections,hostingbesidesa free functionalmorphemevariousoperator-like elementssuchastopics,focalizedelements,interrogativeandrelativepronouns.Below thehighestfunctionalphraseForcePwhichprovidesinformationabouttheclausaltype (or the ”specificationof Force”), heassumesa topic-focusfield, which in Italian involvesa FocPsurroundedby recursive TopPs. In RizzisSPLIT-CP-model(1997),however, only theTOPIC-andFOCUS-projectionshostconstituentswith specificdiscourseinterpretations,andits expansionthusseemsnecessaryby introducinga KONTRAST-P in the left peripheryof the sentence(30):

(30) ForcePKontrP TopP*FocPTopP*FinP

TheKONTRAST-projectionin thearticulatedCP-domainis not verb-relatedandcannotbe licensedby the sametype of “Affective OperatorCriterion” (cf. Rizzi1991)astheFocus.Accordingto my proposal,aK-featureshouldbemaderespon-siblefor thelicensingof theconstituentin theSpecifier-positionof theKONTRAST-projection.ThisK-featuremustguaranteetheso-called“discourseconnection”(cf.

111

Haegeman& Gueron1999:536)andrequires(i) theabsolutelyleftmostposition,(ii) the explicit mentioningof relevant excludedalternatives within a closedset,(iii) - or alternatively, in caseof Contrastive topic, thatthediscourseconnectionbewarrantedby thegivennessor highdegreeof “identifiability” of this constituent.

6 Conclusion

The datasurveyed in this paperhave served to show that the notion of contrastis a linguistically relevant phenomenonandcannotmerelybe regardedas “con-versationalimplicature” arising from focusingin certainconversationalcontexts(Lambrecht1994:291).Eventhoughin certainlanguagesformalmarkingof con-trastis optional– thusnotallowing a clear-cut distinctionbetweencontrastive andnon-contrastive cases– empiricaldatafrom agreatnumberof languagesshow thatcontrasthasa crucial impacton linguistic forms, requiringthe obligatoryuseofcertainphonologicalandsyntacticmeans.

It wasalsoarguedthat further distinctionswithin thenotionof contrastitselfseemnecessary, in orderto makeexplicit theimpactof contrastivenessonlinguisticforms: above all the separationof the pragmaticlevel from the semantic- quan-tificationally based- level is motivatedby linguistic data. The different typesofcontrast– realizedin certainpositionsandwith specificaccenttypes– are,how-ever, not directly boundto certaindiscoursepragmaticinterpretations.Contrastis thuscompatibleboth with focusingandtopicality, with the restrictionthat thepreferredrelationsaredependenton thetypeof contrast.Thepossibilitiesandre-strictionsof theco-ocurrenceof the differentcontrasttypeswith topic andfocusarepartlydueto universal,andpartly to languagespecificregularities.

The most importantclaim of the work presentedis that the linguistic statusof contrastmustbe analyzedin a moredifferentiatedway thanhasbeenthecasehithertoin linguistic research:from a cross-linguisticperspective, contrastivenessis not simply a featureof focusingandtopicality. Contrastmustbe established-on thebasisof evidencefrom certainlanguagesandin certainconstructions- asafurthercategory of informationstructure,superimposedontopicandfocus.

Bibliography[1] Bolinger, D.L. 1961.Contrastiveaccentandcontrastive stress.Language 37

(1): 83-96.

[2] Bußmann,H. 1990.Lexikon derSprachwissenschaft. Kroner.

[3] Buring, D. 1997.Themeaningof topic and focus: the 59th StreetBridgeaccent(Routledgestudiesin Germanlinguistics3]. London:Routledge.

[4] Buring, D. 1999.Topic. In P. Boschand R. van der Sand(eds)FOCUS.Linguistic,cognitive, andcomputationalperspectives.Part II Semanticinterpretationof focusphenomena.[Studiesin NaturalLanguagePro-cessing],142-165.New York and Cambridge: CambridgeUniversityPress.

112

[5] Chafe,W. L. 1976.Givenness,contrastiveness,definiteness,subjects,topics,andpoint of view. In C.N. Li (ed.)SubjectandTopic, 25-55.New York,SanFranciscoandLondon:AcademicPress.

[6] Dretske, F. J. 1972.Contrastive statements.ThePhilosophicalReview 81:411-437.

[7] E. Kiss, K. 1995.Discourseconfigurationallanguages.Introduction.In E.Kiss, K. (ed.) Discourse Configurational Languages.3-27. New YorkandOxford: OxfordUniversityPress.

[8] E. Kiss,K. 1998.Identificationalfocusversusinformationfocus.Language74(2): 245-273.

[9] Etxepare,R. 1997.Two typesof focusin Basque.In B. AgbayaniandS.-W.Tang(eds)Proceedingsof the 15th WestCoastConferenceon FormalLinguistics,113-127.Stanford.CSLI. Distributedby Cambridge:Cam-bridgeUniversityPress.

[10] Haegeman,L. andGueron, J. 1999.EnglishGrammar. A generative per-spective. Malden,Mass.:Blackwell.

[11] Halliday, M.A.K. 1967.Noteson transitivity andthemein EnglishII. Jour-nal of Linguistics3: 199-244.

[12] Hetland,J.2001.Contrast,thefall-riseaccent,andinformationfocus.Ms.

[13] Jackendoff, R.S. 1972. Semanticinterpretation in generative grammar[Studiesin LinguisticsSeries2]. Cambridge,Mass.andLondon: TheMIT Press.

[14] Jacobs.J.1988.Fokus-Hintergrund-Gliederung undGrammatik.In H. Alt-mann (ed.) Intonationsforschungen, 89-134. [Linguistische Arbeiten200]Tubingen:Niemeyer.

[15] Jacobs,J.1997.Bemerkungenzur I-Topikalisierung.LinguistischeBerichte168,91-133.

[16] Jarventausta,M. andMolnar, V. 2001.Diskurskonfigurationalitat im Finnis-chenundim Ungarischen.Ms.

[17] Kenesei,I. 1989.Logikus-ea magyarszorend? [Is Hungarianword orderlogical?].AltalanosNyelveszetiTanulmanyok17: 105-152.

[18] Lambrecht,K. 1994.Informationstructure and sentenceform. CambridgeUniversityPress:Cambridge.

[19] Molnar, V. 1998.Topic in focus.On thesyntax,phonology, semanticsandpragmaticsof theso-called‘contrastive topic’ in HungarianandGerman.ActaLinguisticaHungarica vol. 45 (1-2): 89-166.Dordrecht:Kluwer.

[20] Prince,E. 1984.Topicalizationandleft-dislocation:A functionalanalysis.In S.J.White andV. Teller (eds)Discoursesin readingand linguistics.

113

[Annualsof the New York Academyof Sciences433], 213-225.NewYork: New York Academyof Sciences.

[21] Reinhart,T. 1982.Pragmaticsandlinguistics: An analysisof sentencetop-ics.Reproducedby theIndianaUniversityLing. Club.

[22] Rizzi, L. 1991.Residualverb-secondandthe Wh-criterion.Technical Re-ports in Formal and ComputationalLinguistics2 (Faculte desLettres,Universite deGeneve).

[23] Rizzi, L. 1997.Thefinestructureof theleft periphery. In L. Haegeman(ed.)Elementsof grammar, 281-337.Dordrecht:Kluwer.

[24] Rochemont,M.S. 1986.Focusin generative grammar [Studiesin Gener-ative Linguistic Analysis4]. AmsterdamandPhiladelphia:JohnBen-jamins.

[25] Rooth, M.E. 1985.Associationwith focus.Ph.D.diss.University of Mas-sachusettsatAmherst.

[26] Rooth,M.E. 1992.A theoryof focusinterpretation.Natural Language Se-mantics1: 75-116.

[27] Vallduvı, E. 1992.Theinformationalcomponent.New York andLondon.

[28] Vallduvı, E. andM. Vilkuna.1998.”On rhemeandkontrast”.In P.W. Culi-cover andL. McNally (eds)Thelimits of syntax, 79-108.[SyntaxandSemantics,Volume29].

[29] Vilkuna, M. 1995.Discourseconfigurationalityin Finnish. In K. E. Kiss(ed.)Discourseconfigurational languages, 244-268.New York andOx-ford: OxfordUniversityPress.

[30] Zubizarreta,M.L. 1998.Prosody, focus,andword order [Linguistic InquiryMonographs33]. Cambridge,Mass.andLondon:TheMIT Press.

114

Complementsetreference

RICK NOUWEN

UiL-OTSUniversiteitUtrecht

[email protected]

ABSTRACT. Thepaperconsidersthephenomenonof complementanaphoraandoffersananaly-sisin theframework of optimality theoreticsemantics/ pragmatics.I arguethatfrom aninterpre-tationperspective thereis a preferencefor referenceto the so-calledREFSET over all othersetsassociatedwith quantificationalstructures.Referenceto thecomplementsetis thusonly possiblewhentwo conditionsaremet: (i) the complementsetsatisfiesa witness-constraintand(ii) thereferencesetis ruledoutasanantecedentfor independentreasons.

1 Intr oduction

Givena quantificationalstructureD b Ac�b Bc , therearethreesetsthatcanbeassoci-atedwith it1: (i) the referenceset(or REFSET), correspondingto the intersectionof A andB; (ii) themaximalset(MAXSET) which equalstherestrictorA; and(iii)thecomplementset(COMPSET) which is thesetof entitiesin A whichdo nothavepropertyB. Anaphoricreferenceto all thesethreesetsis possible.

(1) a. Most studentswentto theparty.They hadagoodtime. REFSET

b. Most studentswentto theparty.They like to have agoodtime. REFSET/MAXSET

c. Few of thestudentswentto theparty.They wentto thebeachinstead. COMPSET

In (1a), the plural pronounrefers to the set of studentsthat went to the party,the referenceset. Thepronounin (1b) canalsobe interpretedasreferringto thisset,but could alsorefer to studentsin general,which correspondsto a MAXSET-interpretation.Finally, in (1c) we encounterreferenceto thecomplementset.Thepronounhererefersto the studentsthat did not go to theparty. This is a caseofcomplementanaphora.

1Herearesomeformalconventions:In astructureD d Ae<d Be wecall D adeterminer, D d Ae aquantifier, A therestrictorandB thenuclear

scope.A determinerD is monotoneincreasingor upwardentailingin its right argumentif for any C fB andany A, if D d Ae<d Be holds,thenD d Ae<d C e holdsaswell. A determinerD is monotonedecreasingor downwardentailingin its right argumentif for any C g B andany A, D d Ae9d Beih D d Ae<d C e .

115


Whatmakescomplementanaphorainterestingis that their distribution relatesto formal propertiesof the determinerinvolved in the antecedentsentence.Ref-erenceto COMPSET is not possiblefollowing non-downwardentailingquantifiers.This is in contrastwith REFSET reference.

(2) a. Few of thestudentswentto theparty.They wentto thebeachinstead. COMPSET

b. Few of thestudentswentto theparty.They hadagoodtime. REFSET

(3) a. Most studentswentto theparty.#They wentto thebeachinstead. COMPSET

b. Most studentswentto theparty.They hadagoodtime. REFSET

Referenceto COMPSET is badfollowing cardinaldownwardentailingquantifiers.

(4) a. Lessthanhalf of thestudentswentto theparty.They wentto thebeachinstead. COMPSET

b. Lessthantenstudentswentto theparty.??They wentto thebeachinstead. COMPSET

Referenceto COMPSET is betterfollowing partitive cardinaldownward entailingquantifiers.

(5) a. Lessthantwenty-five of thefifty studentswentto theparty.?They wentto thebeachinstead. COMPSET

Thesearethebasicfactsconcerningcomplementanaphora.In section2 we willdiscussthe sourceof the discussionsurroundingcomplementanaphora:a rangeof psycholinguisticexperimentsconductedby Moxey andSanford.Section3 dis-cussestwo analysesthattry to explaincomplementanaphora.In section4 we willpresentsomenew data,focusingon the interpretationof plural pronouns.I willarguethat from this perspective referenceto a complementsetis marked. Finally,in section5 I will give ananalysis.

2 Moxey and Sanford’ s experiments

As we have seenanaphoriclinks to COMPSET aresubjectto otherconstraintsthanreferenceto the othersetsassociatedwith quantificationalstructures.This phe-nomenonhasbeenthoroughlystudiedin a seriesof psycholinguisticexperiments(see(SanfordandMoxey 1993)). In theseexperimentssubjectswereconfrontedwith a singlequantifiedstatementandwereaskedto make up a sensiblecontinua-tion beginningwith theplural pronounThey.

(6) Q of theMPswereat themeeting.They ^J^J^116

Subjectswerealsoasked to indicatewhat the plural pronounreferredto in theircontinuation. Here they could make a choicebetweenfive categories: MPs ingeneral, all MPs, MPswhowentto themeeting, MPswhodid notgoto themeetingandnoneof theabove. Independentjudgescheckedall theutterancesandreferenceindications. In 98% of thesecasesthe judgesagreedwith the judgmentsof thesubjects.

An alternative experimenttestedfor intra-sententialcomplementanaphora,us-ing astructurelike (7).

(7) Q of theMPsattendedthemeeting,becausethey ^J^J^Theresultsshowedapreferencefor COMPSET referencefollowing thedeterminers:hardly any, not many, very few andfew. This preferencewasnot presentwith thedeterminersa few andmany. A specialcasewasonly a few which only showedCOMPSET referencein thetaskinvolving astructurelike (7).

In a different studycomplementset referencewasstudiedwith proportionalnumericalexpressions.The continuationmethodand the useof judgeswas asin the experimentabove (cf. (SanfordandMoxey 1993):77). Here it was foundthat COMPSET continuationswere favored following the determiners:lessthann%. The otherdeterminers(n%, only n% and more than n%) showed few to nocontinuationscontainingcomplementsetreference.

It is importantfor the currentpaperto notethat preferencefor a certaincon-tinuationshouldbeinterpretedratherweaklyasroughly indicatingthatmorethanhalf of thesubjectsusea complementanaphor. That is, REFSET continuationsdidoccurfollowing downwardmonotonequantifiers.

In their analysisMoxey andSanford’s basicassumptionis thefunctionaluse-fulnessof COMPSET. Earlier experimentshave shown that expressionslike few,not many, a few andonly a few roughly specifythe samequantity(see(SanfordandMoxey 1993):ch2).Thereis however a fundamentaldifferencebetweenthesedeterminersin thatsomeof themseemto reporton a deviation from expectation,while othersdonot. Moxey andSanfordlink thisobservationto thenotionof neg-ativity (cf. (Klima 1964)).Negativedeterminers,theargumentcontinues,put focusupontheir COMPSET. Thismeansthatin adiscoursetheCOMPSET is morepromi-nentthanthe REFSET. Therewill bea preferencefor referenceto thesetfocusedby thedeterminer. Moreover, this focuswill have thethematiceffect that,follow-ing a COMPSET-licensing quantificationalstructurea reasonis givenwhy REFSET

is smallerthanwasto be expected.This is confirmedby a seriesof experimentsshowing that COMPSET-continuations generallyare(in Moxey andSanford’s ter-minology) of a reason-why-notnature. In theseexperimentsindependentjudgeswereasked to indicatewhich of four typesa producedcontinuationbelongedto.The COMPSET continuationsweredominantlyclassifiedasindicating“the reasonwhy thepredicateis not trueof therefset”(see(SanfordandMoxey 1993):66).

117

3 Explaining COMPSET reference

A commonreactionwithin the (formal) semanticcommunityon the resultsofMoxey and Sanford’s experimentswas that complementanaphorado not reallyinvolve complementsetreference,but insteadarea caseof somesortof pseudo-referenceestablishedby ageneralizationoverthemaximalset(seeespecially(Cor-blin 1996)).Thisview still enjoysconsiderablepopularityalthoughit createssomefundamentalproblems.For one,thegeneralizationseemsto beallowedsincetheantecedentsentenceexpressesthe inferiority of the REFSET. Problematicthenisthat this generalizationshouldalsobeallowedfollowing a sentencelike Not all AB, wherethe COMPSET might consistof but a few exceptions.Otherauthorshaveproposedothervariantsof pseudo-reference(e.g.(Geurts1997))andthe resultisaninterestingdiscussionon thereality of complementsetreference.Dueto spaceconsiderationsI will not go into thatdiscussionhereandpresupposethereality ofcomplementanaphora.(For a detailedelaborationon this issueaswell asmoreargumentsagainstpseudoreferencecf.(Nouwen2001)).

Oncewe acceptthat complementanaphorainvolve referenceto the comple-mentset,thequestionariseshow thisanaphoriclink comesabout.

3.1 Emptiness

In their presentationof optimality theoreticsemantics(de Hoop and Hendriks2001)suggestthat thedatacanbeexplainedby the interactionof pragmaticcon-straints.

The basicassumptionbehindoptimality theoreticsemanticsis the free inter-pretationhypothesis.This is nothingbut the driving force behindthe generatorof theoptimality theoreticsystemof finding themostoptimal interpretationfor acertainlinguistic form. Relevantto ourstoryis thatgiventhishypothesisageneral-izedquantifiercantake any domainof quantification.ThusREFSET referenceandCOMPSET referenceof apronouncompetein thecandidatesetof theinterpretationof acontinuationfollowing aquantificationalstructure.

Thebasisof deHoopandHendrik’s explanationis theconstraintEMPTINESS.

(8) EMPTINESS: As theantecedentof ananaphoricexpression,do not chooseasetthatis or maybeempty

By itself, this constraintcan alreadyexplain the correlationbetweendownwardmonotonicityand complementset reference. The referenceset of a downwardmonotonequantifiercan,of course,beempty. Referenceto thissetwill thusviolateEMPTINESS. Theviolationpatternis reversedfor monotoneincreasingquantifiers.

Theconstraintisespeciallyinterestingwith respecttocardinaldownwardmono-tonequantifiers.In thosecasesbothREFSET andCOMPSET canbeempty:thefirstdueto downwardmonotonicity;theseconddueto thefactthatfor all thesequanti-fiersD(A)(A) holds.Both interpretationsthusviolateEMPTINESS. Theoptimality

118

decisionis now left to a constraintwhich is ranked lower. De Hoop and Hen-drikshaveaconstraintcalledforward directionality(or FORWD), whichexpressesapreferencefor referringto theREFSET.2

Thereis however acomplication.Emptinessis not theonly constraintin com-petitionwith forwarddirectionality. Thereis alsothehighrankedavoidcontradic-tion or AVOIDC. If ranked higherthenforwarddirectionality, avoid contradictioncanexplain the factsin (9). In the secondsentencein (9a), the determineris in-terpretedas quantifying over the referenceset of the precedingsentence.Onlywhena contradictionis encountered,asin (9b), doesinterpretationpreferreturn-ing to the larger domain, namely that of students(cf. (de Hoop and Hendriks2001):(28)/(32)).

(9) a. Tenstudentsattendedthemeeting.Threespoke.

b. Tenstudentsattendedthemeeting.Twelvespoke.

This additionalinteractioncausesproblemsfor theanalysisof cardinaldecreasingquantifiers.Aswehaveseen,boththereferencesetandthecomplementsetof thesestructuresviolateemptiness.Forwarddirectionalitythenprefersthereferenceset.But in theexamplesof interest,like (4b) repeatedbelow, avoid contradictionhasanunwanteddecisive preferencefor thecomplementset.

(4b) Lessthantenstudentswentto theparty.??They weretoobusy.

Thereis anotherproblematicsideeffect of their proposal.Thereseemsto be noroomfor optionality. Rememberthattheresultsfrom Moxey andSanford’s exper-imentsdid not show that REFSET referencedoesnot occur following downwardmonotonequantifiers.They merelyshowedthereto beapreference.Theoptimal-ity modelproposedby deHoopandHendrikssuggeststhatREFSET continuationswill never follow non-cardinaldownwardentailingquantifiers.

We will return to de Hoop andHendrik’s analysisbut next consideranotherexplanationof thecomplementanaphoraphenomenon.As we will seein this ap-proachoptionalitydoesplayanimportantrole.

3.2 Dynamic Quantifiers

RodgerKibble ((Kibble 1997a))givesa technicalexplanationfor Moxey andSan-ford’s experimentalresults. In dynamicsemantics,quantifiersaremadedynamicby giving themanexistentialstructure,asin (10).

(10) Q b Ac�b Bckjml x nol y n maxy b A p xq yr�c@n maxx b x s y n Bctn Qu-b yc�b xcThis saysthat a quantificationalstructureQ b Ac�b Bc is to be interpretedasthe in-troductionof two (maximal)sets,onesatisfyingthe restrictorA andthe other, a

2This constraintis actuallya muchmoregeneralconstraintexpressingthepreferencefor anon-goingreductionof topic range.

119

subsetof the former, satisfyingthescope.Thequantificationsucceedsif the twointroducedsetsarein the quantificationalrelationQu (the extensionof the deter-miner Q). The effect is that every quantificationalstructureintroducesMAXSET

andREFSET into thecontext. A well-known problem,however, is that this mech-anismonly worksfor MON v determiners.Becauseof theexistentialnatureof theconstruction,it merelydesiresthereto besomepair of maximalsetssatisfyingre-strictorandscopeandtheQu -relation.It doesnotenforceall pairsof maximalsetssatisfyingA andB to bein this relation.Kibble takesthisproblemto bethesourceof complementanaphora.A naturalway of solving the problemis by definingMON

Wdeterminersin termsof their duals.Thereare,however, two possiblecon-

structionsof dualsfor eachdeterminerD: onebasedon theso-calledcomplementw D andonebasedonthecontradualD w . Thefirst canbecomparedwith ordinarywide scopenegationof the quantificationalstructure.This way, the complementof few b Ac�b Bc will be not few b Ac�b Bc or manyb Ac�b Bc . To returnto few we simplynegateoncemore: not manyb Ac�b Bc . Now we have an alternative to few in termsof the(unproblematic)MON v determinermany. Theotherway of doingthis is byusing the contradual:few b Ac�b notBc . It is easyto seethat D w (just like w D) isMON v whenever D is MON

W. Negatingoncemoregivesusthesecondalternative

to few. All this is illustratedin (11).

(11) Few of thestudentswentto theparty.

a. Not many of thestudentswentto theparty.

b. Many of thestudentsdid notgo to theparty.

Using theseinsightsto cometo a dynamicinterpretationfor MONW

determinersresultsin anambiguity. Whenthecomplementis used,thedynamicstructurein-troducesthereferencesetA x B into thecontext. But whenthecontradualis usedthereferencesetwill beA w B. This,accordingto Kibble, explainswhy (andwhyonly) weencountercomplementanaphorawith downwardentailingquantifiers.

At first sight,Kibble’s solutionto thecomplementanaphorpuzzlemight seema sideeffect of a technicaloddity. But thereis an importantinsight hiddenbe-hindall this. Thecorrelationbetweenmonotonicitypropertiesof adeterminerandits possibility to licensecomplementsetreferencefollow from logical inference.As Kibble putsit: “[A] plural pronouncanpick up anantecedentwhich is eitherexplicitly introducedor logically inferablefrom antecedentinformation,providedthat the referentfunctionsasa witnessset which serves to verify the antecedentsentence(s).” ((Kibble 1997b):126).Thus,while a smallproportionof MPsbeingabsentis notawitnessfor MostMPsattendedthemeeting(therecouldbemoreofthem),a largesetof absenteeswill verify thatfew of theMPsattendedthemeeting(Kibble refersto suchwitnessesasnegativewitnesses). With cardinalquantifiersthereis no way of knowing whethera large set verifies the antecedentsentenceor not, simply becausethe domainis not known. Thus,COMPSET canonly be awitnesssetfor adownwardentailingproportionalquantification.

Thereis however a gapbetweenthe formal explanationandKibble’s concep-tual story involving witnesshood.Formally, nothingpreventsusfrom forming the

120

dualsof amonotoneincreasingquantifierandsuchadualwouldalsooutputacom-plementset. The only reasonwhy it is not thereis becauseMON v quantifiersdonot have thesameproblemswith existentialstructureasthedownwardoneshave.On the conceptualside, then, it is not clearwhat it meansfor a set to verify itsantecedentsentenceandwhy we wouldn’t assumesucha setto bemaximal. OnealsocanstartwonderingwhetherREFSETs arealwayswitnesses.Following Kib-ble’s analysisa smallsetof partygoingstudents,for instance,would not verify asentenceexpressingthatfew of thestudentsattendedtheparty, but still pronominalreferenceto this set is possible.So apparentlyreferencesetshave unconditionalwitnesshood.Moreover, a set of studentsnot going to a certainparty can onlycountasa negative witnessfor few of the studentsattendingthat party, oncewehave establishedthatthis setis a largeproportionof thewholesetof students.Butin orderfor us to know that, we have to know the cardinalityof the (contextual)domainof students.Genuinecomplementsetsarethusalsoin needof cardinalityinformation. But exactelythis waswhy Kibble excludednon-proportionaldeter-minersfrom producingnegative witnesses.

AlthoughKibble makesclearthatsomesenseof witnesshoodis neededto ex-plain complementsetreference,his formulationof this key notionseemsto raisea lot of questions.In the next section,however, we will seethat accountingforthe classof COMPSET licencingdeterminersis not the only problemfor the twoanalyseswehave considered.

4 COMPSET fr om the hearer’sperspective

Noticeadetailin bothHendriksanddeHoop’s analysisandthatof Kibble’s. Bothimplementthe insightsin a framework meantto describenaturallanguageinter-pretation,while Moxey andSanford’s datashows usprimarily factsof production.Althoughthis dataclearlyhasrepercussionsfor comprehension(e.g.theexpecta-tion of acertainkind of referencefollowing acertaintypeof determiner),I will tryto show herethatfrom aninterpretationperspective thedatais moresubtle.

WehavealreadyseenthatMoxey andSanfordfoundathematiceffectin COMPSET

continuations:they usuallyspecifythereasonwhy a relatively largeproportionofthedomaindid not satisfythepredication.Fromaninterpretationperspective it isinterestingto seewhat happensif we usecomplementanaphorain continuationsother than the thematicpreferencefound by Moxey and Sanford. Considertheexamplesin (12).

(12) a. Few of thestudentswentto theparty.I know who they are.

b. Few of theAmericanpresidentsin the20thcenturywerein power fortwo consecutive terms.My historyteachermademelearntheirnamesby heart.

c. Few of theseballsareblue.Canyoupoint themout for me?

121

In all thesediscoursesthepredicationsin thefirst andsecondsentenceareneutralwith respectto oneanother. Resolvingthe plural pronounshows a clearprefer-encefor REFSET reference.It appearsthenthatthedefault interpretationfor pluralpronounsis the intersectionof restrictorandscopeno matterthe formal proper-tiesof thedeterminer. As far asI know, this asymmetrybetweenproductionandinterpretationhasnotbeennoticed.

Thepoint madehereis thatREFSET is (if possible)thepreferredresolutionofa plural pronoun.Thecomplementsethasno suchstatus.During resolutionit isoverruledby thereferencesetin neutralsituations.It seemsto methat COMPSET

interpretationis the result of a last resortstrategy. Notice that a side effect ofreason-why-notcontinuationsis that resolvingthe plural pronounto refer to theREFSET resultsin a contradiction. The only non-contradictoryresolutionis, ofcourse,theCOMPSET.

In (Moxey andSanford1987),we find anexampleshowing that reason-why-not continuationsarenot obligatory. In (13), thereis a clearcaseof complementsetreference,but thefactthatthemembersof COMPSET sendtheirapologiesdoesnot really indicatewhy sofew MPswereat themeeting.

(13) Few MPswereat themeeting.They sentapologiesfor beingabsent.

Still, onceagainwe seethat resolvingthe plural pronounto REFSET referencewould resultin acontradiction.

Notice how both accountsgiven in the previous sectionarenot able to copewith thegeneralpreferencefor REFSET reference.DeHoopandHendriks’analysiswill alwayspreferthenon-emptysetover the(possibly)emptyone,no matterthepredicationsinvolved (i.e. ignoringtheproblematicinteractionwith avoid contra-diction,but moreon thatlater).Following Kibble’s story, COMPSET is a referenceset,sothedesireddifferencebetweenthetwo setsdoesnotexist.

More supportfor theview that the interpretationof anaphorashows a prefer-encefor REFSET comesfrom explicit referenceto thecomplementset.Noticethefollowing contrast.

(14) a. Few of thestudentswentto theparty.Theothersstayedat homeinstead.

b. Few of thestudentswentto theparty.Theothershadagoodtime. non-party-goersyzy ??party-goers

In (14a)we seethatwe canreplacethecomplementanaphorwith anexplicit ref-erenceto thecomplementset: theothers. This definitedescriptiontakesthecom-plementof theREFSET relative to somedomainof quantification(theMAXSET). Ifwe take Kibble’s analysisserious,thereis really no differencebetweenCOMPSET

andREFSET otherthanthat they arethereferencesetof dynamicallydifferentbutlogically equivalentrepresentationsof a quantificationalstructure.In otherwords,thereis no apparentreasonwhy theothers would not acceptthe COMPSET asanantecedentreferenceset.Thishowever givesustheodd(14b).

122

Summarizing,for the interpretationof plural anaphora,we find the followingparadigmin termsof emptiness.(A) correspondsto monotoneincreasingcontexts;(B) to cardinaldownward monotoneonesand(C) representsthe caseswith trueCOMPSET licensers(proportionalmonotonedecreasingdeterminers).

(A) COMPSET possiblyempty, REFSET non-empty: * COMPSET/REFSET

(B) COMPSET andREFSET bothpossiblyempty: * COMPSET/REFSET

(C) REFSET possiblyempty, COMPSET non-empty: REFSET y COMPSET

5 Analysis

Thefirst questionarisingfrom all this is how to derive thepotentialcancellingofthepreferencefor REFSET. Oneway to accountfor this is to returnto anoptimal-ity approachandaddthehigh-ranked constraintavoid contradiction (henceforth:AVOIDC). This constraintseemsfit for accountingfor thelast-resortstrategy nec-essaryfor resolvingpronounreferenceto COMPSET. But it is easyto seethattherecanbeno rankingof AVOIDC, FORWD andEMPTINESS suchthat it accountsforthedata.Thereasonis thatif werankEMPTINESS higherthanAVOIDC wegetnopreferencefor REFSET whenit is potentiallyemptybut lackingcontrastin predi-cation.Sowouldwenow chooseto rankthetwo constraintstheotherwayaround,thenwe predictCOMPSET readingsin caseslike (3a),repeatedhere.

(3a) Most studentswentto theparty. #They wentto thebeachinstead.

Theobviousproblemis thatEMPTINESS isonlysupposedtobedecisivefor COMPSET

reference.As wehaveseen,weactuallypreferreferringto thepossiblyemptyREF-SET overreferenceto theguaranteednon-emptyCOMPSET. ThustheproperroleofEMPTINESS appearsnot to beageneralconstraintonreferenceatall, but it appearsto tell uswhichsetsarepotentialantecedentsandwhicharenot.

Here we stumbleon an important relation betweenemptinessand witness-hood.Bothshow usthelogicalaccessibilityof aset.EMPTINESS seemsto excludepossiblyemptyreferencesets.Wealsosaw thattheformalizationof witness-hoodpredictedthatREFSETs arealwaysavailableasanantecedent.Let usthereforere-placeEMPTINESS with a new constraint,a reformulationof what it meansto beawitness:

(15) WITNESS: Be a witness,i.e. asanantecedentof ananaphoricexpressionchoosean accessiblereferentor choosea constructedreferentwhich isguaranteedto benon-empty.

The intuition behindthis constraintis that referenceto an introducedreferentislessdangerousthanreferenceto a constructedone,sincethepropernessof suchaconstructionis not guaranteed.Kamp andReyle proposedthat in discourserep-resentationtheorysubtractionis not a permissibleoperationfor theforming of an

123

WITNESS AVOIDC FORWD

Most(A)(B). They [ B REFSET *Most(A)(B). They [ B COMPSET * *

Most(A)(B). They Cneutral REFSET

Most(A)(B). They Cneutral COMPSET * *

Few of the(A)(B). They [ B REFSET *Few of the(A)(B). They [ B COMPSET *

Few of the(A)(B). They Cneutral REFSET

Few of the(A)(B). They Cneutral COMPSET *

Lessthanten(A)(B). They [ B REFSET *Lessthanten(A)(B). They [ B COMPSET * *

Lessthanten(A)(B). They Cneutral REFSET

Lessthanten(A)(B). They Cneutral COMPSET * *

Figure13.1:Tableaufor theparadigm

antecedent(see(KampandReyle 1993):307).(Kibble 1997b)alreadysuggestedthat this is a far too stronga constrainton antecedentformationandproposedhiswitnessalternative. I proposeto alterthisagaininto aconstraintwhich is morelikeHendriksanddeHoop’s emptiness,with thekey differencethat it is only appliedto constructedplurals.Following thecurrentconstraintWITNESS andgivensomesortof ontology(like for instancetheuppersemi-latticestructureusedby (KampandReyle 1993)) summationof referentsis alwaysallowed, subtractionis onlyallowedonceit is defined.

WITNESS is thusapowerful constraintonanaphoricreference.RankingAVOIDCover FORWD accountsfor thedatain theprevioussection,i.e. only contradictivemeaningcanforceusto interpretapronounasreferringto acomplementset.From(3a) it follows that if thecomplementsetis not a suitableantecedentdueto WIT-NESS we shouldacceptany readingresultingfrom resolvingthepronounto REF-SET, even a contradictive one. Hence,AVOIDC is ranked below witness-hood.3

Theeffectsof this rankingareillustratedin thetableauin 13.1.WITNESS rules out all referencesto COMPSET whereCOMPSET is possibly

empty. This only leavestheproportionaldownwardentailingquantifiers.In gen-eral REFSET referenceis preferred,but this can be overruledin the non-neutralcases,wherethepredicationin thecontinuationcontradictsthepredicationin theantecedentsentence.

Following a downward entailingcardinalquantifier, WITNESS is violatedbyCOMPSET reference.This is becausethesetA is unknown andthusD b Ac�b Ac couldbevalid. In theintroductionI mentionedthatfollowing thesequantifiersCOMPSET

referenceis notplainly out,but bad. Thereasonfor this is thatcontraryto increas-3See(deHoop2000)for moreonunintelligibility, optimality theoryandavoid contradiction.

124

ing quantificationalstructuresthereis anindirectway thecomplementsetcanstillbea witness.Sometimes,thereis an implicit domainavailabletriggeringa parti-tive readingandmakingthecomplementsetinto a negative witness.Considerforexample(16), wherethe fact that we know how many nomineestherearefor anOscarcategory (five)providestheimplicit partitive reading.

(16) Lessthantwo nomineesfor theOscarfor bestactorworea tuxedo. Theysimplyworeasuit andtie.

Moreover, it seemsto me that sometimesa domaincanbe accommodated.Con-sider(17):

(17) Lessthantenstudentshandedin theiressay.They hadall sortsof excuses.

Wedonotknow how many studentsthereare,but thereis norealreasonto assumethereonly to beten. Thecontinuationseemsto provide evidencefor theexistenceof morethantenstudents.Thisallows theCOMPSET to beawitnesssetfor thefirstsentence.

Let usbriefly considerthecaseof theothers. Noticefirst how it differsfrom acomplementanaphorin thatit canoccurfollowing increasingenvironments.

(18) a. Most studentswentto theparty. Theothersstayedin.

b. Lessthantenstudentswentto theparty. Theothersstayedin.

It seemsplausibleto characterizetheothers assignalinga shift from thepreferredreferenceto its complement.Thisbehavior followsdirectly from ouranalysisoncewe assigntheothers a semanticslike: λPP b X w Ac , whereX is a context set(themaximalset)andA is theantecedentfor theothers. Notice that WITNESS, sinceit is a constrainton antecedent-hood,alreadyblocksCOMPSET asa potentialan-tecedentfor X in increasingandcardinalenvironments.Whatis left to beexplainedis thefactthattheothers cannotreferto REFSET evenin theremainingcases.Oneway to go would be to saythat forward directionality(like WITNESS is andlikeemptinesswas)is aconstrainton antecedents,i.e. it prefersanaphoricexpressionsto selecta forward directionalantecedent.At first, this might seema breakwiththeoriginal intentionof FORWD, viz. reductionof topichood,but noticethatsuchabreakonly occursin casetheanaphoricexpressionexpressesashift of somesort.Of course,that is exactly whatwe want. I leave it to the readerto checkthat thecurrentproposalindeedcoversall thedata.

6 Conclusion,reflection,further research

Complementsetsaremarkedantecedentsfor plural pronouns.Their constructionis only permittedoncethey areguaranteedto benon-empty, but even in this caseresolutionto COMPSET only occursonceREFSET referenceis out for independentreasons.

125

Theinterpretationof pronounsreferringto a partof a quantificationaldomainnow sumsupto thefollowing: interpretthepronounasareferentor asanecessarilynon-emptyconstructedreferent;next –if thereis a choice–interpretthe pronounin sucha way that therearisesno contradiction;finally, if therestill is a choice,chooseREFSET over all otheralternatives.

Reflectingon this resolutionstrategy we couldsaythat the taskof optimalitytheoryhereis nothingbut mediatingbetweendifferentmodulesof languageuse.WITNESS couldthusbeseenasahardconstraintsayingthatweshouldonly refertothatwhich is availableor safelyinferable.Makingsenseis thusonly possibleoncewe have obeyedwitness-hood.Finally, pragmaticpreferencefor a certaintypeofinformationflow is only importantoncewe have anintelligible interpretation.

I have remainedquieton onecomplicatingissue.It concernsa problemwithreferenceto themaximalset. In (19) we prefera maximalsetreadingover a reso-lution to COMPSET.

(19) Not all of theanimalsin thiszooaredangerous.More thanhalf canbestroked.

Clearly, morethanhalf of theanimalscanbestroked,notmorethanhalf of thenon-dangerousones. If we assumethat the maximalset is presentasa referent,thenfollowing ouranalysispresentedherewepredictMAXSET referenceto beimmuneto WITNESS. This,however, posesthequestionhow MAXSET andCOMPSET areincompetitionin exampleslike (19). A possiblesolutionwould beto assumethatinadditionto forwarddirectionality(violatedby bothCOMPSET andMAXSET) thereis a constraintwhich representsthe costsof constructinga new referentout ofexisting ones.This would accountfor exampleslike (19). However, thereis littledataknown involving maximalsetreference,sothedetailswill have to bepartoffurtherresearch.

Bibliography

Corblin, F. (1996). Quantificationet anaphorediscursive: la referenceauxcomplementaires.Langages 123, 51–74.Also as: Peut-onanaphoriserlecomplementaire d’un ensemble? In: Referencetemporelleet nominale.Actes du 3e cycle romand de Sciencesdu langage,Cluny, avril 1996.Moeschler, J.andBeguelin,M. (eds.).

deHoop,H. (2000).Theproblemof unintelligibility in OT semantics.unpub-lishedmanuscript.

deHoop,H. andP. Hendriks(2001,February).OptimaliltyTheoreticSemantics.LinguisticsandPhilosophy24(1), 1–32.

Geurts,B. (1997).Book review of Linda M. Moxey andAnthony J. Sanford.CommunicatingQuantities.1993.Journal of semantics14(1), 87–94.

Kamp,H. andU. Reyle (1993).FromDiscourseto Logic. Dordrecht:D. Reidel.

126

Kibble, R. (1997a).Complementanaphoraanddynamicbinding.In A. Lawson(Ed.),SALT VII, Ithaca,NY: CornellUniversity.

Kibble, R. (1997b).Complementanaphoraandmonotonicity. In G. Morrill, G.-J.Kruijf f, andR. Oehrle(Eds.),Formal Grammar, pp.125–136.

Klima, E. (1964).Negationin english.In J.FodorandJ.Katz (Eds.),Thestruc-ture of language. EnglewoodCliffs, New Jersey: PrenticeHall.

Moxey, L. andA. Sanford(1987).Quantifiersandfocus.Journalof semantics5,189–206.

Nouwen,R. (2001).On the reality of complementsetreference.Unpublishedmanuscript.UiL-OTS,Utrecht.

Sanford,T. andL. Moxey (1993).Communicatingquantities.A psychologicalperspective. LaurenceErlbaumAssociates.

127

DiscourseStructur eand Anaphoric Accessibility

MASSIMO POESIO

Universityof Essex andUniversityof Edinburgh

[email protected]

BARBARA DI EUGENIO

TheUniversityof Illinois atChicago

[email protected]

ABSTRACT. We evaluatedthe predictionsof GroszandSidner’s theoryof anaphoricaccessi-bility usinga corpusof tutorial dialogueswhosediscoursestructurewasannotatedaccordingtoRelationalDiscourse(RDA) Analysis. We found supportfor MoserandMoore’s proposalthatonly segmentswith an intentionalcoreshouldbe viewed asintroducingnew focusspaces;wealsofoundthatbothembeddedcoresandembeddedcontributorsshouldremainopenaslong asthe RDA-segmentin which they occur is open,anddiscussthe implicationsof this finding forGroszandSidner’s theory.

1 Intr oduction

In this paper, we presentthe resultsof anempiricalstudyof therelationbetweendiscoursestructureandanaphoricaccessibility. Seminaltheoriessuchas(Reich-man 1985; Groszand Sidner1986; Fox 1987) have beenaroundfor aboutfif-teenyears. However, only now is it possibleto subjectthemto rigorousempiri-cal testingbecauseof therecentimprovementsin annotationmethodologyleadingto morereliableannotationtechniques,andthe resultingincreasedavailability ofcorporaannotatedfor discoursestructure(Carletta,Isard,Isard,Kowtko, Doherty-Sneddon,andAnderson1997;Moser, Moore,andGlendening1996;Marcu1999).

Much of this recenteffort on discoursestructureannotationhasbeenbasedon RhetoricalStructureTheory(RST) (Mann andThompson1988)or on similartheoriesthat restrict the rangeof possiblestructures. One reasonfor this isthat (eachversionof) RST specifiesa fixed (althoughfairly large) repertoireofintentionsto choosefrom. While it is still anopenquestionwhetherit is possibleto comeupwith anexhaustive list of discourseintentions(GroszandSidner1986),in orderto achieve reliableannotationit is essentialto do so.

As a result,however, thecorporaannotatedin thisway have mainlybeenusedto testnew theoriesof anaphoricaccessibilitythatbuild directly on RST, suchas

129


the recentlyproposedVeinsTheory (Cristea,Ide, andRomary1998). Our ownaim,at leastat this stage,is not to developanew theoryof anaphoricaccessibility,but to studythe empiricalvalidity of existing theoriesof anaphoricaccessibility,andspecificallyof thebest-known amongthese,GroszandSidner’s theory(1986).This usedto bea problembecausealthoughGroszandSidner’s theoryhasorigi-natedacodingmanual(Nakatani,Grosz,Ahn, andHirschberg 1995)thathasbeenusedat leastonce(Nakatani1996),asfar aswe know thereis no sizablecorpuscodedaccordingly. However, recentproposalsconcerningthe mappingbetweenrhetoricalstructureand intentionalstructuresuchas (Moser and Moore 1996b)producedannotatedcorporathat canbe usedto investigatethe predictionsof thetheory.

TheSherlockcorpuscollectedat theUniversityof Pittsburgh is acasein point.This corpusis a collectionof tutorial dialoguesannotatedaccordingto RelationalDiscourseAnalysis(RDA), a theoryof discoursestructurethatattemptsto mergeRST with Groszand Sidner’s theory. One of the featuresof RDA is to distin-guish rhetoricalrelationsinto intentionaland informational(Moore andPollack1992;MoserandMoore1996b):intentionalrelationspertainto theeffectsthatthespeaker intendshis discourseactionsto have on thehearer, whereasinformationalrelationspertainto domainrelationsbetweentheentitiesbeingtalked about. Forexample,amongRST relations,evidenceis intentional,whereascauseis informa-tional (MooreandPollack1992;MoserandMoore1996b).

In this paper, we reporton theresultsof a studyof anaphoricaccessibilityus-ing this corpus.We first briefly review bothGroszandSidner’s focus-spacebasedtheoryof theattentionalstateandRDA. We thendiscusshow RDA structurescanbe usedto specifyopeningandclosingoperationson focusspaces.In theevalu-ationpartof thepaper, we presentfirst of all our resultsconcerningtheeffect onanaphoricaccessibilityof the moredistinctive aspectof RDA, the distinctionbe-tweenintentionalandinformationalrelations,andof MoserandMoore’s proposalthat only the former expressDSPs. We thenlook at a few aspectsof the relationbetweenGroszandSidner’s theoryand RDA left openby MoserandMoore, inparticular, how embeddedsegmentsof differenttypesshouldbe treated.Finally,we analyzetheresultsandevaluateGroszandSidner’s proposal.

2 Background

2.1 Groszand Sidner’s Theory

Accordingto GroszandSidner(G& S), thestructureof a discourseis determinedby the intentionsthat the peopleproducingit intend to convey, or DISCOURSE

SEGMENT PURPOSES (DSPs). In acoherentdiscourse,theDSPsarerelatedto forman INTENTIONAL STRUCTURE by eitherdominance relations(in casea DSP isinterpretedascontributing to thesatisfactionof anotherintention)or satisfaction-precedesrelations(whenthesatisfactionof an intentionis a preconditionfor thesatisfactionof asecondone).

130

Anaphoricaccessibilityof entitiesin a discourseis modeledby its ATTEN-TIONAL STRUCTURE, which, accordingto GroszandSidner, is a stackof FOCUS

SPACES. G& S proposethatwhenasegmentis open,its correspondingfocusspace,whichincludesthediscourseentitiesintroducedin thatsegment,is pushedontothefocusstack;whenthe segmentis closed,the focusspaceis popped,andthe dis-courseentitiesassociatedwith that focusspacearenot accessibleany more. G& S

alsoarguethat the pushingandpoppingof focusspaceson the stackreflectstheintentionalstructure,in the sensethat a new focusspaceis pushedon the stackwhenever thediscourseintroducesa new DSP subordinateto thepresentone,andthefocusspaceof thecurrentis poppedwhenever theassociatedDSP is satisfied.

This claim aboutanaphoricaccessibilitywasillustratedin the original paperwith a few examples;however, as far as we know, it hasnot beenempiricallytested.Thereareno sizablecorporaannotatedaccordingto both G& S’s treatmentof discoursestructureandanaphora. More in general,thereareno guidelinesabouthow to identify the DSPs in a discourse.Our purposeis thereforetwofold:to useRDA to makemorespecificclaimsabouttheDSPs in discourses(of acertaingenreandin agivendomain),andthento testG& S’s claims(within thisgenre).

2.2 Relational DiscourseAnalysis (RDA)

RelationalDiscourseAnalysis(RDA) (MooreandPollack1992;MoserandMoore1996b)synthesizesGroszand Sidner’s approachand RST. RDA owes to GroszandSidnerthe ideathat discourseis hierarchicallystructured,andthat discoursestructureis determinedby intentionalstructure;eachRDA-segmentoriginateswithan intentionof the speaker. But in RDA segmentshave additionalinternalstruc-ture: eachsegmentconsistsof oneCORE, i.e., thatelementthatmostdirectly ex-pressesthespeaker’s intention,andany numberof CONTRIBUTORS, theremainingconstituentsin thesegment,eachof which playsa role in servingthepurposeex-pressedby the core. The notionsof coreandcontributor derive of coursefromthe notionsof nucleusandsatellitein RhetoricalStructureTheory (RST) (MannandThompson1988),which claimsthat in each“segment” (text span,for RST)onecomponentshouldbeidentifiedasthe’main’ one,andtheothersassecondary.However, in RST thereis a distinctionbetweennucleusandsatellitefor (almost)all RST relations,whereasin RDA core and contributors are only identified if asegmentpurposehasbeenrecognized.In this case,eachcontributor is linked tothecoreby oneintentionalrelation,andoneinformationalrelation.This is unlikeRST, in which only a singlerelationcanobtainbetweennucleusandsatellite—see(MooreandPollack1992).

In RDA, segmentconstituentsmay in turn be other embeddedsegments,orsimplerfunctionalelements:theseelementsmay be eitherbasicUNITS, i.e., de-scriptionsof actionsandstates,or relationalCLUSTERS. Clustersarespansthatonly involve constituentslinked by informationalrelations;no core:contributorstructureexists,but they canthemselvesbeembedded.

Unlike G& S’s theoryandlike RST, RDA is basedonafixednumberof relations;

131

1.1 Beforetroubleshootinginsidetheteststation,

1.2 it is alwaysbestto eliminateboththeUUT andTP.

2.1 Sincethetestpackageis movedfrequently,

2.2 it is proneto damage.

3.1 Also, testingthetestpackageis mucheasierandfaster

3.2 thanopeningup teststationdrawers.

1.1 1.2 2.1 2.2 3.1 3.2

step1:step2Cause:effect

prescribed-act:wrong-act

Enable

Act:ReasonAct:Reason

ConvinceConvince

Figure14.1:A tutorial excerptandits RDA analysis

in particular, RDA assumesfour intentionalrelations– convince, enable, concede,joint –anda larger setof informationalrelations;this latter set is expectedto bedomaindependent.In the Sherlockcorpus,23 informationalrelationsareused,of which 13 pertainto causality(they expressrelationsbetweentwo actions,orbetweenactionsandtheir conditionsor effects)(Moser, Moore, andGlendening1996).

Figure14.1shows a small excerptfrom oneof the dialoguesin the Sherlockcorpus,andits correspondingRDA analysis.Thetext is broken into clauses(UUTis “Unit under test”, TP is “test package”). The analysisshows the text to beanalyzedasanintentionalsegmentwhosecorespans1.1and1.2.Thissegmenthastwo contributors,spanning2.1and2.2,and3.1and3.2 respectively. Graphically,thecoreis at theendof the arrow whoseorigin is thecontributor; moreover, thelink is marked by two relations,intentional(in bold), and informational. In thisspecificcase,the two contributors carry the sameintentionaland informationalrelationsto the core,but this doesn’t needto be the case. The coreandthe twocontributorsarefurtheranalyzed.Thecoreandthesecondcontributorareanalyzedasinformationalclusters,whereasthefirst contributor is recognizedashaving itsown intentionalstructure.1 Clustersaremarkedby oneinformationalrelation,but

1According to the manualusedfor the annotation(Moser, Moore, andGlendening1996), anenablerelationholds“if thecontributor [2.1] providesinformationintendedto increasethehearer’s

132

notby intentionalrelations.We will comeback to the analysisof the text accordingto G& S in the next

section.As regardstheanalysisof the text accordingto RST, thestructurewouldpresumablybe thesame,althoughno doublerelationswould exist, andfor everyrelationonerelatumwould beconsideredasthenucleus,theother(s)asits satel-lites.

3 Anaphoric Accessibility in RDA

In orderto useanannotationbasedon RDA to testGroszandSidner’s claimsaboutanaphoricaccessibilitywe have to specify a mappingfrom an RDA structuretofocus spaces:which RDA constituentscorrespondto separatefocus spaces,andwhich focusspacesshouldbeopenwhena givenanaphoricexpressionis encoun-tered. This mappingis not entirely trivial, becausethe structureof a discourseaccordingto RDA is muchmoredetailedthanthestructurethatwould beassignedto thatdiscourseby GroszandSidner. In RDA, eachclauseis treatedasa distinctconstituent,whereasin a G& S-style analysis,multiple sentencescanbe consid-ereda singleconstituent. Furthermore,G& S make no distinctionbetweencoresandcontributors,andonly allow two relationsbetweenintentions,whereasin RDA

many typesof intentionalrelationsarepossible.Moser andMoore partially specify a mappingbetweenRDA notionsand an

intentionalstructurein theG& S sensebasedon thefollowing principles:

{ Every DSP mustbeassociatedwith acore.{ Constituentsof theRDA structurethatdonot includecores- i.e.,clusters(seeabove) - do not introduceDSPs.

In termsof segments/ focusspaces,theseprinciplesmean,first of all, thata seg-ment in the G& S senseshouldalwaysbe a segmentin the RDA sense(an RDA-SEGMENT): an elementwith a coreandoneor morecontributors. I.e., no focusspaceshouldbepushedonthestackunlessacoreis recognized.(MoserandMooreleave openthe questionof whetherthe reverseshouldalsobe true, i.e., whetherthereshouldbea 1:1 mappingfrom G& S-segmentsto RDA-segments.)Secondly,informationalrelationsdo not affect the attentionalstate,unlessassociatedwithintentionalrelations.

Thesegmentstructurethat–accordingto whatwe have seensofar–wecande-rive from theRDA analysisin Figure14.1shouldthenbeasin Figure14.2.Noticethat becauseinformationalrelationsdon’t give rise to intentionalsegments,theinformationalclusters1.1-1.2and3.1-3.2arenot assignedseparatesegments.Inparticular, theunembeddedcorein 1.1-1.2is not assigneda separatefocusspace,sinceit expresses(partof) theDSP associatedwith theoverall RDA-segment.

understandingof thematerialpresentedin thecore,or to increasethehearer’s ability to performtheactionpresentedin thecore.” (p. 6).

133

2.1

2.2

1.1

1.2

3.1

3.2

Figure14.2:A G& S-styleanalysisfor thetext in Fig. 14.1

Even thesefirst simplemappingprinciplesalreadyresult in differentpredic-tions concerningaccessibilitythanonewould get from a pure RST analysis. InFox’s analysis,for example(seebelow), even if theconstituentspanning2.1 and2.2 werea cluster, not an intentionalsegment,no antecedentintroducedin thesepropositionswould be accessiblefrom 3.1 and3.2. This is thefirst claim whosecorrectnesswehave to test.

However, theprinciplesproposedby MoserandMoore leave a numberof as-pectsof the mappingopen. OnequestionMoserandMoore themselves raiseiswhetherembeddedcores,i.e., coresthat are themselves RDA-segments(whosepossibility they consider, but do not analyzein detail) shouldbe treatedasem-beddedG& S-segmentsor aspart of the sameG& S-segmentas their embeddingRDA-segment. (Examplesin which the antecedentof a pronounis containedinanembeddednucleusarediscussedby (Fox), p. 101.) A secondquestionis howshouldembeddedandnot embeddedcontributorsbe treated:whetherall of themshouldcountasseparateG& S-segments,or only theembeddedones.

Third, evenif (some)contributorspushfocusspaces,whenshouldthesefocusspacesbe popped? Immediately, or only when the intentionalrelation is com-pleted?E.g., in theexamplein Figures14.1and14.2,shouldsegment2.1-2.2bepoppedassoonaswe aredoneprocessingit, or shouldit remainon thestackun-til the whole segmentis over, given that it participatesin the intentionalrelationthat determinesthe superordinatesegment? Fox’s dataseemto suggestthat an-tecedentsintroducedby ‘active’ non-embeddedsatellitesshouldbeaccessible;butevenantecedentsintroducedby active embeddedcontributorsmight.

4 The Study

We testedall of thesepossiblewaysof usingan RDA structureto guidethe focusspaceconstructionmechanismusingthe Sherlockcorpus. Our resultsshow thatthe versionin which both embeddedcoreandembeddedcontributors remainonstackhastheleastnumberof referenceviolations.

134

4.1 Existing Data

The corpuswe usedis a collectionof tutorial dialoguesbetweena studentandatutor, collectedwithin the Sherlockproject (Lesgold,Lajoie, Bunzo,andEggan1992). Thecorpusincludesseventeendialoguesbetweenindividual studentsandoneof 3 experthumantutors,for a totalof 313turns(about18 turnsperdialogue),and1333clauses.The studentsolves an electronictroubleshootingproblemin-teractingwith the Sherlocksystem;then,Sherlockreplaysthe student’s solutionstepby step,schematicallycriticisingeachstep.As Sherlockreplayseachstep,thestudentscanaskthehumantutorsfor explanations.Thestudentandtutor commu-nicatein written form.

The Sherlockcorpushadbeenpreviously annotatedusing RDA to studycuephrasesgeneration(MoserandMoore 1996a;Di Eugenio,Moore, andPaolucci1997). Theresearchgroupwhich proposedRDA discussesthefollowing reliabil-ity results(MoserandMoore1996a).25%of thecorpuswasdoublycoded,andtheκ coefficient of agreementwascomputedon segmentationin a stepwisefashion.2

First,agreementat thehighestlevel of segmentationwascomputed.After comput-ing agreementat level 1, thecodersresolvedtheirdisagreements,thusdetermininganagreeduponanalysisat level 1. Thecodersthenindependentlyproceedto de-terminethesubsegmentsat level 2, andsoon. Thedeepestlevel of segmentationwaslevel 5; theκ valueswere.90, .86, .83,1, and1 respectively (from level 1 to5).

4.2 Our Methods

We annotatedabouthalf of the Sherlockcorpusfor anaphoricinformation, us-ing a muchsimplifiedversionof theannotationschemedevelopedby the GNOME

project(Poesio2000b).Morespecifically, wemarkedeachNP in thecorpus,spec-ified its NP type (propername,pronoun,the-np,indefinite NP, etc) and thenwemarkedall ‘direct’ anaphorsbetweentheseNPs (i.e.,nobridges).Thisschemehasgoodresultsfor agreement(Poesio2000a)andhasalreadybeenusedfor studyinganaphoricaccessibility(Poesio,Cheng,Henschel,Hitzeman,Kibble, andSteven-son2000).Weannotateda totalof 1549NPs,507of whichwereanaphoric.

Oneproblemwehadto addresswasthatin theRDA annotation,only tutorturnshadbeenannotated(becausethestudents’questionsarevery short),but many ofthe antecedentsof anaphoricreferenceswerediscourseentitiesintroducedin theprecedingstudentturn askingthequestion. We followed Fox andmadethefirstelementsof adjacency pairs part of the accessibilityspaceof anaphoricexpres-sionsin the secondpart. To do this, we enclosedeachstudentturn in a specialstudent-turn element,markedtheNPs it contained,andmadethis turn aspecialfocusspaceaccessiblefromtheentitiesin thetutorturn. Wedealtwith antecedentscontainedin ‘tied’ adjacency pairs(Fox 1987),i.e., in turnsfurtheraway, by count-

2It is unknown to uswhetherκ wasalsocomputedon clusters,andon thespecificinformationalrelationsused.

135

ing theanaphoricexpressionswhoseantecedentwasunaccessiblefor thesereasonsandfactoringthemout.

A secondproblemto solvewasthefactthata largeproportionof theanaphoricexpressionswhoseantecedentis noton thestackarepropernames.Becauseit canbe arguedthat theseexpressionsdo not accessthe stackto find their antecedent,we alsocountedthemseparately.

We rana scriptover theannotatedcorpusthatsimulatesfocusspaceconstruc-tion undertheseveralpossiblewaysof mappingRDA structuresinto focusspacesthatweconsidered,andattemptsto find theantecedentfor ananaphorin thefocusspacestackaccessibleto theanaphoraccordingto eachof thesepossibilities.Thevariantsconsideredare:

1. All: Pusha new focusspaceon the stackwhenever a non-atomicRDA unit (bothintentionalsegmentsandinformationalclusters)is encountered,andpopthis focusspacewhentheconstituentends.(E.g., in Figure14.1,pusha new focusspaceforall threeconstituentsof thetop segment:1.1-1.2,2.1-2.2,and3.1-3.2.)

2. Intentional Only / Imm Pop: Only pusha new focusspacewhenan intentionalsegmentis encountered;popit assoonasthesegmentis completed.(E.g.,in Figure14.1,only pusha new focusspacefor segment2.1-2.2,andpop it assoonasthatsegmentis completed.1.1,1.2,3.1,and3.2arejustaddedto thetop focusspace.)

3. Intentional Only / Delay pop of cores: Only pusha new focus spacewhen anintentionalsegment is encountered.Pop focus spacesintroducedfor contributorsegmentsimmediately;but only popthefocusspaceintroducedfor a coresegmentwhenpoppingthewholesegment.

4. Intentional Only / Partial delay pop of tribs Like the previous version,but inadditionkeepa focusspaceintroducedfor a contributor on thestackaslong asthesegmentin which it occursis still on. (E.g., in Figure14.1,do not pop the focusspacefor segment2.1-2.2beforeprocessing3.1-3.2.)

5 Results

Table14.1 illustratesthe impacton anaphoricaccessibilityof the distinctionbe-tweenintentionalandinformationalrelationsbyshowing thepercentageof anaphoricantecedentswhich areon thestackaccordingto thefirst two variantsof themap-ping algorithm. The line indicatedas ‘All’ shows the resultsobtainedby treat-ing both informationalandintentionalrelationsasintroducingnew focusspaces:’OK’ indicatesthenumberof anaphoricantecedentswhichareaccessible,’NO’ in-dicatesthenumberof antecedentswhich arenot accessible,‘Out of AP’ thecasesin whichtheantecedentis notaccessiblebecauseit’soutsidethecurrentAdjacencyPair, and’PN’ the numberof casesin which the antecedentis not accessiblebuttheanaphoricexpressionis apropername(which,presumably, canaccessit deno-tation throughlong termmemoryratherthanthroughthestack).Thetableshowsthat separatingintentionalconstituents(which introducenew focusspaces)frominformationalclusters(thatdon’t) makesmoreantecedentsaccessible;theresultishighly significantby theχ2 Test(χ2 | 29} 47~ p � 0 } 001).

136

OK NO Out of AP PNAll : 199 74 63 158Intentional only / Imm Pop: 280 20 63 131

Table14.1:Theinformational/ intentionaldistinctionandaccessibility

OK NOInt. only / Imm pop of emb core and trib 280 20Int. only / Delaypop of emb cores 287 16Int. only / Delaypop of emb trib 310 8

Table14.2:Effectof thedifferencesin popping.

Table14.2showsthedifferencesamongthedifferentwaysof fixing theoptionsleft openby MoserandMoore(variants2-4above). Thefirst line shows theresultsobtainedif both embeddedcoresand embeddedcontributors were to pushnewfocusspaces,poppedassoonastheseembeddedconstituentsarecompleted;thesecondline theresultsif we keepembeddedcoresopenuntil theendof the RDA-segment; the third line the resultsif we treat embeddedcontributors within anRDA-segmentas remainingon the stackuntil the segmentis closedoff. In thistablewe have ignoredboth casesin which the antecedentis inaccessiblebut theanaphoricexpressionis apropername,andthe63casesin which theantecedentisinaccessiblebecauseit’snot in thesameadjacency pair(seediscussionabove). Thedifferencesarenot sogreatin thiscase,but thecorrelationis still significant(χ2 |6 } 09~ p � 0 } 05) andin particularthereis a clearly significantdifferencebetweenthe simplestpossibletreatmentof embeddedintentionalrelations,in which theyarealwayspopped,andthelastmodel.

6 Discussion

As saidabove, we studiedtwo separateissues:how bestto useideasfrom RDA

to make G& S’s theorymorespecific,andto evaluateG& S’s proposalsconcerninganaphoricaccessibilityunderthismapping.

6.1 Mapping RDA into FocusSpaces

Thefirst goalof our work is to useRDA structuresto gaina moredetailedunder-standingof whenfocusspacesshouldbeopenedandclosed. In this respect,ourfirst resultis a significantlybettercharacterizationof anaphoricaccessibilityif weassumethatnew focusspacesareonly pushedon thestackwhencoresarerecog-nized,asopposedto alsobeingpushedwhena purely informationalstructureisobserved. This resultis especiallyinterestingwhencomparedwith Fox’s proposal(discussedbelow). Accordingto Fox, informationalrelationsalsoaffectaccessibil-ity. It is alsointerestingto contrastthis resultwith theproposalsof VeinsTheory,

137

whichalsomakesno distinctionbetweeninformationalandintentionalrelations.Thesecondinterestingfinding in this respectis thatthebestresultsconcerning

accessibilityareobtainedwhenembeddedcontributors,aswell, areonly poppedwhenan RDA-segmentis closed. This is not somethingthat would be predictedon the basisof eitherFox’s work, or any obvious interpretationof the notion ofRDA-segment;in fact,aswe will seeshortly, it’s not immediatelyobvioushow toaccountfor this resultin termsof GroszandSidner’s theory, either.

6.2 An Evaluation of Groszand Sidner’s Theory

If we assumethat contributors stayon the stackuntil the RDA-segmentis com-pleted,only 8 anaphoricantecedentsareoutsidethecurrentlyopenfocusspaces.Of these,5 arecasesof definitedescriptionsthatmight beviewedasreferringde-ictically to partsof thecircuit, 1 is acataphoricdiscoursedeixis,oneis a temporaldeixis (at this point), andonea bit unclear. In otherwords,underthe suggestedrulesfor openingandclosingfocusspaces,virtually all anaphoricantecedentsareaccessible.

The questionthenis whethertheserulesareconsistentwith the G& S frame-work: the answer, we think, is that while the simplesttreatmentsof contributorsandcoresprobablyare,themostcomplex onesprobablyaren’t. In fact,we didn’teven testwhat somepeoplemight think of as the mostnaturalmapping- maketheunembeddedcoreof anRDA segmentpartof thefocusspacefor thatsegment,but treatunembeddedcontributorsasintroducingsubordinatefocusspaces- sincethis would violate one of MoserandMoore’s basichypotheses(that only RDA-segmentsresult in G& S-segments). The approachof leaving unembeddedcoresandcontributorson thestack,but poppingall embeddedRDA-segments,canprob-ably be still viewed asconsistentwith the theory;3 but delayingthe poppingofembeddedcoresandcontributorson thestackmaybeconsideredbeyond its lim-its, whetherwe view the problemin algorithmictermsor in termsof intentionalstructure.Leaving thesefocusspacesopenmeansthat theattentionalstatecannotbeproperlyseenasa stackanymore,but hasto beseenasa list, sincethereis noguaranteethat thefocusspaceassociatedwith thecorewill bethelastelementonthestack. Figure14.3shows anexamplein which it is necessaryto leave theem-beddedcontributor 24.13-24.14on thestack,in orderto solve the“other” voltageto someothervoltage; attheendof thesegment,this focusspacehasto beremovedwhile leaving the focusspacefor thecoreon top of thestack. Furthermore,notethat in thecaseof Figure14.1we needa ‘discontinuousfocusspace’to make theembeddedcontributoraccessiblewithouteliminatingtheprinciplethatnew entitiesarealwaysaddedto thefocusspaceon topof thestack.

Looking at theproblemfrom thepoint of view of intentionalstructure,thesecasesof accessibilitycanonly behandledwithin theG& S framework by hypothe-

3This however meansthat the intentional relationscontainedin the segment- e.g., convincein Figure14.1–shouldbecomepart of the DSP for that segment,which againmay or may not beconsistentwith G& S’s ideas.

138

24.13a SinceS52putsa return(0 VDC) on it’soutputs24.13b whenthey areactive,24.14 theinactive statemustbesomeothervoltage.24.15 So even thoughyou maynot know what the ”other” voltageis,24.16 youcantestto ensurethat24.17a theactive pinsare0 VDC24.17b andall theinactive pinsarenot0 VDC.

Enable

effect:cause

24.13a 24.13b

Enable

24.14

cause:mental-effect

Concede

criterion:act

24.15 24.16

24.17a 24.17b

contrast1:contrast2

criterion:act

Figure14.3:A contributor thatprecedesthecore

sizingthattheDSPsassociatedwith embeddedcoresandcontributorssatisfaction-precedethecoreandtheothercontributors. More in general,onehasto wonderwhetherFox’s (and beforethat, Reichman’s) suggestionthat the material intro-ducedby ‘active’ propositions(i.e., theotherconstituentsof a rhetoricalscheme)remainsaccessiblecanreally beformulatedin termsof GroszandSidner’s inten-tional structure.

6.3 Open Issues

Keepingmoreentitieson the stackis obviously going to increasethe likelihoodthatanantecedentwill befoundon thestack,but we have to becarefulnot to losethecrucialpropertyof theattentionalstate,reducingsearchambiguity(otherwise,wecouldjustmakeeverydiscourseentityaccessibleatall times).In orderto prop-erly comparemodelsof anaphoricaccessibilityit is necessaryto computewhatwemight call their ‘perplexity’; unfortunately, in orderto properlydo this it is nec-essaryto have ananaphoricresolutionmechanism. We planto do this evaluationnext.

A major problemwith any attemptat looking at the empirical import of dis-

139

coursestructureis thatit’s very hardto getresearchersto agreeon whatthestruc-tureof agivendiscourseis. Work suchasMarcu’shasshown thatin generalwecanexpectpeopleto roughlyagreeon themainsegmentsof adiscourse,but it is muchharderto getagreementontherelative importanceof theconstituentsegments,andevenharderto getagreementon theparticularrelations.At themoment,theonlyway to achieve a reliableannotationis by extensive training anddiscussions,asdonewhenconstructingtheSherlockcorpus.So,we do not expecteverybodytoagreeon theparticularanalysesproposedhere;we would want to point out how-everthattheseannotationsdohaveacertainamountof objectivity in thatthey werenotproducedby us.

7 RelatedWork

7.1 Fox

(Fox),althoughonly concernedwith referencesto singularandhumanantecedents,is perhapsthemostextensivestudyof theeffectsof discoursestructureonanaphorain both spoken andwritten discourses.Fox usesdifferentmethodsfor analyzingthe two genres:sheusesconceptsfrom ConversationAnalysis,andin particularthenotionof Adjacency Pair, for spokenconversations,andRST to analyzewrittentexts. Hermainproposalaboutwritten texts is asfollows:

A pronounis usedto referto a personif thereis a previousmentionofthat personin a propositionthat is ACTIVE or CONTROLLING; other-wisea full NP is used.

(Wherea propositionis ACTIVE if it’s partof thesameRST schemeasthepropo-sition in which thepronounoccurs;whereasapropositionis CONTROLLING if it’spartof a schemewhichdominatestheschemein which thepronounoccurs.)

Fox’s proposalsconcerningpronominalizationapplylesswell to referencestoobjects(andevenin hercorpustherearemany referencesfor which thehypothesisabove would licencethe useof a pronounareactuallyrealizedby a definite NP,which sheexplainsby arguing that theprincipleabove is only oneof many inter-actingprinciplesthat determinethe realizationof a NP); nevertheless,shemakesa lot of compellingpointsaboutstructure. In particular, shemakes it very clearthatactivepropositionsshouldbeaccessiblefor aslongastheschemeis open;andproducesseveral examplesshowing that materialintroducedin active embeddednuclei is accessible.Fox didn’t find referencesinsideactive embeddedsatellites(but thenagainnoneof theseis madevia a pronounin our corpus). In addition,our studysuggeststhatpropernamesbehave differentlyfrom definitedescriptionsin that the formeraremuchlesssensitive to discoursestructurethanthe latter, sothe two classesshouldnot be conflatedlike Fox does;andnot separatinginfor-mationalrelationsfrom intentionalonesrestrictstoo muchtherangeof accessibleantecedents,evenif it maybecorrectasfar aspronominalizationis concerned.

140

XNucleus

Nucleus YInfo rel

Intentional rel

Figure14.4: A treedistinguishingVeinsTheoryfrom the proposalsdiscussedinthispaper

7.2 VeinsTheory

VeinsTheory(VT) (Cristea,Ide,andRomary1998;Cristea,Ide,Marcu,andTablan2000;IdeandCristea2000)is a recentlyproposedtheoryof theeffectof discoursestructureon anaphoricaccessibility, which relieson RST for its definition of dis-coursestructure,andwhosepredictionshave beentestedusingan RST annotatedcorpusof newspapertexts (Cristea,Ide, Marcu, andTablan2000). The proposi-tions accessibleto an anaphoricexpressionsarecomputedby an algorithm thatoperatesdirectly over an RST tree and involves two steps: a bottom up stepinwhich the‘heads’of eachnodein thetreearecomputed(wheretheheadof anon-terminalnodeis theconcatenationof theheadsof its nucleardaughters)followedby a top-down computationof the VEIN EXPRESSIONS. The crucial ideaof VT

is thatmaterialintroducedin nuclearnodes‘percolatesup’ veins,whereveinsarepathsin the treeall of whosearcsconnectnuclearnodes;the antecedentsintro-ducedin any nodealongthevein areaccessiblefrom all thenodesof thesubtreewhich hasthetop of thevein to which thatnodebelongsasits root. Thesecondideaof the theoryis thatantecedentsintroducedin a satellitenodeto the left of anucleusremainaccessibleto all nodescontrolled(in Fox’ssense)by thatnucleus.

In somerespects,theproposalpresentedherecanbeviewedasageneralizationof theproposalsof VT: thematerialintroducedin coreconstituentspercolatesupinasimilarway, but we alsoallow antecedentsintroducedby embeddedcontributorsto the right of thenucleusto beaccessibleaslong asthesecontributorsareactive(in VT, only binarytreesareconsidered).

Theonepoint of contrastbetweenthetwo theoriesis that in our proposal,wedo not considerall nuclei, but only the nuclei of intentionalrelations. We haveseenthat treatinginformationalrelationsasintroducingfocusspacesmakesa bigdifferencein termsof accessibility;the differenceis significanteven if we allow

141

theseadditionalrelationsto remainopenaslongasthedominatingrelationis open.Thecrucialcasedistinguishingthetwo theoriesaretreeswith thestructurein Fig14.4: in our theoryX would beavailableasanantecedentof Y, whereasin VeinsTheoryit wouldn’t.

Bibliography

Carletta,J.,A. Isard,S. Isard,J. Kowtko, G. Doherty-Sneddon,andA. H. An-derson(1997).Thereliability of a dialoguestructurecodingscheme.Com-putationalLinguistics23(1), 13–32.

Cristea,D., N. Ide, D. Marcu, andV. Tablan(2000).Discoursestructureandco-reference.In Proc.of COLING, pp.208–214.

Cristea,D., N. Ide, andL. Romary(1998).Veins theory: A modelof globaldiscoursecohesionandcoherence.In Proc.of COLING, pp.281–285.

Di Eugenio,B., J. D. Moore, andM. Paolucci (1997).Learningfeaturesthatpredictcueusage.In Proc.of the35thACL, Madrid.

Fox, B. A. (1987).DiscourseStructure andAnaphora. Cambridge.

Grosz,B. J. andC. L. Sidner(1986).Attention, intention,andthestructureofdiscourse.ComputationalLinguistics12(3), 175–204.

Ide, N. and D. Cristea(2000, October).A hierarchicalaccountof referentialaccessibility. In Proc.of ACL, HongKong.

Lesgold, A., S. Lajoie, M. Bunzo, and G. Eggan(1992). SHERLOCK: Acoachedpracticeenvironment for an electronicstroubleshootingjob. InJ. Larkin andR. Chabay(Eds.),Computerassistedinstruction and intel-ligenttutoring systems, pp.201–238.Hillsdale,NJ: Erlbaum.

Mann,W. C. andS.A. Thompson(1988).Rhetoricalstructuretheory:Towardsa functionaltheoryof text organization.Text 8(3), 243–281.

Marcu, D. (1999, May). Instructionsfor manually annotatingthe discoursestructuresof texts.Unpublishedmanuscript,USC/ISI.

Moore,J.andM. Pollack(1992).A problemfor RST:Theneedfor multi-leveldiscourseanalysis.ComputationalLinguistics18(4), 537–544.

Moser, M. andJ. D. Moore(1996a).On thecorrelationof cueswith discoursestructure.Unpublishedmanuscript.

Moser, M. and J. D. Moore (1996b).Toward a synthesisof two accountsofdiscoursestructure.ComputationalLinguistics22(3), 409–419.

Moser, M., J.D. Moore,andE. Glendening(1996).Instructionsfor CodingEx-planations:Identifying Segments,RelationsandMinimal Units. TechnicalReport96-17,Universityof Pittsburgh,ComputerScience.

142

Nakatani,C. H. (1996).Discoursestructuralconstraintson accentin narrative.In J. P. H. van Santen,R. W. Sproat,J. P. Olive, andJ. Hirschberg (Eds.),Progressin Speech Synthesis. SpringerVerlag.

Nakatani,C. H., B. J.Grosz,D. D. Ahn, andJ.Hirschberg (1995).Instructionsfor annotatingdiscourses.TechnicalReportTR-25-95,Harvard UniversityCenterfor Researchin ComputingTechnology.

Poesio,M. (2000a,May).Annotatingacorpusto developandevaluatediscourseentity realizationalgorithms:issuesandpreliminaryresults.In Proc.of the2ndLREC, Athens,pp.211–218.

Poesio, M. (2000b, July). The GNOME Annotation Scheme Manual.University of Edinburgh, HCRC and Informatics. Available fromhttp://www.hcrc.ed.ac.uk/ ˜ gnome.

Poesio,M., H. Cheng,R. Henschel,J.M. Hitzeman,R. Kibble, andR. Steven-son (2000, October).Specifying the parametersof CenteringTheory. InProc.of the38thACL, HongKong.

Reichman,R. (1985).GettingComputers to Talk Like YouandMe. Cambridge,MA: TheMIT Press.

143

Topic Structur e in RouteExplanation Dialogues.

LAURENT PREVOT

IRIT- Universite Paul Sabatier

118RoutedeNarbonne

31062ToulouseCedex 04

France

[email protected]

ABSTRACT. This paperdealswith the representationof dialogue,including the semanticas-pects.We have beenworking on a realspeechcorpusof routeexplanationdialogues.Thepaperfollows thetraditionof dynamicsemantics(KampandReyle 1993),andwewill usetheSDRT of(Asher1993)which proposesa treatmentof thesemantic-pragmaticinterface.Dynamicseman-tic theorieshave traditionally focusedon discourseissues,but a few studieshave begun to lookat dialoguewithin that framework in the pastfew years. (AsherandLascarides1998). To therhetoricalandintentionalstructureproposedby SDRT we adda ”topic structure”closelyrelatedto theinformationalstructure.

1 Intr oduction

Weaddressheretheproblemof representinghumandialogue1. Weareinterestedinaspecifickind of dialogue: routeexplanationdialogue2. Weplaceourselvesin thetraditionof dynamicsemantics(GronendjikandStokhof1991)(KampandReyle1993)(Asher1993). Theseapproaches,which have mainly focusedon theprob-lem of monologue,arenow addressingthe domainof dialogue(Ginzburg 1994)(AsherandLascarides1998). From anotherperspective, works alsoexist whichenrich dialogue-acttheorieswith semanticcontent,rhetoricalstructureand dia-loguegames(PoesioandTraum1997).

Here, the chosenframework is SDRT (SegmentedDiscourseRepresentationTheory)(Asher1993). Severalworksbasedon this theoryhave studiedthe inter-actionbetweensemanticcontentandDiscourseStructure (Asher, Aurnague,Bras,Sablayrolles,andVieu1995)(Asher, Busquets,andVieupear).

Thenotionof “topic” is usedin this framework but it is not preciselydefined.Recentlywe have proposeda treatmentof dialogueglobal topic in routeexplana-tion dialogues(Prevot 2000).After thecollectionof arealcorpus,ouranalysishas

1Dialogueis seenhereasa discoursewith two speakers.2Seesection2 for a morecompletedefinitionof this kind of dialogue.

145


shown thenecessityof takingInformationStructure into account,aswaspredictedby (AsherandTxurruka1995)(Engdahl,Larsson,andEricsson2000)(Engdahl2000).

Now, in a prospective work, we want to studythe interactionbetweenInfor-mationStructure andDiscourseStructure andbetweenInformationStructure andTopicGlobal Structure.

First, we briefly presentthe corpusthat we have recordedand that we arestudying(section2). Then our analysisof somepiecesof thosedialogueswillbepresentedin threepoints: first from theSDRT point of view, strictly speaking(section3.1) ; thenin thelight of Txurruka’s accountof InformationStructure andDiscourseStructure (section3.2); thirdly, takingglobaltopic into account(section3.3). Lastly, weproposeaconstructionof a topicstructure(section4).

2 The Corpus

We introducein this sectionthecorpuswe have recorded.At this stage,it is com-posedof 15 dialoguesfor a total time of 30 minutes.To ensurea large coverageof the kind of phenomenawe plan to collect moredialogues.As we have said,it is a corpusof routeexplanationdialogues.This kind of dialoguesinvolvestwoparticipants: a giver anda receiver. Thegiver explainsto thereceiver how to gofrom onelocationto anotherone.Wehavechosenthedomainof routedescriptionbecausea lot of work hasalreadybeendoneon it, especiallyin lexical semantics(motion verbs,spatio-temporalprepositions,}J}J} ) (AsherandSablayrolles1995)(Muller 1998). It give usa basisfor therepresentationof thesemanticcontentofutterances.

Many featuresareimportantin thecompletedefinitionof thisdialogues(char-acteristicsof placesused,of the participants,}J}J} )3. Here, for reasonswe willexplain below, we focuson two of thesefeatures: dialoguesituationandpartici-pants’knowledge.

As far asthe dialoguesituationis concerned,we will focuson the followingtwo questions:{ Are thespeakerslocatedin the“explanationplace”?{ Are thespeakerstogether, in asameplace?

In our work, we want to focuson purelyverballinguistic phenomena.Therefore,wehave to minimizetheutterancecontext dependency (deicticuses,gestures,faceexpressions).Thuswe have notplacedtheparticipantstogetherin asamelocation

3See(Denis1997)for a classificationof routeexplanation.

146

andhave recordedtelephoneconversations.

During a previous corpuscollectionwe hadnoticedthat if the receiver didn’tknow anythingaboutthegloballocationincludingtheroute(for examplein atown)therewasno exchangebetweenparticipants.As we wantedpreciselyto studyin-teractionsbetweentheparticipantsmentalstates,wechosesubjectswhohadsomeknowledgeaboutthelocation.

Theanalysisof our corpusis directedalongthreelines:

1. StudytheDiscourseStructure/ InformationStructureinteraction.

2. Within eachstructure,studytheinteractionof sententialanddiscursive level.

3. StudytheTaskStructure/ LinguisticStructureinteraction.

Below is a transcriptof a representative examplefrom ourcorpus.

A 1 A: Allons-y! donctu parsd’ou?Let’sgo ! sowhereareyougoingto comefrom?

2 B: Jeparsdoncde chezLaurent,c’est dansla granderue St-Micheleuhapeupres200metresavantla placeSt-Michel.I will comefrom Laurent’s place, it is in St-Michel Streeter... about200mbeforeSt-MichelPlaza

3 A: 200metresavant la placeSt-Michel. Donccte plutt versle boule-varddesRecollets,enpartant....200m before St-Michel Plaza. Therefore, rather on the side of RecolletsBoulevard, leavingfrom....

4 B: Ouais.yeah.

5 A: D’accord. Donc pour venir chezPhilippec’est assezsimple. Tudescendsla granderueSt-Michel.Tu traversesle carrefour.Ok. Soto cometo Philippe’s,it’ srathersimple. YougodownSt-Michelstreet.You crossthejunction.

6 B: Ouais.yeah.

7 A: Tu... euh...Tu prendsla ruePharaon.you... er... youtakePharaonStreet

A 8 B: La ruePharaon?PharaonStreet?

9 A: Oui.Yes.

10 B: Jeneconnaispascetterue.I don’t knowthat street

147

11 A: Tu neconnaispasla ruePharaon.You don’t knowthat street.

12 B: C’estapresle carrefour, euh?It’ safter thejunction,..?

13 A: Dansle prolongementeuhoui c’estapresle carrefour.continuing... er yesit’ safter thejunction.

14 B: Ouais.yeah.

15 A: Dansle prolongementde la granderueSt-MichelTu asla ruequis’appelleje saisplusrueduLanguedocou la rueOzenne.ContinuingSt-Michelstreet,you’vegotthatstreetcalledi don’t know, Langue-docStreetor OzenneStreet

16 B: Ouaisd’accord.yeah,okay

17 A: Bon euh100 metresapres le carrefour, le grandcarrefourde St-Michel...euh...il y aunefourche,la ruedulanguedocoula rueOzenneje saispluscommentelles’appelleobliqueversla droite.well er 100mafterthejunction,after thebig junctionat St-Micheler... there’sa fork, LanguedocStor OzenneSt,i don’t rememberits name, it turnsright.

18 B: Ouais.yeah.

19 A: et tu asdevanttoi unepetiterueasensunique...qui s’appellela ruePharaon.andyou’vegota smallone-waystreetin front of you...calledPharaonStreet

20 B: Ah ouaisd’accord.OK. Jesitue.Oh,yeah,okay. I see.

21 A: Donc, tu parcourstoutecetterue... euh...Tu vasarriver a la placedesCarmes.So,you take this streetand go all the way, er... you will arrive at CarmesPlace

22 B: muhm...muhm......

A 23 A: Arrive a la placedesCarmeseuh...tu tournesa gauchedansla ruedesPolinaires,c’est unerue qui debouchesur la placedesCarmes...Arrive au27 de la ruedesPolinaires...sur ta droite tu trouveras...euhle 27ou habitePhilippeMuller.Whenyouare a theCarmesPlace,... youturn left in PolinairesStreet,it’ s astreetthatgoeson CarmesPlace. whenyouareat the27of Polinairesstreet,on your right youwill seeer... the27 where PM lives.

24 B: OK. Donc,ruedesPharaonsjusqu’a la ruedesCarmeset apres...Ok. So,PharaonStreetup to CarmesStreetandthen...

148

25 A: jusqu’a la placedesCarmes.up to CarmesPlace

26 B: Oui placedesCarmes...et a gauche.yes,CarmesPlace,... andto theleft

27 A: Alors la premierea gauche,c’estpasla bonnequandtu debouchessurla placedesCarmes.thenthefirston theleft is not theright onewhenyouarrive at CarmesPlace

28 B: Ouais.Yeah.

29 A: la secondea gauchec’est la bonne. C’est au bout de la placedescarmesenfait. Est-cequetu vois la ruedesFilatiers?Thesecondoneon the left is the right one. It’ s at the endof CarmesPlaceactually. Do youseeFilatiers Street?

30 B: Ouais.yeah.

31 A: Eh bientu vasjusqu’a la ruedesFilatiers,tu t’engagespasdedans.Tu regardesa tagaucheet tu verrasla ruedesPolinaires.well, you go to Filatiers Street,you don’t take it. You look on your left andyouwill seePolinairesStreet

32 B: D’accord.Okay.

33 A: C’estbon?All right ?

34 B: Ouaisc’estbon.yeah,all right.

3 Elementsof dialogueanalysis

3.1 DialoguestructuresasDiscourseStructures

We will now give an SDRT-orientedanalysisof a pieceof this dialogue,but firstwe will give a very roughoutline of this theory. Asher’s theoryassumesthat adialogueis a kind of discourseinvolving two participants.SDRT alsoassumes,asRST(MannandThompson1987),thatadiscoursecanbeseenasasetof segmentslinked with (rhetoricalor intentional)relations.Theserelationscanbehierarchi-cal or not. Hierarchicalrelationsbetweensegmentsinducea treestructurewhichimposesconstraintson theinterpretationof currentutterances(e.g.anaphorareso-lution canbelimited to certainsegmentsof thecurrentinterpretation).More gen-erally, constraintsbasedon the structureinducedby the segmentationdeterminewhat is a coherentdialogue.New utteranceswill beattachedto somesegmentin-crementallywithin analreadyexisting dialoguestructureasthey come,by taking

149

into accountlexical semantics,world knowledgeandsemantic-pragmaticrulesse-lectinganappropriaterelationin context (see(AsherandLascarides1998)).Theseaspectsareoftentakenfor grantedin theoriestakingasprimitivespropositionsandspeechactsrelatedto thesepropositions.Our goal is to accountfor theconstruc-tion of suchstructureswithin anincrementalinterpretationprocessof thedialogueseenasa public constructionof a commongroundbetweenparticipants.We willmake useof topicsin thatperspective, focusinghereon questionsandanswersse-quences.Ourapproachis thusrather“top-down” comparedto (Kruijf f-KorbayovaandKruijf f 1997)who placesheranalysisin the samediscursive perspective butcloserto the sentencelevel. That beingsaid,even thoughwe useSDRT asourframework, we departsomewhat from someof its choicesmadein thecontext ofdialoguese.g. in (AsherandLascarides1998). We will keepthe following prin-ciples,while leaving asidethepresentlylessstableaspectsof thestill in progresstheory:{ theglobalrepresentationof adialogueis composedof asetof labelledspeech

acts(SA) andrhetoricalrelationsbetweentheseoccurrencesof SA.A speechactwill beof theform � Speaker~ Mode~ Content � , wherethemodecanbein-terrogative, declarative or imperative. Contentcanbeany partof the repre-sentationof thedialogueor asub-dialogue.A basicsemanticcontentwill bea DRS, that is a setof linguistic referentsandof conditions(predicates)onthosereferents.Speechacts,seenin other frameworks ashaving an inten-tional content,take on an intentionaldimensiononly whenlinked togetherby relationsbearinganintentionalcontent.{ In orderto dealwith dialogue,SDRT hadto beenrichedwith new relations,Thelist of relationsneededby adialogueversionof SDRT is still veryunsta-ble in theprospective work of (AsherandLascarides1998).Wehavedividedtheminto thefollowing categories:

— discursive relations4: narration,elaboration, explanation,continuation,background,parallel, contrast,comment,result,consequence;

— dialogic relations: we departfrom the work of Asher by consideringthefollowing relations,dividing questionelaborationsinto precisionrequest5

andconfirmationrequest6. (At this time we don’t make any differencesbe-tweendirect andindirect partial answer.7. no-answer(linking replies,such

4Thesetpresentedin (Asher, Busquets,andVieupear).5DividedbetweenYes/NoquestionsandWH-questions.6This kind of requestdoesn’t introduceany information, it only aska confirmationof already

introducedinformations.SeeB8.7Probablythesedistinctionsarenecessaryfor theevaluationof the“answerness”of anutterance.

But for themoment,within theapplicationof our rules,we only considerthat thereareutteranceswhich give an answer(direct, indirect or partial) and somewhich don’t. In technicalterms,therelationanswerrecoversQAP (QuestionAnswerPair), IQAP (IndirectQuestionAnswerPair), andPQAP(PartialQuestionAnswerPair) SDRT relations.

150

as,“I don’t know”), acknowledge, andcorrection. For instanceanswer� α ~ β �relatesa turn β to its correspondingquestionα.

conf irm-r — answer

B 7 A: Tu...sss...Tu prendsla ruePharaon.You ...sss...You take Pharaonstreet.

8 B: La ruePharaon?PharaonStreet?

9 A: Oui.Yes.

π7

π8

π9

It shouldbenotedthat intonationis crucial in this example.Theinterrogativemode,thatweusefor theinferenceof theconfirmation-requestrelation,is givenbyan interrogative intonation. This kind of questionsis preciselyidentifiedby Reis(citedby (Engdahl2000))as“echodecidingquestion”.Hereif we don’t have therising intonation,(8) couldbeanacknowledgementof (7).

3.2 Inf ormation Structure,Asher and Txurruka analysis

In this sectionwe aregoingto presenttheaugmentedSDRT of (AsherandTxur-ruka1995).

As wassaidbefore,dialogueanalysishasto take into accountinformationalstructure. Thequestionis how to explain why a singlepropositionalcontentmayhave different realizationsin dialogue(Lambrecht1994) (Vallduvı and Engdahl1996).Thepartitioningof thecontentof a sentenceinto two parts–sentencetopicandsentencefocus–structuresthesentencein termsof informativeness.Thepar-tition criteria variesamongauthors: given/in discussion,presupposed/asserted,}J}J} 8 andthesedifferentdefinitionsraisedifferentdenomination: topic/comment,topic/focus,ground/focus. In spite of theseheterogeneityin the sentencetopicdefinition,we will usethecommonnotionof topic emphasis.This InformationalStructure is assumedto begivenby thesyntacticandintonationalanalysis(Lam-brecht1994)(vanKuppevelt 1995).

Now we focuson AsherandTxurruka’s propositions.They study the inter-actionbetweenDiscourseRelationsand InformationalStructure9 of utterances.They assume,as their main hypothesis,a doubleinteractionbetweenDiscourse

8For a completetaxonomyof topic/focuspartition,see(vanKuppevelt 1998).9They speakof informationalpartition with the aim to underlinethe two separatepartsof the

sentence.

151

Structure and InformationalStructure. The informationalstructurewill be influ-encedby thepreviouspieceof discourse,but thediscourserelationwhichlinks twosegmentshelpsto disambiguatetheinformationalstructureof thesecondone.

They claim for a systematicstudy of context impact on the determiningofinformationalstructure.

Very roughly, wecansaythatin their theory:{ Discourserelationsforceconstraintson InformationalStructure. (DS � IS){ Theconstraintsatisfaction is an importantclue for theestablishmentof thediscourserelation.(IS � DS)

Now we aregoing to considerour examplefrom AsherandTxurruka’s pointof view. We now look morepreciselyat (25) turn of (A) (a correction).Thecor-rectedutterancewill be calledthe target (here24). AsherandTxurrukasaythata correctionintroducesan inconsistentpropositionwith someprevious discourseproposition. In the example,thereis no inconsistency in a strict sense. (Asher1998)extendsthis definition andproposesa semanticfor thecorrectionrelation.Herewe have only studiedonekind of correction. In the future we will have tolook at theotherones.

¿FromAsher and Txurruka’s point of view the interactionbetweenthe dis-courserelationof correctionandthe InformationStructure of (25) is fundamentalfor dialoguecoherence.In particular, thedistinctionbetweenthe linguistic mate-rial sharedby correction(25) andtarget (24) (until theCarmes)andcontradictinginformation(place/square) is givenby Informationstructure. Theinconsistentin-formation(relatedto a sharedsituation)will be in focuswhereassharedinforma-tion will appearassentencetopic element.

In our example,it is the intonationwhich stressesthefocus. The InformationStructure is crucial here; indeedif the stressedelementwas not only the word”place”, thediscourserelationwouldbedifferent.

C 24 B: OK. Donc,ruedesPharaonsjusqu’a la ruedesCarmeset apres...ok soPharaonstreetup to Carmesstreetandafter...

25 A:...jusqu’a la PLACEdesCarmes....upto CarmesSQUARE10

26 B: Oui, placedesCarmes.Yes,Carmessquare.

The(24-25)pair satisfiesthecorrectionconstraint11: Thecorrection(25) andthe target (24) differs in onepoint. The focusof the correctiondoesnot match

10We write in capitalsthestressedpartof utterances.11See(AsherandTxurruka1995), (Asher1998) for a morecompleteaccounton correction in

SDRT.

152

with the correspondingelementin the target. Here the stressedelement“place”(square)doesnotmatchthe“rue” (street).If wedidn’t have this intonationalstresson “place”, (25)wouldbeacontinuationof theexplanationof theroute(figure1.)insteadof a correction(figure 2.). We noticethat here,we don’t have an incon-sistency12 betweenthecorrection(25) andthe target (24). So,a strictly semanticanalysiscannotprovide uswith thecorrection.HereInformationStructure is notonly acuein theanalysisbut really theonly way to obtaintheright interpretation.

correction — ack

continuation — ackπ24 π25

π26

� Fig1 }��

π24

π25

π26

� Fig2 }��

3.3 Global Topics

Dialoguesof our corpusareall task-oriented.This kind of dialoguearestudiedby a lot of researchers,mainly in a human-machineinteractionperspective. Awidespreadideaaboutthesedialoguesis that they obey a generalpurpose,whichconstrainsthe structureof the whole dialogue. (Groszand Sidner1986) speakabouttheDiscoursePurposedistinguishedfrom theDiscourseSegmentPurpose.(vanKuppevelt 1995)enrichesthisview with hisdefinitionof topic in termof topicconstitutingquestions. Heclaimsthatin aTaskOrientedDialogue,asin anarrativediscourse,“the utterancesbelongingto themainstructureform togetherananswerto a singletopic-constitutingquestion”. We follow this ideathat thereis a topicwhich archesover theothertopics(which becomesub-topics).In this sectionwestudytheeffectof globaltopic,notonly theoverarchingonebut alsothesubtopicscorrespondingto sub-dialogues.Froma generalpoint of view, thedialoguetopicis givenby thecontext of dialogue(linguistic or not; verbalisedor not).

In our corpus,theoverarchingtopic is a routeexplanation. We have to repre-senttheroutecomponentsrelatedto theexplanation: thereis astartingpoint,and

12In fact,if therewasaCarmesstreetleadingto aCarmessquare,therewouldbenocontradictionbetween(24)and(25). It is theintonationthatpointsout thecorrectionnatureof (25). SoAsherandTxurruka’s propositionfor theconstraintsoncorrectionrelationmustbeextended.

153

endpoint, anda prescriptionsequence(Denis1997), introducingrecursive sub-itinerary descriptions.The two participantsaresupposedto fill in an incompletestructuremadeof theseelements.Thispresupposedstructuregivesusacanvasfortheconstructionof thefinal, sharedstructure.Wecall theseelementsthedialoguetopics (DT). It will beusedin a top-down manner. Wegive thefollowing form :

ab c d

start(a)end(b)path(c)direction(d)

Fig3 }

We have to explain what is a “path” in this context. Thereareseveralwaysofdefiningpathsin theoriesof space13. But in theroutecontext, apathis usuallyseenasanactionsequence.Eachactionof thesequenceis a movebetweentwo places.Weareinterestedin thelinguisticaspectof routes,sowefocusontheprescriptionsof theseactions.Participants(A andB) aregoing to establisha sequenceof pre-scriptionof theactionsthatB hasto do in orderto follow thepath(c) in adirection(d) from (a) to (b).

At the beginning of the dialogue(turns1 to 5), thereis a first sub-dialogtoestablishthestartingpoint. It is a subtopicelaboratingtheoverarchingone. It isrelatedto the referent“a”. This sequenceis a specificationof “a” which fills outthecanvas. We don’t give herethedialogueanalysisof this establishmentof thestartingpoint,but wegive therepresentationfor thisdialoguesegmentwith anap-proximatetranslationin English(Fig4 } )1415.

13See(Muller 1998)for a presentationof thesedefinitions.14The“d1” directionis presupposedby the“avant” use.It is not thesamewith “d” (from overar-

chingglobaltopic)whichcomesfrom thepreliminarycontext.15For a morepreciseaccountof thesemanticrepresentationof a route,see(Krause2000).

154

e f d1g h

Laurent(e)chez(e,a)rue(f)nom(f,”St-michel”)dans(a,f)place(g)nom(g,”St-michel”)direction(d1)avant(a,g,d1,c)distance(a,g,200m)rue(h)nom(h,”Recollets”)du-cote-vers(a,g)

e f d1g h

Laurent(e)a is Laurent’s placestreet(f)name(f,”St-michel”)in/on(a,f)square(g)name(g,”St-michel”)direction(d1)before(a,g,d1,c)distance(a,g,200m)street(h)name(h,”Recollets”)near(a,g)

Fig4 }

All the informationsexchangedin this sub-dialogconcernsthestartingpoint.This factis crucialinformationin theanalysisof thedialogue.

The(8-20)sequenceis theestablishmentof Pharaonstreetintroducedby (7).We do not study herethe constructionof this establishment.At the end of thesequence,Pharaonstreetis groundedandthesharedinformationis only thetakingof thestreetby B (Fig5 } ) .

g e1

street(g)name(g,“Pharaon”)take(B,g,e1)

Fig5 }

Wecanseein thefinal routestructurethatsomeof theinformationintroducedin thedialogueis preservedandsomeis left out. Typically thepieceof informationkeptis whatis shared(i.e. accepted)by bothparticipants.

Thebuilding of thissharedinformationismainlyguidedby therhetoricalstruc-ture of the dialogue. However, trying to processit as a sharedplan within therhetoricalrepresentationof thedialogue,asin (AsherandLascarides1998),yieldsclumsyrepresentationsof the dynamicsof the processitself (e.g. for correction,no-answers,...),andobscuresthedifferencebetweenthecourseof a dialogueandits underlyingobjectives. We thereforeproposeto createanotherstructure,basedon thenotionof topic, for integratingthis information.

As wehaveseenbefore,in taskorienteddialogues(in whichrouteexplanationscanbeincluded)these“topics” comesfrom two sides: sentencelevel (see3.2)anddiscourselevel (see3.3). We think thatwe mustintegratethesetwo sidesof Infor-mationstructureto contribute to theconstructionof ahierarchictopic structure.

155

4 Constructing topic and rhetorical structur esin parallel

Our proposalis to enrichtherhetoricalstructurewith a parallelstructurekeepingtrackof topic constructions.For this we considertopicsaspropositionalcontentslinkedby specificrelations,calledtopical relations, eithertopic sequenceor topicelaboration. In classicalSDRT, theserelationsaresomewhat relatedto the planrelationsof SDRT (e.g. thereis a plan-elaborationrelation),but arenot separatedfrom rhetoricalrelations,while we think it is importantto distinguishthecourseofadialoguewith theconstructionof acommon-ground(whichis oftenthesharingorcommunicationof a plan in task-orienteddialogues).Besides,the topicsof somepartsof thedialoguesarejust thesepartsof the taskto beachieved. In this sensewe join the(TraumandNakatani1999)point of view on themulti-level approachof dialogue.Ourglobaltopicrecoversthemesoandmacrolevelsof TraumandandNakatani. Nonethelesstopic relationsintendto go beyond plansin task-orienteddialogue,by introducingamoregeneralinformationstructure.

An importantnotion for defining the topic structureis the resolvednessof atopic. Thecontentof a resolvedtopic is sharedbetweentheparticipants:this kindof topic is in the commongroundof the conversation. By contrast,the contentof an unresolved topic is still underdiscussion,it canbe pulled out of the topicstructure.Wewill markanunresolvedtopicwith “*” andweremove “*” whenthetopicbecomeresolved.

We now give a setof constructionrulesfor thetopic structure.Topic structureconstructionis donein parallelwith semanticinterpretationandestablishmentofrhetoricalrelations.For everyrhetoricrelationwehaveto triggeroff a topicupdaterule.{ Confirmation request(α,β) : unresolvestheα associatedtopic.{ Precisionrequest(α,β) : introducesa new unresolved topic, attachedto the

topicassociatedto α with aTopic-Elaborationrelation.Theactualcontentofthis topic will dependon thenatureof thequestion(Y/N, WH).{ Answer(α,β) :— to aconfirmationrequest: resolvesthetopic associatedto α.— to a precisionrequest: Resolves the topic associatedto α anddestroysunresolvedsub-topicelaboratingit.16 Moreoverwehave to addnew focused(information)from β in theDT.{ No Answer : This rhetoricrelationhasno effectson thetopicstructure.{ Acknowledge(α,β) : An acknowledgementresolvesthe topic associatedtoα.

16Wedestroy theunresolvedtopicbecausethey arenoteestablishedin thecommongroundandsothey don’t participatein theconstructionof theglobaltopicstructure

156

{ Continuation(α,β) : A continuationresolves the topic associatedto α andintroducesanunresolvednew topicattachedto thetopicassociatedto α by atopicsequencerelation.{ Correction(α,β) : A correctionreplacesin thetopic associatedto α thecor-rectedelementby thefocusof thecorrection(β)17.

Consideringthatwe have thecorrectionrelationin therhetoricalstructure(in-ferredwith anIS account);consideringtheprecedingrules;we give thefollowingconstructionfor the(24-26)turns.Thisstructurerecoversthemesoandmacrolev-elsof (TraumandNakatani1999).In ourstructure,themeso-level is down andthemacro-level is up.

* follow(B,a,e1) street(b)

until(a,b,c,d) name(b,"Carmes")

*

(25)

(24)

(26)

follow(B,a,e1)

until(a,b,c,d) name(b,"Carmes")

** follow(B,a,e1)

until(a,b,c,d) name(b,"Carmes") square(b) square(b)

Fig6 }

5 Conclusion

We have proposedherea way of structuringthe analysisof a dialogueby takinginto accountworks on topicsin a semantic-pragmaticframework. We adoptedastill-opentheory, theAsher’s SDRT, showing how to modify it to build in parallela dialoguestructureanda topic structurereflectingtheconstructionof a commongroundbetweentwo participantsin a task-orienteddialogue. Our approachhasremainedglobal(ignoringthespecifictaskof building a dialoguetopic from a di-alogueturn) andto a certainextent prospective. We claim that topicshave to beconsideredin a systematicmannerat any segmentof a dialogue,andnot only oc-casionallyto solve specificlocal problems.

One of our preoccupationsis to stay as closeas possibleto the real speechcorpus. Indeed,if we want to take InformationStructure into accountwe mustbepreciseandintegrateprosodicandintonationalcomponents.We don’t want toavoid the problemsrelatedto the oral natureof the corpusandwe think that onthis subject,we have to do a lot of work. In the future we plan to modify our

17The correction rule that we give hereconcernsa specificuseof correction. The area lot ofdifferenttypesof corrections(counterevidence, contradiction) whicharementionedin (Asher1998)(Prevot 2000).

157

transcripts.Insteadof usingthe“written” punctuation,wearegoingto usespecificspeechnotations,developedin speechcorpusdomain(Blanche-Benveniste.1997)(Valli andVeronis1999)for thespeechphenomenon(pauses,overlaps,truncatedwords).

Many openproblemsremainto befully investigatedin thatperspective. Firstsentencetopicandsentencefocushave to bemorepreciselylinkedto thediscoursetopic, to seehow they contribute to thewholestructure.Then,we have to seeinwhichcontexts thecomputationof discoursetopicsinteractwith thedeterminationandacceptabilityof rhetoricalrelations.We have so far only looked at the caseswhereselectingtheserelationsprecedesconstructingthetopics.

Bibliography

Asher, N. (1993).Referenceto Abstract Objectsin Discourse. Kluwer.

Asher, N. (1998).Commonground,correctionsandcoordination.submittedtoJournal of Semantics.

Asher, N., M. Aurnague,M. Bras, P. Sablayrolles,and L. Vieu (1995). Del’espace-tempsdansl’analysedudiscours.Semiotique(9), 11–63.

Asher, N., J.Busquets,andL. Vieu (to appear).La SDRT : uneapprochedelacoherencedu discoursdansla tradition de la semantiquedynamique.Ver-bum.

Asher, N. andA. Lascarides(1998).Questionsin dialogue.LinguisticsandPhi-losophy21, 237–309.

Asher, N. andP. Sablayrolles(1995).A typology anddiscoursesemanticsformotionverbsandspatialppsin french.Journalof semantics1(12),163–209.

Asher, N. andI. G. Txurruka(1995).ExtendingSDRT to integratea theoryoftheinformationalpartition.Technicalreport,ILCLI.

Blanche-Benveniste.,C. (1997). Approchesde la langueparlee en francais.Gap:Ophrys.

Denis,M. (1997).Thedescriptionof routes: a cognitive approachto thepro-ductionof spatialdiscourse.CPC4(16),409–458.

Engdahl,E. (2000).The role of syntacticfeaturesin the analysisof dialogue.Technicalreport,INDI.

Engdahl,E., S.Larsson,andS.Ericsson(2000).Focus-groundarticulationandparallelismin a dynamicmodelof dialogue.Technicalreport,TRINDI.

Ginzburg, J. (1994).An updatesemanticsfor dialogue.In Proceedingsof theinternationalworkshopon computationalsemantics, Tilburg.

Gronendjik,J.andM. Stokhof(1991).Dynamicpredicatelogic.LinguisticsandPhilosophy14, 39–100.

158

Grosz,B. andC. L. Sidner(1986,July-September).Attention, intentions,andthestructureof discourse.ComputationalLinguistics12(3), 175–204.

Kamp, H. andU. Reyle (1993).From Discourse to Logic. Kluwer AcademicPublishers.

Krause,P. (2000).A cooperative conventionnallanguagegamefor identifyingintendedreferent.Technicalreport,IMS.

Kruijf f-Korbayova, I. andG.Kruijf f (1997).Topic-focusarticulationin DRT. InProceedingsof the11thAmsterdamColloguium,December17-20.

Lambrecht,K. (1994).Informationstructure andsentenceform.CUP.

Mann, W. andS. Thompson(1987).Rhetoricalstructuretheory : a theoryoftext organization.Technicalreport,InformationScienceInstitute.

Muller, P. (1998).Elementsd’une theorie spatialedu mouvementpour la for-malisationduraisonnementspatio-temporel desenscommun.Ph.D. thesis,Universite PaulSabatier, Toulouse.

Poesio,M. andD. R. Traum(1997).Conversationalactionsanddiscoursesitu-ations.ComputaionalIntelligence13(3), 309–347.

Prevot, L. (2000).Lesdialoguesd’orientation,uneapprochepar la semantiquedu discours.Master’s thesis,Universite PaulSabatier, Toulouse.

Traum,D. andC. Nakatani(1999).A two-level approachto codingdialoguefordiscoursestructure: activitiesof the1998dri workinggrouponhigher-levelstructures.In Proceedingsof theACL’99 WorkshopTowardsStandardsandToolsfor DiscourseTagging, pp.101–108.

Vallduv, E. andE. Engdahl(1996).The linguistics realizationof informationpackaging.Linguistics34, 459–519.

Valli, A. andJ. Veronis(1999).Etiquetagegrammaticaldescorpusde parole:problemeset perspectives.Revuefrancaisede linguistiqueappliquee. 4(2),113–133.

van Kuppevelt, J. (1995).Discoursestructure,topicality andquestionning.J.Linguistics31, 109–147.

vanKuppevelt, J. (1998).Topic andcomment., pp. 191–197.Elsevier Science,Oxford.

159

Presuppositionor Abstract Object Anaphora?:Constraints on Choiceof FactiveComplementsin SpokenDiscourse

JENNIFER SPENADER

ComputationalLinguistics,StockholmUniversity

[email protected]

ABSTRACT. Usingresultsfrom anempiricalstudyof factivesin spokenEnglish,acomparisonismadebetweenfactive verbsandtheirpresupposedcomplements,bothboundandaccommodated,andtheirnon-presupposingalternatives: abstractobjectanaphorathatderiveanantecedentfrom adiscourse-given linguistic expression.Therole of discoursestructuralconstraintsandgroundingin choiceof expressionandthecontexts in which they areusedis discussed.Finally, theresultsarerelatedto theoreticalissuesin presuppositiontheoryhaving to dowith thedifferencebetweenanaphoraandpresuppositions,the ability of presuppositionsto accommodateandthe natureofaccommodation.

Intr oduction

Theaimof thiswork is to give a betterdescriptionof thedistribution andfunctionin spoken dialogueof oneexpressiontype in onespecificconstruction,comple-mentsof factive verbsthatcanbeusedto referto anabstractobject.

Factive verbspresupposetheir sententialcomplements.For introducingthesameinformation the non-presupposingalternative is to assertthe information,makingit part of the discourserecord,andthenrefer to this informationwith anabstractobjectanaphoricexpressionsuchasapronounor definitenounphrase.Anexamplemay make thingsclearer. Herefactive verbsaremarked with bold andabstractobjectanaphoraarewritten in CAPITALS.

Example (1)a. Computational linguists are in demand.b. Students apply for our program because they knowthat (computational linguists are in demand).c. Students apply for our program because they know THIS.d. Students apply for our program because they havenoticedTHIS TREND � THIS FACT.

161


Sequenceab illustratesa boundpresuppositionusagewith the presuppositioninparentheses,ac illustratesabstractobjectanaphoricreference,asdoesad, thoughhereafull NPis used.It is possibleto communicatethesameinformationby usingb alone,in which casethe presuppositionwould be consideredto be accommo-dated.In orderto learnmoreabouttheuseof thesedifferentexpressionsin naturalspokendiscourse,acorpusstudywasdone.

1 Background

Thisfirst sectionexplainstheanaphorictheoryof presupposition,andhow it func-tions with the factive verbsstudiedhere. The secondsectionpresentsabstractobjects,andhow they canbe referredto aswell asproposedconstraintson theirusagebasedon discoursestructure.

1.1 PresuppositionsasAnaphora

Theanaphorictheoryof presupposition,developedby vanderSandt(1992),arguesthatpresuppositionscanbe treatedjust asanaphoraaretreatedin DRT (Kamp&Reyle 1993).Presuppositionresolutioninvolvesexaminingthepreviousdiscoursecontext for an antecedent.If an antecedentis found then the presuppositionsisboundto it. If anantecedentcannotbefoundthenthepresuppositionis accommo-dated,andtheability to accommodateis whatdistinguishespresuppositionsfromotheranaphoricexpressions.

Factive verbspresupposefull propositions,andit is not immediatelyevidenthow binding shouldbe identified here. The potentialantecedentmust be suffi-ciently similar to the presupposedpropositionso that the latter cantruly be con-sideredto functionasanantecedent.Exampleswherethepotentialantecedentisidentical with the presupposedproposition,as in Example(1) ab, are probablyrare. Speakerstendto vary their speechandavoid uninformative repetitions,andthis meansthat therearepotentiallygreatdifficulties in identifying propositionalpresuppositionalbinding.

The ability of presuppositionsto accommodateis suggestedin van der Sandt(1992)to berelatedto their greatersemanticcontentandinternalstructure.Otheranaphoricexpressionsaresaidnot to beableto accommodate.

Despitethepopularityof theanaphorictheoryof presupposition,therehasnotbeenmuchwork that studiespresuppositionalusagefrom the sameperspectivesthatareconsideredcentralto characterizingtheuseof anaphoricexpressions,suchasthechoicebetweena presuppositionor a non-presupposingform, e.g.,parallelto thechoicebetweena pronounor a name,or constraintson accessibility1 of po-tentialantecedents.Generally, in theoriesof anaphoraaccessibility, thenumberofpotentialantecedentsin thepreviouscontext decreasesasthesemanticcontentof

1HereI meanaccessibilityin termsof salience,attentionalstate,etc.,andnot thestructuralac-cessibilityof discoursereferentswithin embeddedDRSsasthetermis usedin DRT.

162

theanaphoricexpressionincreases.This in turn allows a greaterdistancebetweenanaphorandantecedentwith semanticallyhighly specifiedanaphorabecausetheantecedentshouldbeeasyto identify evenwith a certainamountof distance.Be-causethepresuppositionsof factivesaresemanticallyvery rich, wecanbegin withthe hypothesisthat they will allow longerdistancesbetweenantecedentandpre-supposedpropositionsthanabstractobjectanaphoraandthelinguisticexpressionsfrom which their antecedentsarederived.

A puzzlingquestionthough,is why speakerswould ever chooseto usea fac-tive verbwith a presupposedsententialcomplementif it will getbound,i.e. whyrepeatan alreadygiven proposition?This would seemto breakthe informativityconstraintondiscourse.RepeatedlyusingdefiniteNPs,for example,to referto al-readyintroducedentitiesseemsmorenaturalbecausewe oftenneedto referto thesamediscourseindividualsor objects,in orderto describethemin new situationsandrelationships.But propositionswouldseemto beunnecessaryto repeat.

However, Walker (1996)hasshown thatinformationallyredundantutterances,or IRUs,arenotatall infrequentin spokendiscourse,citing thatthey madeupabout12 % of the utterancesin her corpusstudyof collaborative taskdialogues.Shegivesthreecategoriesof communicative functionsbasedonthecorpusstudy, thesewere1. Attitude: whichseemsto betheuseof anIRU for explicitly grounding,2.Attention: useof an IRU to make a propositionssalient,or 3. Consequence: anIRU is used”to augmenttheevidencesupportingbeliefsthatcertaininferencesarelicensed”(Walker, 1996,p. 5).

Canboundpropositionalpresuppositionsbe consideredto be informationallyredundantutterances?I think their statusis somewhatdifferent. Examinethefor-maldefinition(takenfrom Walker(1996),p. 5): An utteranceui is INFORMATION-ALLY REDUNDANT in adiscoursesituationS

(i) if ui expressesa propositionpi , andanotherutteranceuj thatentailspi hasalreadybeensaidin S.(ii) if ui expressesapropositionpi andanotherutteranceuj thatpresup-posesor implicatepi hasalreadybeensaidin S.

Presuppositionsdiffer from IRUsin two importantways.Thefirst is thatthey havebeenpresupposed.Thespeaker is thereforeexplicitly signallingthat the informa-tion is known, givenandshouldbeacceptedasbackgroundedby thehearer. Theseconddifferenceis that presupposedpropositionsoccur in utterancesthat alsohave an assertionalcomponent,in which casethe utteranceitself is not informa-tionally redundant,but thepresupposedpartis.

Still, even if the useof given or boundpresuppositionsis distinct from theuseof IRUs, thefunctionsof propositionalpresuppositionsthatareboundmaybesimilar. Hopefully empiricalstudycananswersomeof thesequestions.

163

1.2 Abstract Object Anaphora and DiscourseStructure

Thetypeof higherorderentity anaphorabeingreferredto hereasabstractobjectanaphorahasbeencalled many different things in the literature,e.g. discoursedeixis(Webber1991,Eckert & Strube,2001),situationanaphora(Fraurud,1992),andfinally, asabstractobjectanaphora,by Asher(1993).

All thesetermsareusedto refer to anaphoricreferenceto situations,factu-alities, eventualities,propositions,speechacts,aswell asmoredeictic functionsreferring to the referenceof a discoursesegment,thoughmany of thesedistinc-tive usesareoftendifficult to distinguishin practice.We canoftenfind examplesof differentaspectsof thesameantecedentbeingreferredto by thesametype ofanaphoricexpression,and wherethe actual type of the referentis codedin thepredicationof thesentencein which theanaphoricexpressionis used.This obser-vationis madeby Asher(1993),Dahl& Hellman(1995)andis animportantpartofEckert & Strube(2001).Thetypeof entity alsomaybeindicatedby theNP used,as in Example(1) d. The continuationof Example(1) below illustratessomeotherpossibilities.Thesequenceae indicatesthatwe shouldconsiderthetypeofreferentaproposition,becauseusuallyonly propositionscanbeconsideredtrueorfalse.Continuingwith thesequenceaf meansthatthetypewill bea situation,ascodedin thenounphrase.Eckert & Strube(2001)usethis informationto distin-guishbetweenindividualandabstractobjectanaphoricreferencewhich thenhelpsguidetheir resolutionalgorithm.

Example (1) continuede. Everyone at the university knows that THIS is true.f. THIS SITUATION has led to an alarmingly high drop-out rateas students leave school to take industry jobs.

Dahl & Hellman(1995)discussabstractobjectreferenceasa typeof anaphorthatinstigatesa processof reference-creation, or referencecoercion;that is, usinganabstractobjectanaphorsignalsto theinterpreterto look for establishedinformationof theappropriatetypein thediscourserecordandcreatea discoursereferentthatcanfunction asan antecedentfrom this linguistic information. Dahl & Hellman(1995)further list threereferentcreatingoperations,1) SummationandComplexCreation,2) ’Type-coercion’and3) AbstractionandSubstitution.Theprocessofreferent-creationhasbeencomparedto thatof accommodationby Eckert& Strube(2001), thoughit is unclearif this shouldbe consideredaccommodationin thesamesenseasit is usedwithin theanaphorictheoryof presupposition.For presup-positionalaccommodation,new informationis addedto thediscourserecord,butin referent-coercion,alreadygiveninformationis consideredin a new way.2

2Someresearchershave arguedthat thereis only accommodationwhenthe informationis new,andinformationthatis derivablefrom analreadygivenrepresentationis technicallynotnew. HereIthink it is unclearwhatthestatusof informationthat is availablein thediscourserecordin differentformsactuallyis. Developinga distinctionsimilar to that in computersciencebetweenInformation

164

Abstractobjectshave beenshown to beseverelyconstrainedin their ability tobeaccessedby anaphoricmeans,andtheselimitationsseemto bedueto discoursestructuralconstraints.For example,Webber(1991)hasarguedthatabstractobjectanaphoracanonly accessdiscoursesegmentson theright frontier of thediscoursestructurefor referents;discoursesegmentson the right frontier arealsothoseinfocus,or salient.

Fraurud(1992)hasarguedthatWebber’s descriptionof constraintson acces-sible antecedentsis for the mostpart correct,but that in addition to unfocusseddiscoursesegments,propositionsthat aresubordinateto a main propositionalsoneedto be excludedfrom supplyingan antecedentfor abstractobjectanaphora.Shegivesexamplesthat show that the simpletreestructuresusedby Webbertorepresentdiscoursestructuredon’t correctlycapturethe full rangeof distinctionsthatwill be relevantbecausethe right frontier constraintdoesn’t give a meansbywhichto factoroutsubordinatedor modifyingpropositionsthatareconjoinedwithamainpropositionaspartof theright frontier. Makingadistinctionbetweenmainandsubordinatingpropositionswould solve this problem,thougha differentdis-coursestructurewith differentaccessibilityruleswouldbenecessary.

As this earlierwork indicates,discoursestructureis generallyagreedto con-strainaccessibilityfor abstractanaphoricreference.In orderto beableto discusspossiblediscoursestructuralconstraintson accessibility, it is necessaryto definewhat units of discoursestructurewill be consideredrelevant, i.e. the discoursesegmentation.Sofar, mostwork on discoursestructure(includingthatmentionedabove)hasfocussedonwrittendiscourse,andwork onspokendiscoursehaslookedmainlyatutterancesandhow they form adjacency pairs,only sometimesconsider-ing amorehierarchicalstructure.It is notclearwhatlevelsof structurearepresentin spoken dialogue,nor what units will be relevant to the discourseparticipantsanddiscourseinterpretation.Generally, the levels thathave beendiscussedin theliteratureincludeaspeechactor discoursemove level. Theseareusuallythendis-cussedasadjacency pairsor asinstancesof dialoguegames,e.g. asin Carlettaetal.(1997),or calleddiscourseunits.Finally, ahigherlevel is sometimespostulatedthat in Carlettaet al.’s codingschemehasto do with domainspecificgoalsin thediscourse.Which, if any of theselevels, is relevant for characterizinganaphoric(andpresuppositional)accessibilityis still unclear.

Eckert& Strube(2001)is thereforeparticularlyrelevantto thework herein thatthey have doneempiricalwork on thespecialproblemsof anaphoricresolutioninspoken discourse. Their work is a corpusstudy of the SWITCHBOARD corpus,a spoken corpusof telephoneconversationsbetweentwo participantswho wereunaquaintedwith eachotherbeforetheir conversation.

Eckert & Strubeusea very simplified discoursestructurebut seemto get re-liable resultsin usingthis simplifiedstructureto delimit potentialantecedentsfor

ExtractionandDataMining, i.e. betweenretrieving informationyouknew existedandgleaningnewinformationthatyouweren’t awareof from anexistingknowledgerepresentation,mightbehelpfullyrelatedto accommodation,anddiscourseinterpretationin general.

165

personalpronounsaswell asfor abstractanaphoricpronominalreference.Theyanalyzedeachutteranceas being a dialogueact of one of threetypesbasedonthetop dialoguemovesin Carlettaet al.(1998):Initiation, I, Acknowledgment,AandA/I for thoseutteranceswhichservedbothaninitiating andanacknowledgingfunction. Thesedialogueactsarethenpairedinto what arecalledsynchronizingunits(SU’s). Certaindialogueactsdon’t requireacknowledgement,andthesemayberepresentedby assingleI andstill beinterpretedasanSU. Centralto theirworkis the ideathat groundingconstrainsthe accessibilityof antecedents.Groundingis the processby which discourseparticipantssignal that informationintroducedin the discourseis part of the commonground(Traum 1994), by somesort ofacknowledgment.A ’s (Acknowledgments)aregroundingacts. Discourserefer-entsor information introducedin utterancesthat are not acknowledged,that is,not grounded,arenot availablefor anaphoricreference.Notehowever thatwhileacknowledgementis a signalof grounding,lack of acknowledgmentis not neces-sarilyassignof lackof grounding,andthelackof aprotest,or allowing thespeakerto continueto have thefloor, canbeconsidereda typeof implicit grounding.Eck-ert & Strube(2001)alsogiveanexamplewhereaclearindicationthataSpeaker’sutterancewasnot groundedby theotherparticipantmeansthat thediscourseseg-mentandreferentswithin thatsegmentarenot addedto thecommonground,andthereforearenot available for anaphoricreferenceof any kind. In this way theydefinewhatis salientor in focusasthatwhich is in commonground,andtheneedto clearlyidentify groundingis alsoreflectedin how they chooseto codediscoursestructure.

2 Empirical Data and Method

Examplesof factiveverbsin context wereexcerptedfrom theLondon-LundCorpusof SpokenEnglish(LLC)3. 50multi-speakerdialogueswereused,whichcontainedroughly233,000words.Thefollowing factive verbswereexcerpted:

Factiveverbs: subjectcomplementscount,makesense, suffice, amuse,bother, matterFactive verbs: object complementsdiscover, find out, know4, notice,realize, regret, resent,see

Exampleswithout a complement,or with anNP-object,non-abstractobjectcom-plement,werediscarded.For eachexample,therelationbetweenthepotentialan-tecedent,if therewasone,5 anddiscoursestructurewasexamined,notingspeaker

3Information on obtaining this corpus can be be found on the ICAME website athttp://www.hd.uib.no/icame.html

4”know” with a sententialcomplementnot markedby that is not includedin theanalysisduetothe greatnumberof falsehits found becausethe corpusis untagged,meaningthat theseexamplesmustgonethroughby hand.Theauthorplansto do thisat a laterdate.

5Presupposedpropositionswithoutantecedentsareaccommodatedfor thatvery reason.

166

TYPE TOTAL ACCOM PRESUPP BOUND PRESUPP ABS-OBJ ANA

objectcomp 75 48 6 20subjectcomp 13 0 0 14TOTALS 88 48 6 34

Table16.1:Typeof AbstractObjectComplement

shiftsbetweenantecedentandanaphor, distance,in utterances,andany otherrele-vantaspects.Theanalysiswasdonefrom theperspective of theannotator.

The dialoguesstudiedseemedto be morecomplicatedthanthosestudiedbyEckert & Strubein thatmany werebetweenmorethantwo participants.Becausethey take placein personso turn-takingis lesspreciseandtherearemany casesof overlappingspeech.Additionally, theparticipantsoftenknow eachotherwell,whichalsoseemsto supportinterruptionsandoverlaps.

To adaptEckert & Strube’s codingof discoursestructureto dialoguesbetweenmorethantwo participants,thefollowing guidelineswereused:An utterancemaybegroundedby morethanonespeaker andthis wasoftenthecase- soa synchro-nizing unit (SU) canbe madeup of a sequenceof multiple A’s as long as theyareacknowledgingthe sameI. Sometimesonespeaker will clearly acknowledgeanotherspeaker while aftera third speaker hascontributedan I, which meansthatSU’s mustbe ableto overlap(crossstructures)(e.g. we canhave Speaker A: I1

, Speaker B: I2 , Speaker C: A1 , Speaker A or C: A2 ). Segmentationwasdonemakingeachnew turn a new utteranceandsplitting a turn into morethanonceactif partof theturn clearlyonly hasanacknowledgementfunction. Of course-if anutterancewith anabstractanaphoricreferenceoverlapswith anotherutterancethiscannotbea potentialantecedentasit is not yet partof thediscourse,andfor thisreasonoverlapsmustbetakeninto consideration.

3 Results

Table1 presentsthe results. By far the most frequentusagewasa presupposedpropositionthathadto beaccommodated(48 examples).6 Looking at anexamplewouldn’t reallycontributeto thediscussionheresobecauseof spacelimitationsnoexampleswill begivenhere.

The next largestgroupfound wereabstractobjectanaphoricreference.Hereparticularattentionwaspaid to the discoursestructureandwhetheror not utter-ancesweregrounded,usingthemodifiedcodingsystembasedon theonegiveninEckert & Strube(2001)anddescribedabove.

6Note that accommodationcanoccurglobally (39 examples),or intermediatelyand locally (9examples)accordingto vanderSandt’s (1992)theory.

167

Example (2) Samespeaker, previous discoursesegment,chain of referencesSpeaker A: So that it’s the faculty of arts, I �or the faculty of economics or both that’ll SUbe putting him forward (1)Speaker B: Mmm. (2) A �But they can put it forward for any titlethat they like apparently.(3) II didn’t realize THIS. (4) I � SU*1 to 2 sylls*. so this �Speaker A: No, I didn’t know THAT. (5) A

Here,utteranceswereconsideredto bediscoursesegments.In theexampleaboveit seemsclearfrom thespeaker’s useof ”No”, in ”No, I didn’t know that,” thatheis referringto B’s lastutterance,andnot to theentireinformationalcontentof whatB hasjust said. THIS (4) refersto an abstractobjectderived from the linguisticexpressionin discoursesegment(3). THAT in (5) referseitherto thesamething asTHIS in (4), makingthisachainof abstractreference.

Example(3)7 Overlapping speech,sourceof abstract object could besynthesisof several utterancesmadeby different speakers

Speaker C: University of the Air (1) ISpeaker D: *that would be S* (2) n.g. �Speaker C: *are doing a series* on various Isorts of � communication which struck me SUimmediately as **disparates** (3) �Speaker B: � disparates, surely � ? (4) ASpeaker D: ** � but but � ** but it would be,um it would be non-surreptitious Iwouldn’t it ? (5) �Speaker C: *presumably* (6) ASpeaker B: *yes -*, it would be SU� non-surreptitious� (7) A �Speaker A: Yes, (8) Abut THAT wouldn’t matter. (9) A/I

In the above example,four differentspeakers take part, andidentifying synchro-nizing units wasnot totally straightforward. Eckert & Strube’s codingsystemisexpandedto allow differentspeakersto eachgroundthesameutterance(here(6),(7), and(8), grounding(5)) andstill considerit to beoneSU. Utterance(2) is la-belledas”n.g.” for ”not grounded”andthis discoursesegmentdoesnot introduceany referentsavailablefor anaphoricreference.Note that it overlapswith partof

7Note that thediacritic marksthatencapsulatepartsof theutterancesmarkwherethespeechofspeaker’s overlapped.Here,for example,”that would beS” and”aredoinga series”overlapped.

168

(3). HereI have split Speaker A’s utteranceinto two discoursesegments,(8) and(9), because(8) seemsto solelyhave a groundingfunction,whereas(9) seemstobe informative, thoughit is not clear if this is the correctsegmentation.SpeakerA’s abstractanaphoricreferencein (9) seemsto referto theimmediatelyproceed-ing groundedSU (5-8), or could perhapsbe consideredto referring to only thegroundedportion, (6-8). It is not clearwhatanalysiswould becorrect,thoughinbothcasestheabstractobjectanaphorderivesits antecedentfrom animmediatelyprecedingdiscoursesegment,within the sameSU, and it is impossiblethat theanaphoracouldbereferringto thepreviousSU (1-4).

Theformsof abstractobjectanaphorausedmayalsobeof interest.Therewere8 casesof it, 2 casesof this, 18 casesof that and3 casesof zero anaphora and2casesof definitenounphrases. For almostall examplesa linguisticexpressionthatcould be a sourcefor anabstractobjectcould be found in theprevious SU, or inthe sameSU, thoughtherewerea few exceptions.In 19 casesthe samespeakerwhoutteredtheabstractobjectanaphoralsohadsaidthelinguisticexpressionfromwhich its antecedentcanbederived; in 12 remainingcasesthespeakersweredif-ferent,andin 3 casesit is impossibleto pinpointexactlywhatutterance(s)providedtheantecedentare(cf. Eckert & Strubewho alsofounda greatnumberof vagueor difficult to identify abstractanaphoricoccurrences),thoughit is clearthat it ispart of the previous context (e.g. previous or sameSU), it is often (asillustratedin example(3)) a questionof determininghow much of the previous context isintendedas the antecedent.It alsoseemedthat the simpleanalysisof discoursestructuremodifiedfrom Eckert& Strubewashelpfulandadequateto organizeandunderstandthedata.

The next exampleillustratessomethingthat could potentially be consideredpresuppositionalbinding. In determiningpresuppositionalbinding theentiredis-courserecordup until the useof the factive was taken into consideration.Thecriteriausedto determineif bindingwasa potentialanalysiswastheauthorsownintuitionsasto whetherthepresupposedinformationwasnew, or hadalreadybeengivenin someform in thediscourse.

Example (4) Presupposedpropositions- boundSpeaker A: It was lethal to expectant motherswith small children. (1) (38 intervening lines of text).Speaker A:After all, I mean you can’t go down and shop if youKNOW that you’re going to knock out an expectant mother ��it was some violent streptococcus that he’d got (2)

Herethepresuppositionis that”you” (or anyone,in a genericsense)would knockout an expectantmotherif having beeninfectedwith the streptococcusreferredto by ”it” in utterance(1). Here it is arguablethat the informationpresupposedin (2) is not really new, in that utterance(1) refersto the samesituation,thoughin a different way and with someother conclusionsthrown in. Note that therereallyseemsto benopoint in doingadiscoursesegmentanalysishere,becausethe

169

intervening38 lines meansthat the presupposedpropositionis very far removedfrom utterance(1). Anotheroneof the examplesof potentialbinding alsohada900line gapbetweenpotentialantecedentandpresupposition.

4 Discussion

To summarizetheresults:Factiveverbs,for theusesstudiedhere,overwhelminglytend to occur with full propositionalpresuppositionsthat needto be accommo-dated,i.e. they presupposediscoursenew information. Factive verbsalsoappearwith abstractanaphoriccomplementsthat thenrefer to abstractobjects. The useof factive verbswith presupposedinformationthat is alreadypartof thediscourserecord,e.g.presuppositionalbinding,is minimal (only 6 examples!).

Thebindingexamplesaremostinterestingto discussfirst asoneinterestingre-searchquestionwaswhy speakerswould chooseto usefull presuppositionswhenthepresupposedinformationcanbeconsideredto bealreadypartof thediscoursecontext. Theanswerseemsto bethatthey seldomdo,but whenthey do, thefunc-tion of thepresuppositionin thediscourseseemstentatively to beoneof the fol-lowing:

(1) The bound presupposition has a summation function Full sententialcomplementsareusedto explicitly expressasa whole an idea that waspresentonly in bits andpiecesin the earlierdiscourse,perhapseven contributed by oneor morespeakers. This usageseemsto relateto thereferent-creatingoperationofSummationproposedin Dahl & Hellman(1995),andmayalsobetheeffect of themulti-speaker discoursesettingin that information is being contributed from somany differentdirections.

(2) The bound presuppositionstatessomekind of conclusion that is de-ductableor inferrable fr om the discourserecordFully inferrableinformationisconsideredto be known information. If a strict definitionof accommodationthatlimits its applicationto new materialis used,thenthepresupposedpropositionisknown andmustboundbecauseit mustbetruein eachattentive speakersinforma-tion state. However, it is beingexpressedin the discourseexplicitly for the firsttime here. All conclusionsmay or may not have beenrealizedby all discourseparticipants.Multi-speaker dialoguemayneedto usepresuppositionsthiswaybe-causeestablishinginformationasmutually known is a morecomplex taskwhenseveraldiscourseparticipantsareinvolvedthanwhenonly two participantsarein-volved. In fact, it would be strangeif we would not needto explicitly concludethingson occasion,asa form of grounding.

(3) The bound presuppositionhas another pragmatic function than theoriginal, assertedusage. Therewere two exampleswherethis could be consid-eredto have beenthe case. In oneexamplethe presupposedinformationwasanearrepetitionof the other speakers immediatelyproceedingstatementandhadthefunctionof showing agreement.In theotherexample,therepetitionwasusedto relatewhat anotherspeaker hadsaid,so shouldperhapsnot be consideredas

170

very goodexampleof apresupposition.8

(4) The linguistic expressionfrom which the abstract object could be de-ri ved is in a discoursesegmentthat is no-longer accessiblefor reference.Thiswasmy original hypothesisaboutwhenspeaker would choseto usea boundpre-supposition,given the greatnumberof constraintson abstractobject anaphoricreference.Indeed,for fiveof thesix examplesfoundhere,it wouldbeawkward,ifnot impossible,to referto thesameinformationwith anabstractobjectanaphor.

Two of thecommunicative functionsof IRUs identifiedby Walker (1996)(seesubsection1.1) were quite similar to the usagesidentified above. Category (1)seemsvery closeto theConsequencecommunicative functionandAttitudeseemssimilar to category (2), a kind of groundingfunction. Attention, would have mostlikely fallenundercategory (4), but therewereno examplesof category (4) found.

In light of the infrequentcasesof potentialbinding,andthedubiouswaysinwhich thepotentialantecedentsoftenwererelatedto thepresupposedproposition,is it defensibleto considerthecasesfoundasactualbinding?

If we take the ideaof presuppositionsasanaphoraseriously, thenperhapsweshouldalsoseriouslyentertaintheideathatbindingpresupposedpropositionalin-formationmayalsobeconstrainedby thesamediscoursestructuralconstraintsasabstractobjectanaphora.Until now I have beenworking underthenaive assump-tion thatpresupposedpropositions,becauseof their greatersemanticcontent,willfreely allow referencefrom almostany positionin thediscourse.9 Usuallythedis-coursestructuralconstrainton accessibilityof anaphoracanbe testedby creatingexamplesentencesandconsideringwhat interpretationthe anaphoricexpressionwill get. Unfortunatelywe cannottestfor constraintson theaccessibilityfor pre-supposedpropositionsin thiswaybecausewe can’t reallydistinguishbetweenourown processesof bindingor accommodation.

While wecan’t reallytestaccessibilityin thewaywecanfor abstractanaphoricreference,wecanhowevermakeconclusionsaboutwhetheror notthepresupposedinformationhasenteredinto thecommongroundof thediscourseparticipants,themostimportantcriteria for anaphoricreferenceaccessibilityin Eckert & Strube’sstudy. Indeed,theexplanationsfor theuseof a boundpresuppositiongivenin (1),and(2) above could both be describedascaseswherethe speaker wasunsureifthepresupposedinformationhadenteredinto thecommonground,andthathis/herusageof a presuppositionwith this informationwasin a sensea way to groundit.Presupposingasan informationpresentationdevice may be quite apt in that theinformationpresentedis uncontroversial informationthat may alreadybe knownto somediscourseparticipants.

Theanalysisof theexamplesof bindinggivenabove canthenberevised.Par-allel to constraintsonabstractobjectreference,wecouldconsiderthattheseexam-

8In thisexamplethespeakerfirst describedasituationwith aproposition,andthenrelatedastoryaboutanotherperson’s reactionto thesamesituation.Whenreportingontheotherperson’s reaction,sheusedthesameproposition,but asthecomplementof a factive verb

9Excludingof coursestructurallyinaccessiblepositionsdueto context createdby logical opera-tors,suchasmodalcontexts,modalsubordinatingcontexts,andbeliefcontexts

171

plesrepresentinformationthat is not consideredby theparticipantsto have trulybeenestablishedin the commonground,and thereforeis unavailable for bind-ing. We could alsodisallow binding for someexamplesbasedsolelyon the ideathe greatdistancesbetweenpotentialantecedentandpresuppositionmake refer-enceunacceptable.In an interestingtwist, taking the analysisof presuppositionsasanaphorafull out, andapplyingthesametypesof constraintsto them,actuallygivesusananalysiswherepresuppositionsandanaphoricalternativesareusedinfunctionallycomplementarydistribution, but not asfirst thought.Presuppositionspresentnew informationor groundunclearinformation,abstractobjectanaphorain the samecontext (here,as the complementof factives) refer to alreadygiveninformation.

If thecasesearlieridentifiedasbindingarenow categorizedasaccommodation,thenall examplesfound in thedataof presupposedpropositionswereaccommo-dated.Whatdoesthis tell usabouttheuseof factivesin particular, andthenatureof presuppositionaccommodationin general?

Remember, vanderSandt’s (1992)theorytells us thatpresuppositionsshouldtheoreticallybeableto bothbind andaccommodate,andthattheability to accom-modatehassomethingto do with thesemanticcontentof thepresupposition.Theconclusionsreachedheresuggestthat in practicenot all triggersareusedto doboth. And for factivesparticularly, binding doesnot seemto be a normalusage.Factiveverbswith apresupposedcomplementarethenusedprimarily to introducediscoursenew information.

In terms of how much semanticinformation is being presupposed,factiveverbs, becausethey presupposeentire propositions,must be one of the richesttriggers.Becauseanaphoricexpressionsandpresuppositionswith lessersemanticcontentareconsiderednot to beableto accommodate,or to accommodatebadly,it is temptingto interpretthe resultshereasevidencethat thereverseis alsotrue:Presuppositiontriggerswith a rich semanticcontentcannot only accommodatewhennecessary, but this is theirpreferredusage.

All this hasrelevanceto our view of thenatureof accommodation.Therearetwo competingviews of accommodation:thatit is arepairstrategy, andsomethingthatshouldbeavoidedandtheview that it is a normalmethodof communication.Most work on presuppositionsseemsto have takentheformerview. Of course,inpresuppositionresolutionbinding mustbe preferredover accommodation,hencethepreferencesin vanderSandt(1992)andin Blutner(2000)’s constraintAVOID

ACCOMMODATION in his bi-directionalOT treatmentof presupposition.Zeevat(2001),building on Blutner(2000)hasarguedthatit is not therichnessof seman-tic contentthatdetermineswhatexpressionscanaccommodate,but theavailabilityof non-presupposingalternatives. If an alternative is available,AVOID ACCOM-MODATION will inhibit, or evenblock thespeaker’s useof a trigger to beaccom-modatedbecausethereis a simpleralternative. Thespeaker shouldchooseto usethesimpleralternative, andnon-presupposingalternativesareby their very naturesimpler. For factives,thereis anexpressionalternative: assertandthenreferwithan abstractobjectanaphora,but this alternative hasbeenshown to not alwaysbe

172

available,andthereforeit maybeincorrectto considerit atrueblockingalternative.Thisview of accommodationis still onethatconsidersit to bearepairstrategy

by the hearerwhena non-presupposingalternative wasavailable to the speaker.Accommodationis consideredto bea costlymethodof communicationthatcouldlead to misinterpretation,and somethingthat both speakers and hearersshouldavoid.

Again,I think theresultsheresuggestadifferentview. Accommodationis bothanexploitable10 communicationstrategy andarepairstrategy, but it dependsonthetrigger involvedandthecontext. Sometimesaskingyour listenerto accommodateis thebestmeansby which to expressyour ideaandwhenthesemanticcontentofthe trigger is rich enoughthataccommodationcanproceedwithout thedangerofmisinterpretationon thepartof thehearer, thenit is themostoptimalway to com-municate,andthatis why factiveswith presupposed,accommodatedcomplementsarethenorm.

Sothepresuppositiontriggersthatarebestequippedfor accommodation,e.g.thosewith rich semanticcontentandstructure,will alsobeexploitedby speakersto the fullest as this will be the mosteffective andeconomicalway to introduceinformation. In fact,moreeffective thannon-presupposingalternatives in certaincases.Thosetriggersthatareworstequippedfor accommodation,e.g.thosewhosemeaningis underspecifiedto thedegreethataccommodationis a strainon thelis-tener, anda real potentialsourceof confusion,e.g. mostpronominalanaphora,will not be exploited by speakers, thoughthesewill be ableto be interpretedbylistenersby accommodationwhennecessary- and then it is beingusedasa re-pair strategy. Accommodationshouldbeconsideredto beanavailableoption forhearersbothfor anaphoraandpresuppositions,but speakerswill tendto limit theirexploitationof thehearer’s ability to accommodateto semanticallyrich anaphoricandpresuppositionalexpressions.

5 Futur e research

Differencesbetweenotherpresuppositionsandtheirnon-presupposingalternativesshouldbelookedatmorecarefully. Theresultshereshouldalsobecomparedwithwritten discourse,in particularbecausewritten discourseseemsto have a morehierarchicaldiscoursestructurethandialogue,andit would be interestingto seehow thiswouldaffect thechoiceto usepresupposedcomplements.

Acknowledgment

I would like to thankBart Geurts,Emiel KrahmerandSofiaGustafson-Capkovafor discussingthis work with me. Thanksalso to the anonymousreviewers for

10Actually Lewis’s (1979)original paperon accommodationseemsto characterizeit asa typeof’exploitation’ of thehearer’sability to makecertaininferencesby thespeaker, hencethemaster-slaveanalogyheuses.

173

helpful comments.All errorsaremy own.

Bibliography

Asher, Nicholas(1993).Referenceto AbstractObjectsin Discourse,Kluwer Aca-demicPublishers.

Blutner, Reinhard(2000). SomeAspectof Optimality in NaturalLanguageInter-pretation,In Optimality Workshop,Utrecht,January2000

Eckert, Miriam & Strube,Michael (2001). DialogueActs, SynchronisingUnitsandAnaphoraResolution,to Appearin Journalof Semantics.

Carletta,Jean,Isard,Amy, Isard,Stephan,Kowtoko, Jacqueline,Doherty-Sneddon& Anne Anderson(1997). The reliability of a dialoguestructurecodingscheme.ComputationalLinguistics,23: 13-32

Dahl, Osten,& Hellman, Christina. (1995). What happenswhen we use ananaphor?.Presentationat theXVth Scandinavian Conferenceof Linguistics,Oslo,Norway.

Fraurud,Kari (1992).SituationReference,(Whatdoes’it’ referto?). GAP Work-ing PaperNo. 24. Hamburg: FachbereichInformatik,Universitat Hamburg.

Kamp,HansandReyle,Uwe(1993).FromDiscourseto Logic,Dordrecht,KluwerAcademicPublishers

Lewis, David (1976).Scorekeepingin aLanguageGame,Journalof PhilosophicalLanguage,vol 8.,D.ReidelPublishingCompany.

van der Sandt,Rob (1992). PresuppositionProjectionas AnaphoraResolution,Journalof Semantics,Vol 9, pp 333-377.

Traum,David (1994).A ComputationalTheoryof Groundingin NaturalLanguageConversation,Ph.D. Thesis,ComputerScienceDept., U. Rochester, December,1994

Walker, Marilyn A. (1996). TheEffect of ResourceLimits andTaskComplexityon Collaborative Planningin Dialogue,Artificial IntelligenceJournal,85:1-2.

Webber, BonnieLynn(1991).StructureandOstensionin theInterpretationof Dis-courseDeixis. LanguageandCognitive Processes.6:107-135

Zeevat, Henk (forthcoming).ExplainingPresuppositionTriggers. In InformationSharing: GivennessandNovelty in InformationProcessing, KeesvanDeemter&RodgerKibble, (Eds.)

174

Contrast and ContrastiveTopic

CARLA UMBACH

Universityof Leipzig

[email protected]

ABSTRACT. Thestartingpoint of this paperis theobservation that in a question-answerdialogthe useof but insteadof and is obligatory if the answeris overinformative in that it includesanadditionaltopic. A focus-semanticanalysisof but is presentedshowing that (a) but is focus-sensitive and (b) but requiresa denialwith respectto the appropriatequaestio. This analysisprovidesa uniform basisfor explaining the differentuses,e.g. semanticopposition,denial-of-expectation,andthetopicchangeuseof but. Beyondthatit givessomeinsightinto theinteractionbetweeninformationstructureanddiscourserelationsin constructingthediscourse.

1 Intr oduction

Considerthequestion-answerdialogsin (1)-(3). Dueto thecontrastive accentsinthe topic the answersin eachof (1)-(3) have to compriseat leasttwo conjuncts,otherwiseAdamwould beinclinedto askfor a continuation:”And/ but what...?”In (1) Adam asksaboutall of the children, and Ben addressesone part of thechildrenin thefirst conjunctandtheotherpart in secondconjunct.In (2), thoughAdamasksaboutthesmall childrenonly, Ben first refersto thebiggerones,andAdamhasto wait for thesecondconjunctto gettherequiredinformation.In (3) itis theotherwayaround:Adam’squestionis alreadyansweredby thefirst conjunctandthesecondconjunctgivesinformationAdamdid notaskfor. Anyway, in eachof theexamplesin (1)-(3)Adam’s questionis completelyansweredin theend.

(1) a. Adam:Whatdid thechildrendo today?

b. Ben:The small childrenstayedat HOME and/but the bigger oneswent totheZOO.1

(2) a. Adam:Whatdid thesmallchildrendo today?

b. Ben:Thebigger childrenwent to the ZOO, but/*andthesmall onesstayedat HOME.

1Boldfacetypedenotesa contrastive topicaccentandCAPS denotea focusaccent.

175



b. Ben:Thesmall childrenstayedat HOME, but/*andthebigger oneswenttotheZOO.

It is commonlyassumedthat in a coherentquestion-answerdialogtheanswerhasto refer to thesubjectmatterof thequestiononly. In (2) and(3) informationaboutanadditionaltopic is providedwithout renderingtheanswersunacceptable.But comparing(1) and (2)/(3) we observe that in the latter casethe useof butinsteadof and is obligatory. The useof but to indicatea topic changehasbeenmentionedin theliterature.Yet thereis no explanationwhy a contrastcanbeusedthatway: Whydoestheuseof but insteadof andrenderanover-informativeansweracceptable?

To addressthis question,first, I will briefly considerthenotionof contrastivetopic. ThenI will presenttheoutlinesof a focus-semanticanalysisof but inferringthe different usesof but from a uniform semanticbasis. (For a comprehensivediscussionseeUmbachin prep.). It will turn out thatby usingbut insteadof andthespeakerpresentstheadditionaltopicasbeingcloselyrelatedto theoriginalone,thusminimizing thedeviation.

2 ContrastiveTopic

Following e.g. Eckard(1996)andVallduvi/Villkuna (1998),I assumetwo kindsof foci, sentencefocusandcontrastive focus. Sentencefocus is expressedby asentencedefault accentandpartitionsthe sentenceinto a topic anda comment.Contrastive focusis dueto, e.g.,focus-sensitive operatorsandwh-questions.Forcontrastive focus,I adoptthe ideaof Alternative Semantics(cf. Rooth1992):2 Acontrastive focustriggersthe presuppositionthat thereexists at leastoneproperalternative, i.e. an elementdiffering from the ordinarymeaningof the focussedphrasewith respectto the accenteditem. For example,the contrastive focus inthe small children triggers the presuppositionthat there exist other (groupsofthe aforementioned)children in addition to the small children. Following thepresupposition-as-anaphor theory(vanderSandt1992)theproperalternative trig-geredby thecontrastive focusis regardedasananaphorwhichhasto bebound(orbeaccommodated).

A contrastive focusmayoccurin any positionin asentence.If, however, it oc-cursin thetopic part,it representsacontrastive topic andwill typically bemarked

2I will not opt for a particularframework herebecauseit’ s not relevant for the point I want tomake in this paper. In fact, the semanticsof but proposedin the next sectionwill needaccesstobothfocusandbackgroundof theconjunctsthusrequiringa fairly expressive framework, cf. Krifka(1999b).

176

with a rising accent.3 Being in the topic position,a contrastive topic refersto anentity thespeaker wantsto talk about.Moreover, dueto its contrastiveness,it indi-catesthatthereexist alternativesthespeakerwantsto talk about,too. This intuitionis,e.g.,capturedby thepartial-answeraccountsuggestedin Krifka (1999):Assum-ing that a sentenceis an answerto somequestion,the role of a contrastive topicconsistsin indicatingthat the answeris a partial one. (Roughly, a sentenceis apartialcongruentanswerto aquestionif it is entailedby somepropositionp in thequestionmeaningQ, but it is not a (complete)congruentanswerentailingsomepin Q.) In (1b), for example,the answergiven in the first conjunctis partial withrespectto thequestionin (1a)sinceit is entailedby theentireanswer.

Therearetwo notoriousproblemswith contrastive topics:First, in a sequenceof answersthelastanswercompletestherequestedinformation,so,intuitively, it isnotpartialany longer. Second,in theadditional-topicanswersin (2b)and(3b)oneof theconjunctsis acompleteanswerandtheotheroneis notevencongruentwithrespectto thequestion.In Krifka (1999),thefirst problemis handledby requiringeachanswerin asequenceto bepartialin isolation.Thesecondproblem,however,is notdiscussed.4

I will suggesta solution for the additional-topicproblemdistinguishingbe-tweenthe overt questionposedin a dialog andthe implicit ”quaestio”addressedin the answer. The quaestioof an utteranceis supposedto be a questionwhichis a posteriorireconstructedfrom the utterance.In the unmarked casein (1), forexample,thequaestioreconstructedfrom Ben’s answeris equalto Adam’s ques-tion. But in (2) and(3) thequaestioof theansweris slightly differentfrom theovertquestion.Ignoringtheconnective for themomentthequaestiorelatedto (2b)couldbe ”Whatdid thesmall children do, and whatdid thebigger onesdo?”. Beingaposteriorireconstructedthequaestioshowswhichquestionis actuallyansweredbyan utterance,even if thereis a deviation from the questionthat hasbeenposed.5

Dependingon the specificinterestthe quaestiomay be reconstructedeitherasaconstituentquestionor asapolarity question.

Adopting the notion of the quaestio,the role of the contrastive topic canbedefinedasindicatingthat theansweris partialwith respectto thequaestiorecon-structedfrom the entireconjunction. Congruencethenhasto refer the quaestio,too: An answeris congruentif the respective quaestioentailsthe question.Thisaccountsfor theacceptabilityof thedialogsin (2) and(3). But wehave to becare-ful not to throw thebabyout with thebathwater: Bringing in anadditionaltopicobviously requiressomeextraeffort, for example,usingbut insteadof and. Sothe

3In termsof Steedman(2000)a contrastive topic is calledtheme-focus,anda contrastive focusin thecommentpart is calledrheme-focus.Steedman’s notionof focusmatcheswith thenotionofcontrastive focusemployedhere.

4Theopen-question/strategy accountof Buring (1998)handlesbothproblems.However, it failsto cover ”crossed”contrastive topics,ase.g.in (12a).

5It hasbeensuggestedto view thequaestioasthe”questionunderdiscussion”QUD (cf. Ginzburg1996).But thenwe would needa moreliberal protocolfor querying.Accordingto Ginzburg, whenAdam posesa question,Ben caneitheracceptit asthe top mostQUD or reject it. In (2) and(3),however, Benjustslightly deviatesfrom Adam’squestionneitherrejectingit nor (fully) acceptingit.

177

questionis why theuseof but facilitatesa topic change.

3 Standard accountsof but

Thesemanticsandpragmaticsof but hasbeenthetopicof a lively discussionstart-ing with Lakoff ’sseminalpaperin 1971.Lakoff distinguishedbetweentwo usesofbut, semanticopposition(John is tall, but Bill is short) anddenial-of-expectation(John is tall, but he’s no goodat basketball). Sincethentherehasbeena hostofinvestigationspointingout furtherusesof but, e.g. for topic-change,andgeneral-izing theanalysisto othercontrastive connectives.Few approaches,however, havetried to examinetheunderlyingnotionof contrastandtracethevarioususesof butto auniform semanticmeaning.

Recentaccountsof themeaningbut mainlydraw ondefault knowledge.Asher(1993),for example,assumesbut to belicencedby differentpolaritiesin thecon-juncts wherethe polaritiesaredue to linguistic and commonworld knowledge.Winter andRimon (1994)usea default implication interpretedin possibleworldsemanticsto capturethenotionof contrast.Gaerdenfors(1994)presentsa seman-tics of but within his generalframework of reasoningwith expectations.Commonto theseanalysesis the ideathat the useof but indicatesa denial-of-expectation,theexpectationbeingdueto default world knowledge.

On theotherhand,it is easyto show thatcommonworld knowledgecannotbedecisive for theuseof but. Suppose,for example,youarenotversedin botany andyoudon’t know whatloosestrifeis. Nevertheless,youwill interpret(4) asdenyingtheexpectationthatloosestrifeis foundin July. This expectation,however, cannotbelongto your commonworld knowledge,simply becauseyou cannothave anyknowledgeaboutanentityor kind youaren’t acquaintedwith.6

(4) It wasJulybut wecouldn’t find any loosestrife

Theexamplein (5) refersto thefilm ”The EnglishPatient”.7 Thesituationisthis: Lord Almasy hasan affair with Katherine.Katherine’s husbandJeffrey hasto pick up Lord Almasy by planefrom somewherein the desert. Katherinewillbeon theplane,too. Jeffrey, knowing abouttheaffair, decidesto crashtheplaneon thegroundandkill themall. (5a)-(5d)tell theoutcomeof his plan,describingexactly thesamesituation.Nevertheless,in respondingto differentquestions,thesentencesdiffer with respectto thecontrastthey involve.

(5) a. (Whathappened?)Jeffrey is dead,Katherineis seriouslyinjured,andAlmasyis unhurt.

b. (Did Jeffrey succeedin killing themall?)Jeffrey is dead,but Almasyis unhurtandKatherineis alive, too.

6More thanyou ever wantedto know aboutloosestrife,thanksto Kathryn Bock: loosestrife=Lysimachia;invasive perennial,canin somevarietiesdisplacenative plants.

7adoptedfrom Brauße(1999).

178

c. (Have all of theparticipantsbeenaffectedby theaccident?)Jeffrey isdeadandKatherineisseriouslyinjured,but Almasyis unhurt.

d. (Do all of theparticipantsneedadoctor?)Jeffrey isdeadandAlmasyis unhurt,but Katherineisseriouslyinjured.

The”loosestrife”exampledemonstratesthattheexpectationdeniedby theuseof but neednotbegivenby commonworldknowledgeandis thereforenotaprereq-uisitefor theinterpretationof thesentence.Instead,it is triggeredby theinterpreta-tion of but, comparableto apresuppositionor (conversational)implicature.Takingtheexpectationasapresuppositionwouldallow for accommodation,whichwouldworkfinefor the”loosestrife”example:Accommodatethat,normally, loosestrifeisfoundin July. But whatto accommodatein the”EnglishPatient”examples?Sincethesituationis thesamein eachof (5b)-(5d),world knowledgecannottriggerdif-ferentexpectations.Theexpectationsseemto beevoked,instead,by thequestions.But do we really want to accommodate,e.g. in thecaseof (5c), thatnormally, ifJeffrey/someonesucceedsin killing himself,thenhesucceedsin killing theothers,too?This is clearlyabsurd.Theexpectationsinducedby thequestionsin (5b)-(5d)areby far tooadhocto becapturedby wayof accomodation.

Both examplesmake it plain thata contrastive relationis neithergivenby themeaningof theconjunctsnor inducedby commonworld knowledge.Theexpecta-tion deniedby theuseof but is obviously dueto a questionexplicitly or implicitlyposedby thepreceedingdiscourse.Sotheuseof but primarily hasto complywitha questionposedby thepreceedingdiscourse.Insteadof readilyaccommodatingadhocexpectationswe will investigatethe role of thesequestionsandtry to findouthow they relateto thebut-sentencesandwhy they reflectanexpectation.

4 Two novel observations

Theanalysisof but proposedin thispapertakesits startingpoint from two charac-teristicswhich have up to now beenneglectedin consideringthemeaningof but:First, but is focus-sensitive. This is evidentwhenyou compare(6a) and(6b). In(6a)theverbphraseis focussedwhereasin (6b) thesubjectis focussed.Dueto thefocusweexpectdifferentcontrasts:In (6a)washingthedisheshasto becontrastedwith someotheractivity. In (6b) Bill hasto becontrastedwith a differentperson.This suggeststhatwe shouldexaminethealternativesinducedby thefocussedex-pressionsandtake therespective setsof alternativesinto account.

(6) a. ... but Bill haswashedtheDISHES.

b. ... but BILL haswashedthedishes.

Thesecondobservationrelatesto thequestionsansweredby abut-conjunction.If the questionin (7) is answeredby confirming both conjuncts,the useof but

179

insteadof and is unacceptable,cf. (8a),(8b). If theanswerdeniesbothconjunctsbut is equally unacceptable,cf. (8c). If, however, one part of the questionisconfirmedandtheotherpartdenied,theuseof but is perfect(andtheuseof and isat leastmarked),cf. (8d)-(8f). Denial,by theway, doesnot hingeon thepresenceof anexplicit negation,cf. (8e). So,obviously, if a but-sentenceis anappropriateanswerto aquestioncomprisingtwo conjuncts,oneof themwill beconfirmedandtheotheronewill bedenied.

(7) Adam:Did Johnclearup his roomandwashthedishes?

(8) Ben:

a. [yes]Johnclearedup his roomand[yes] hewashedthedishes.

b. # [yes] Johncleareduphis room,but [yes] hewashedthedishes.

c. # [no] Johndidn’t clearuphis room,but [no] hedidn’t washthedishes.

d. [yes] Johncleareduphis room,but [no] hedidn’t washthedishes.

e. [yes] Johncleareduphis room,but [no] heskippedthewashing-up.

f. [no] Johndidn’t clearup his room,but [yes] hedid thewashing-up.

For thequaestioof abut-sentenceto reflecttheconfirm+deny characteristicsithasto comprisepolarity questionconjunctsinsteadof constituentquestions.Thepolarity questionconjunctswill relateto thealternativescontrastedby but askingwhetherboth alternatives apply simultaneously. Hence,in accordancewith theconfirm+deny characteristicsoneof theconjunctsof thequaestiowill beconfirmedby thecorrespondingbut-sentenceandtheotheronewill bedenied.

5 The focus-semanticanalysisof but

Thefocus-semanticanalysisof but makesusebothof its focus-sensitivity anditsconfirm+deny characteristics.The basicideais asfollows: In a but-conjunctiontherearetwo correspondingfoci (in the first andin the secondconjunct,respec-tively) which establishalternativeswith respectto eachother.8 Thesemanticsofbut, beyondbeinga mereconjunction,requiresthatoneof thealternativesrendersa truepropositionandtheotheroneis deniedwith respectto thefirst alternative’sbackground.In short:but excludesanalternative. This,by theway, doesn’t meanthatbut introducesa negation(but is not a ”nand”!). Instead,but requiresa nega-tion, in thesameway a verbselectsan argumentof a certaintype. If thereis noovert negationin oneof theconjuncts,thentheheareris requestedto reconstructit. Thefactthatthequaestiohasto beansweredby ”Yes,...but, no, ...” reflectstherequirednegation.

To show thatthis ideaappliesto but-conjunctionsin generalwe have to distin-guishfour cases:9 Eitherthesubjectof theconjunctsis thesameandthepredicates

8Thusthey arecontrastive foci.9I assumethatcommasmaybesubstitutedby full stops.Due to limitationsof spacecorrection

casesareleft out.

180

differ from eachother(A), or the predicatesarethe sameandthe subjectsdifferfrom eachother (B), or both subjectsand predicatesare different yet compara-ble (C), or subjectsandpredicatesarenot comparableto eachother, i.e. we haveto comparethe entirepropositions(D). To simplify matterslet us assumethat ifthereis an overt negationit occursin the secondconjunct(i.e. take only yes-nosequencesinto account)andconsiderbriefly thefour cases.

The A-caseis illustratedin (9a) and (9b). Normally, the predicateswill befocussed.10 In (9a)thenegationis overt inducingthequaestioin (9c). In (9b) thereis no overt negation. Thequestionin (9d), however, would not betheappropriatequaestiobecause(9b) cannotbeananswerto (9d). Instead,thequaestiohasto bethesameasin thenegatedexampleindicatingthat thereis animplicit negationtobereconstructedfrom thecomplementof thepredicate(i.e. skipthewashingup issupposedto bethesetcomplementof washthedishes). Sotherelevantalternativesin (9a)and(9b) arethesame:clear up theroomasagainstwashthedishes.

(9) a. [yes] Johncleareduphis ROOM, but [no] hedidn’t washtheDISHES.

b. [yes] Johncleareduphis ROOM, but [no] heskippedtheWASHING-UP.

c. Did Johnbothclearuphis roomandwashthedishes?

d. Did Johnbothclearuphis roomandskip thewashing-up?

TheB-caseis given in (10a): Thesubjectsarefocussedandestablishalterna-tiveswith respectto eachother. The quaestiois given in (10b). Contraryto theA-caseexamplesexplicit negationis obligatory, cf. (10c). The reasonfor this iseasyto see:Individuals,asopposedto predicates,don’t have complements(thereis no ”non-John”). However, B-caseexampleswill beacceptablewithout explicitnegation if the particle too is added,cf. (11a). Note that, whenaddingtoo, thequaestiowill beadifferentone,cf. (11b).Similarly, B-caseexampleswith anega-tion in both conjunctswill be acceptableif theparticleeither is added(cf. (11c)andthequaestioin (11d).

(10) a. [yes]JOHN clearedup his room,but [no] BILL didn’t.

b. Did bothJohnandBill clearup their rooms?

c. *Johnclearedup his room,but Bill did.

(11) a. Johnclearedup his room,but Bill did, too.

b. Did Johnclearup his room,andwashetheonly onewhodid?

c. Johndidn’t clearedup his room,but Bill didn’t, either.

d. Did Johnleave his roomin amess,andwashetheonly onewho did?

TheC-caseis morecomplex becausewe have to considertwo foci in eachofthe conjuncts,oneof thembeinga contrastive topic. The contrastive topicsmay

10Theremaybeanadditionalfocuson thenegation,thenthepredicatein secondconjuncthastoberegardedasacontrastive topic. In German,in thiscase,thewordorderwill bereversed:JohnhatAUFGERAUMT, aberabgewaschen hater NICHT.

181

be”parallel” comprisingeitherthesubjectsor thepredicates,or they are”crossed”comprisingthe subjectof the first conjunctandthe predicateof the secondcon-junct, or vice versa.(12a)presentsoneof thecrossedvariants.The it-cleft para-phrasein (12b)makesit clearthatthefocus(in thecommentpartof thesentence)is on Bill , anddoing the dishesrepresentsa contrastive topic.11 For this reasonin Germantheword orderis reversed,cf. (12c). In (12a),asin all of theC-caseexamples,thereis no explicit negation in eitherof the conjuncts. Nevertheless,thereis a denial,because(12a)clearly entailsthat Johndid not washthe dishes.Takingtheentailmentinto account,thequaestioin (12d)is againpartly confirmedandpartly denied,asdemonstratedin (12e). Note,that thequaestiois exactly thesameastheonein (9c). In fact,(12a)and(9a)areverymuchalike bothconveyingthe informationthat Johnclearedup his room anddid not do the dishes. But in(12a)we additionallylearnwhofinally did thewashingup.

(12) a. John clearedup his ROOM, but BILL did thedishes.

b. John clearedup his ROOM, but it wasBILL whodid thedishes.

c. John hatAUFGERAUMT, aberabgewaschenhatBILL.

d. Did Johnbothclearuphis roomandwashthedishes?

e. [yes] Johnclearedup his room,but [no, Johndid not do the dishes]thedisheswerewashedby Bill.

Let us skip over the secondcrossedvariant and look at one of the parallelvariants. In (13a)thecontrastive topicsareparallelbeingthe subjectsin both oftheconjuncts.Thequaestiois givenin (13b). Thedenialof partof thequaestioisentailedby telling whatBill did insteadof clearingup theroom,cf. (13c).

(13) a. John clearedup his ROOM, but Bill did theDISHES.

b. Did bothJohnandBill clearup their room?

c. [yes] Johnclearedup his room, but [no, Bill did not clearup his room]Bill did thedishes.

The last of the four casesconcernsbut-sentenceswith wide foci in the con-juncts,cf.(14a).In thesecasestheentirepropositionshave to beregardedasbeingalternativeswith respectto eachother. If thereis noexplicit negationin oneof theconjuncts,e.g.(14b),it hasto bereconstructed.Notethattheappropriatequaestiofor (14b)hasto be(14c)insteadof (14d).

(14) a. � It is raining� F , but �we arenotgoingto stayat home� F.

b. � It is raining� F , but �we aregoingto go for awalk� F.

c. Is it raining,andarewegoingto stayathome?

d. Is it raining,andarewegoingto go for awalk?11In Englishtheit-cleft variantis clearlypreferredover (12a).In Germanthereversedwordorder

variantin (12c)will bethepreferredone.

182

To sum up thesefindings:12 First, in the secondconjunctof a but-sentencethereis a focusassociatedwith but (Fbut). It is eithera contrastive topic, if thereis one,or a regular focus. The focusassociatedwith but representstheexpectedalternative (EA), thatis, thealternative deniedwith respectto thequaestio.13 Sec-ond,in thefirst conjuncttherehasto beacorrespondingfocus(Fcorr) thatcontainstheexpectedalternative in its setof alternatives.Third, thereis adenial condition,statingthat thepropositionresultingfrom substitutingtheexpectedalternative forthecorrespondingfocusis false,i.e � [...Fcorr/EA...]C1.14 Supposethemeaningofthefirst conjunctis givenby [...Fcorr ...]C1, andthemeaningof thesecondconjunctis given by [...Fbut ...]C2. Thenthe meaningof a but-conjunction”C1 but C2” isgivenby:

� ...Fcorr...� C1 � � ...Fbut...� C2 � �� ...Fcorr/EA...� C1

Thecrucialpoint in thesemanticsof but, which distinguishesbut from a mereconjunction,is the denialcondition. In the A-case,with an overt negationin thesecondconjunct,thedenialconditionis trivially satiesfied,becauseit is given bythesecondconjunct. For example,in (9a), repeatedin (15), the focusassociatedwith but is washthe dishes, andthis is alsothe expectedalternative. The corre-spondingfocusis clear up the room. So thedenialconditionis ”It’ s not thecasethatJohnwashedthedishes”,which is equivalentto thesecondconjunct.

(15) (= 9a)John[cleareduptheROOM]Fcorr, buthedidn’t [washtheDISHES]Fbut.

If an A-caseexampleoccurswithout overt negation the negation hasto bereconstructedusingthepredicate’s complement,cf. (9b). This time, theexpectedalternative is givenby thepredicate’s complementinsteadof beingdirectly givenby Fbut. Thedenialconditionwill thenbeentailedby themeaningof thesecondconjunct.

With B-caseexamplesandD-caseexamplessatisfactionof thedenialconditionis similarly trivial. In the C-caseexamples,however, the denialcondition is notgiven by oneof theconjuncts.For example,in (12b), repeatedin (16), the focusassociatedwith but is a contrastive topic, i.e.washthe dishes, which is also theexpectedalternative. Thecorrespondingfocusis clear up theroommatchingwiththe type of the expectedalternative. Hencethe denial condition is the sameasabove: ”It’ s not the casethat Johnwashedthe dishes”. But this time, the denialcondition is an entailmentresultingfrom the additional information concerningwho/whatinsteadof theexpectedalternative satisfiestheproposition.

12As saidin thebeginning,thepresentationin this paperis restrictedto confirm+deny sequences.To includedeny+confirmsequencesthedefinitionshave to beextended.

13In caseslikes(9b) and(14b)wherethenegationhasto bereconstructedby meansof thecom-plementof thefocussedexpression,theexpectedalternative is givenby thecomplement.

14Dots indicatethe part of the conjunctwhich is not subsumedby Fcorr or Fbut, i.e. either thebackgroundor anotherfocus.Fcorr/EA meansthesubstitutionof EA for Fcorr.

183

(16) (= 12b)John [cleareduphis ROOM]Fcorr, but it wasBILL who[washedthedishes]Fbut.

Sinceit is thedenialconditionwhich distinguishesbut from a mereconjunc-tion, expectationsdeniedby a but-sentenceareobviously inducedby the denialcondition.This is in accordancewith a well-known propertyof negatedsentencesin general:Negatedsentencescommonlytrigger the implicaturethat the speakerexpects(or assumesthehearerto expect)thattheaffirmativepropositionholds(cf.Givon 1978). This is the reasonwhy we reconstructthequaestioaskingwhetherbothof thealternativeshold simultaneously, e.g.,whetherJohndid both,clearuphis room and also washthe dishes. Note that, due to this quaestiothere is anexpectationthat both of the alternativesdo hold simultaneously:If Johnclearedup his room,hewill have washedthedishes,too. So,finally, the focus-semanticanalysisconfirmsthe idea that thereis an expectationdeniedby the useof but.However, contraryto what is saidin the literature,theexpectationis not givenbycommonworld knowledge.Instead,it is triggeredby thedenialconditioninducingthe specialform of the quaestio.The quaestio,of course,hasto be in line withtheprevious context. But that is a generalproblemof discourseconstruction,notrestrictedto theuseof but.

6 ”concessivebut”?

In somecontexts a concessive marker apparentlycanbeaddedor evenbesubsti-tutedfor but without affectingthemeaningof thesentence.Fromthat it hasbeenconcludedthat thereis a concessive useof but (e.g. Groteet al. 1997). How-ever, regardingbut asbeing interchangeablewith a concessionin thesecontextspresupposesthata concessionis interpretedasindicatinga denialof expectation.As opposedto that,Konig (1991)convincingly arguesthata concessionexpresses”incausality”, thusaccountingfor thecloserelationsshipbetweencausalandcon-cessive statements.Following Konig’s incausalityanalysisit is easyto show thata concessionis not a specialcaseof a contrast:First, accordingto the incausalityinterpretation(17a)hasto beparaphrasedas(17b). Second,it is well-known thatand-conjunctionsmay be interpretedin many differentways,e.g. astemporalorascausalrelations,cf. (18a),(18b).But this is anoverinterpretationby thehearer,not includedin the meaningof and (cf. Posner1980). Third, dueto the focus-semanticanalysisinterpreting(19a)requiresto reconstructa negation(analogousto (14b)). Now, (19b)maybeoverinterpretedin a causalway, too, cf. (20). Dueto thenegationcausaloverinterpretationresultsin incausality, compare(17b)and(20).

(17) (incausalityanalysisof concessives)

a. Althoughit is rainingMary is happy.

184

b. It it raining, andit is not the casethat Mary is not happy becauseofthat.

(18) (causaloverinterpretationof and)

a. It is raining,andMary is happy

b. It is raining,andMary is happy becauseof that.

(19) (reconstructionof thenegation)

a. It is raining,but Mary is happy.

b. It is raining,andit is not thecasethatMary is nothappy.

(20) (causaloverinterpretationof but)It is rainingandit is not thecasethatMary is nothappy becauseof that.

Hence,thereis no ”concessive but”, just asthereis no ”causaland”— inter-pretingbut asa concessionis dueto overinterpretation.At thesametime, a con-trastis perfectlycompatiblewith aconcession,justasacausalrelationis perfectlycompatiblewith aconjunction.Thissuggeststhatcontrastandconjunction,on theonehand,andconcessionandcausality, on the otherhand,aredifferenttypesofdiscourserelationsexploiting differentfeaturesof thediscourse.Concessionandcausalityrepresentrelationsbetweenpropositions,or statesof affairs, i.e. seman-tic/externalrelations(cf. Mann,Thompson1988). Thusthey mayberealizedbyan adverbial containinga propositionalanaphor(becauseof that, in spiteof thatetc.)andestablishananaphoriclink (cf. Webberetal. 1999).

A contrast,on the other hand, is basedon the information structureof thesentencecombiningsubsequentfoci. Making useof the informationstructure,acontrastis a genuinestructuralrelation,i.e. no semantic/external relation. Yet itshouldnot be subsumedunderthenotion of pragmatic/internalrelationsbecausethe latter areusuallytied to the illocutionaryaspectsof their arguments(Sanderset al. 1992). The relation of contrastgiven by but seemsto resistthe standardclassification,which may help to clarify the ongoingdiscussionaboutthe typesof discourserelations(cf. e.g. Moore andPollack1992,BatemanandRondhuis1997)

7 Topic change

Letusfinally comebackto thedialogsin (1)-(3),repeatedin (21)-(23).Thestartingpoint of this paperwasthequestionwhy in (22b)and(23b) theuseof but insteadof and is obligatory. Considerthe quaestionesgiven below.15 In the unmarkedcasein (21b)bothandandbut areacceptablebecauseBenmayintendhis answeraseitherreferringto thequestion(21c)or to theonein (21d). Note,however, thatthereis a crucial difference: The but-quaestiobut not the and-quaestiotriggers

15For easeof comparisonthequaestiois givenhereconsistingof apolarityandaconstituentques-tion conjunct,assumingthatansweringaconstituentquestionsimultaneouslyconfirmstherespectivepolarityquestion,sothis form of quaestiois alsoansweredby a confirm+deny sequence.

185

theexpectationthat thebiggerchildrendid thesamething asthesmall onesdid.By usingbut in (22b)and(23b)Bendeliberatelyconveys this expectation.In thisway, althoughactuallydeviating from theoriginal topic of Adam’s question,Benpresentstheadditionaltopic asbeingcloselyrelatedto theoriginal one. Thus,byusingbut Bensuggeststhat theadditionaltopic is relevant, too, andthedeviationis reasonable.

(21) a. Adam:Whatdid thechildrendo today?

b. Ben:The small childrenstayedat HOME and/but the bigger oneswent totheZOO.

c. Ben’s quaestiowhenusingand:Whatdid thesmallchildrendo andwhatdid thebiggeronesdo?

d. Ben’s quaestiowhenusingbut:Whatdid thesmallchildrendo,anddid thebiggeronesdo thesame?


b. Ben:The bigger children went to the ZOO, but the small onesstayedatHOME.

c. Ben’s quaestio:Whatdid thebiggerchildrendo,anddid thesmallonesdo thesame?


b. Ben:The small childrenstayedat HOME, but the bigger oneswent to theZOO.

c. Ben’s quaestio:Whatdid thesmallchildrendo,anddid thebiggeronesdo thesame?

To conclude,the dialogsin (22) and(23) clearly demonstratethat an answerneednot refer to only the topic of the question.This suggeststhat a naturallan-guagedialog shouldnot be conceived asa server-client relationwhereB hastoanswerall andonly A’s questions.Partnersin a dialogseemto be”peer-to-peer”:They areentitledto introduceanadditionaltopic, but they areboundto relatetheadditionaltopic to theoriginal one,thusminimizing thedeviation. Oneway to dothis is by usingtheconjunctionbut.

Bibliography

[1] Asher, N. Referenceto Abstract Objectsin Discourse. Kluver, 1993.

186

[2] Bateman,J., Rondhuis,K. CoherenceRelations:Towardsa General Specifi-cation. DiscourseProcesses24,1997.

[3] Brauße,U Was ist Adversativitat?ABER oder UND. DeutscheSprache2,1999.

[4] Buring, D. Focusand Topic in a Complex Model of Discourse. In: Geurts,Krifka, van der Sandt(1999)Focusandpresuppositionin a multi-speakerdiscourse.ESSLI-99WorkshopReader, 1999.

[5] EckardtR. IntonationundPredication. ArbeitspapieredesSFB 340.Univer-sitat Dusseldorf,1996.

[6] Foolen, A. Polyfunctionalityand the semanticsof adversativeconjunctions.Multilingua 10-1,1991.

[7] Gaerdenfors,P. The role of expectationsin reasoning. In: Masuch, M.,Plos,L. (eds)KnowledgeRepresentationandReasoningUnderUncertainty.Springer, Berlin/New York, 1994.

[8] Ginzburg, J.Interrogatives:Questions,FactsandDialogue. In Lappin,S.(ed.)TheHandbookof ContemporarySemanticTheory, Blackwell,Oxford,365-422,1996.

[9] Givon,T. Negationin Language: Pragmatics,Function,Ontology. In Cole,P.(ed)SyntaxandSemantics:Pragmatics.Vol. 9, AcademicPress,New York,1978.

[10] Grote,B., Lenke, N., Stede,M. Ma(r)king Concessionsin EnglishandGer-man. DiscourseProcesses24,87-117,1997.

[11] Konig,E.Concessiverelationsasthedualof causalrelations. In Zaefferer, D.(Ed.) Semanticuniversalsanduniversalsemantics.Groningen-Amsterdamstudiesin semantics12,Foris,1991.

[12] Krifka, M. Additiveparticlesunderstress. Proceedingsof SALT 8. Cornell,CLC Publications,1999.

[13] Krifka, M. Frameworksfor associationwith focus. Workshop”IntonationandInterpretation”,ESSLLI 99,Utrecht,1999.

[14] Lakoff, R. If ’s, and’s andbut’s aboutconjunction. In Fillmore, Langendoen(eds.):Studiesin linguistic semantics.deGruyter, New York, 1971.

[15] Mann,W., Thompson,S. Rhetoricalstructure theory: Towardsa functionaltheoryof text organization. TEXT 8 (2), 1988.

[16] Moore,J.D.,Pollack,M.E. A Problemfor RST: Theneedfor multi-level dis-courseanalysis. ComputationalLinguistics,18(4),1992.

[17] Posner, R. Semanticsand pragmaticsof sentenceconnectives. In Searle,Kiefer, Bierwisch (eds.)Speechact theory and pragmatics.Reidel, Dor-drecht,1980.

187

[18] Rooth,M. A Theoryof FocusInterpretation. NaturalLanguageSemantics1,1992.

[19] Sanders,T., Spooren,W., Noordman,L. Toward a Taxonomyof CoherenceRelations. DiscourseProcesses15,1992.

[20] vanderSandt,R. PresuppositionProjectionasAnaphora Resolution. Journalof Semantics9, 1992.

[21] Steedman,M. Information Structure and the Syntax-Phonology Interface.Linguistic Inquiry 31,4, 649-689,2000.

[22] Vallduvı, E., Vilkuna, M On Rhemeand Kontrast. Syntaxand Semantics,Vol.29.AcademicPress,1998.

[23] Webber, B., Knott, A., Joshi,A. Multiple discourse connectivesin a lexi-calizedgrammarfor discourse. Third InternationalWorkshopon Computa-tional Semantics,Tilburg, 1999.

[24] Winter, Y., Rimon,M. ContrastandImplication in Natural Language. Jour-nal of Semantics11,1994.

188

Intonational Phrasingand DiscourseSegmentation

KLAUS VON HEUSINGER

Universitat Konstanz

[email protected]

ABSTRACT.

Theoriesthatrelateintonationalstructureanddiscoursestructureconcentrateon intonationalandinformationalunits thateithercorrespondto a clause(proposition)or to a singlereferent.How-ever, intermediatephrasesof the intonationalphrasingoftensegmentunits thataresmallerthana clause,but thatdo not introducea referent.Eventhoughthey areacknowledged(cf. Selkirk’s”senseunits”), therearenoapproachesthataccountfor their role in building thediscoursestruc-ture. The paperpresentsa new approachthatdescribesthe representationof suchintermediatephrasesand accountsfor their functions in the constructionof a more fine-graineddiscoursestructure.

1 Intr oduction

Intonationcontoursarerepresentedby phonologistsasasequenceof abstracttonesconsistingof pitch accentsand boundarytones. Pierrehumbertand Hirschberg(1990,308)assigndiscoursefunctionsto theparticulartones:“Pitch accentscon-vey information about the statusof discoursereferents(...). Phraseaccents[=boundarytonesof intermediatephrases]convey informationabouttherelatednessof intermediatephrases(...). Boundarytonesconvey informationaboutthedirec-tionality of interpretationfor the currentintonationalphrase(...).” The statusofdiscoursereferentscanbe accountedfor in termsof given vs. new; the bound-ary tonesof intonationalphrasesindicatehow the propositionexpressedby thewholephraseis integratedinto thediscourse.Similarly, boundarytonesof inter-mediate(or phonological)phrasesthat correspondto a full propositionindicatetheway thesepropositionsareinterpretedwith respectto thelinguistic context, asillustratedin (1) and(2). However, in thisview thereis noaccountof treatinginter-mediatephrasesthatcorrespondto unitsbelow theclauselevel andabove thelevelof simplediscoursereferents,suchasthemodification”im achtzehntenJahrhun-dert” (”in theeighteenthcentury”)or theunsaturatedphrase”lebte in Frankreich”(”li vedin France”)in example(3):

189


(1) (Georgeatechickensoup) (andgot sick)L LL%

(2) (Georgeatechickensoup) (andgot sick)H LL%

(3) � (Im achtzehntenJahrhundert)� (lebtein Frankreich)] � [(ein Mann,)( � derzu dengenialstenundabscheulichstenGestaltendieserangenialenundabscheulichenGestaltennichtarmenEpochegehorte.)]

“In theeighteencenturyFrancetherelivedamanwhowasoneof themostgiftedandabominablepersonagesin anerathatknew no lackof giftedandabominablepersonages”

In orderto accountfor theseintermediateunitsandtheir functionsin building adiscoursestructure,I modify thegeneralassumptionsof discoursetheories,namelythat the discoursestructureis exclusively realizedbetweenpropositions(Hobbs1990,Roberts1996,Buring2000amongothers),andtheview of DRT (KampandReyle 1993)thatdiscoursestructureonly relatesdiscoursereferentsto eachother.I extendAsher’s (1993,1999)segmentedDRT (SDRT) by assumingthatwe canalsocreatediscourserelationsbetweensub-clausalelementssuchasmodificationsor unsaturatedclauses.The remainderof the paperis organizedas follows: Insection2, I discussdifferentapproachesto thefunctionof intonationalcontours.Ishow thatnoattentionis givento intermediatephrasesandtheirdiscoursefunction.In section3, I presentonefamily of approachesto discoursestructure:DRT, a the-ory that focuseson anaphoricrelationin termsof accessibilityrelationsbetweendiscoursereferents. In section4, I introduceAsher’s segmentedDRT, a modifi-cationof classicalDRT thatallows to expressrelationsbetweenpropositionsin adiscourse.In section5, I modify Asher’s discoursemodelin orderto implementeven smallerunits thoseunits that correspondto the intermediatephrasesof theintonationalstructure.

2 The meaningof intonational structur e

Pierrehumbert(1980) representsan intonationalcontourby a tuneconsistingofabstracttones(4), whicharegeneratedby afinite stategrammar(Figure18.1)thatcombinesthetoneslistedinto legal tunes:

(4) Phonological tones(Pierrehumbert1980)

a. Eachphraserequiresat leastonepitch accent(for English:H*, L*, orbitonalasH*+L, H+L*, L*+H, L+H*, andH*+H)

b. Eachphrasereceivesa phraseaccent(H � , L � ) at theendof thewordthatis associatedwith thelastpitch accent

c. Eachphraseendswith aboundarytone(H%, L%)

190

Figure18.1:Thefinite stategrammarof Pierrehumbert(1980)

This grammaraccountsfor the legal tuneandthereforedescribesthepossiblecompositionof tonesto tunes. However, Pierrehumbert(1980)doesnot accountfor therelationsbetweentheelementsin thetune. Therehave beendifferentsug-gestionshow to interpretthe relationbetweenthe elementsof the tuneandtheirfunctions: (i) Themostclassicaldescriptionis in termsof phonologicalwords,atraditiongoingbackto Liberman’s view of thecontouras’ideophonicword’. (ii)Steedman(1991,2000)describesthepitchaccentasfunctionsthatrequireanargu-mentin orderto yield a tune.His approachreflectsthereconstructionof linguisticconfigurationsin categorial grammar. (iii) PierrehumbertandHirschberg (1990)proposethat thedifferenttoneshave independentfunctionsrangingover differentdomains.

2.1 The tune describing an ideophonicword

HayesandLahiri (1991)executetheprogramof theintonationallexicon proposedby Liberman(1975)in themostcomprehensiveway. They accountfor thedifferenttunesin Bengali and decomposethe tunesinto intonational stems, suffixesandprefixes(they useI for intonationalphraseandP for phonologicalor intermediatephrase(ip)):

(5) Theintonationallexicon of BengaliHayesandLahiri (1991)

191

(a)Accents L* questionaccent(“stems”) H* declarative accent

L*H P focusaccent(b) BoundaryTones LI neutral

(“suffixes”) LI HI continuationriseHI offeringHI LI yes/no

(c) Prefix L+ finality marker(form L+H* whenattachedto H*)

Thesetonescombineto thefollowing tunes,assumingthattherearethreepitchaccents,andoptionalTP andoneobligatoryTP (HayesandLahiri 1991,72). NotethatafocusedphraseismarkedbyaL*-pitch accentandanobligatoryHp boundarytone(of anintermediatephrase)

(6) Possibletunesin BengaliL*H I OfferingL*H ILI Yes/noquestionL*H PHI FocusL*H PLILH I Focuswith continuationriseH*L I DeclarativeH*L IHI Declarative with continuationriseL+H*L I DownstepL+H*L IHI Downstepwith continuationrise

Thisview is influencedby thephonologicaltraditionof analyzingunitslike thesyllableor thephonologicalword, which consistof oneprominentpartandothermoreor lessimportantparts.Here,thepitchaccentconstitutesthecentralpartandtheboundarytonesadditionalaffixes.

2.2 Tune representingthe functional contribution to the utterance

Steedman(1991,2000)executesHalliday’sthematicstructurein termsof combina-tory categorialgrammar(CCG).Thiscanbeillustratedwith thefollowing examplewhich receivesthethematicstructurein theme-rheme. Both thematicunitsarefur-therdividedinto givenmaterialandnew material;the latter is associatedwith thepitchaccent.

192

(7) Q: I know thatMary’sFIRSTdegreeis in PHYSICS.But whatis thesubjectof herDOCTORATE?

A: (Mary’s DOCTORATE) (is in CHEMISTRY)L+H*LH% H* LL%

Given New Given Newtheme rheme

Thebasicunitsarethethemeandtheutterance.All otherpartsaredefinedwithrespectto thesebasicelements.For example,therhemeis a functionthattakesthethemeas argumentyielding the utterance(this is of course,the instantiationofthesubject-predicatestructurein termsof functionalapplication).Steedmannowdefinesthefunctionof thepitchaccentL+H* asthemethatmissesaboundarytone,i.e. asa function that needsa boundarytoneto yield a theme. Analogously, thepitch accentH* indicatesa function that needsa boundarytonein orderto yielda rheme.Thusin thedescriptionof tones,Steedmanassumestheboundarytonesandthewholetuneastheprimaryunits.

(8) Categorial functionsof tones(Steedman1991)LH% boundarytone simpleargumentLL% boundarytone simpleargumentL+H* pitch accent functionfrom boundarytoneinto themeH* pitch accent functionfrom boundarytonesinto rhemeL+H*LH% contour simpleargument:themeH* LL% contour functionfrom themesinto utterance

Steedmanusesthetermsthemeandrhemeaswell asgivenandnew. Thefirstpaircanbedefinedwith respectto thesentenceunderanalysis.Yet thesecondpaircanonly bedefinedby thediscoursein which thesentenceis embedded.

2.3 Tonesrepresentingdifferent discoursefunctions

PierrehumbertandHirschberg (1990)give a list of functionsof pitch accentsandboundarytones.Thelatterindicatewhetherthephraseto which theboundarytoneis associatedshouldbeinterpretedwith respectto theprecedingdiscourseor to thefollowing discourse.PierrehumbertandHirschberg (1990,304)illustratethispointon thefollowing contrastbetween(9) and(10). Thelow boundarytoneL% in (9)indicatesthat this sentenceis relatedto the discourseby its own, while the highboundarytoneH% in (10a)indicatesthat it is to beinterpretedwith respectto thefollowing sentence.This differenceinfluencesthechoiceof theantecedentof thepronounit: In (9) it refersto thefollowing propositionI spenttwo hours figuringouthowto usethejack, while in (10) it refersbackto thenew car manual(seealsoexamples(1) and(1) above):

(9) a. My new carmanualis almostunreadableLL%

193

b. It’s quiteannoyingLH%

c. I spenttwo hoursfiguring outhow to usethejackLL%

(10) a. My new carmanualis almostunreadableLH%

b. It’s quiteannoyingLL%

c. I spenttwo hoursfiguring outhow to usethejackLL%

PierrehumbertandHirschberg (1990,308)assignthefollowing discoursefunc-tionsto theparticulartones:

Pitch accentsconvey informationaboutthe statusof discourserefer-ents,modifiers,predicates,andrelationshipsspecifiedby accentedlex-ical items. Phraseaccentsconvey information aboutthe relatednessof intermediatephrasesinparticular, whether(thepropositionalcontentof) oneintermediatephraseis to form partof a largerinterpretative unitwith another. Boundarytonesconvey informationaboutthedirection-ality of interpretationfor the currentintonationalphrasewhetherit is”forward-looking”or not.

In explaining the functionof intermediateboundariestones(their “phraseac-cents”),they referto the“propositionalcontent”of thecorrespondingphrase.How-ever, notall intermediatephrasesexpressapropositionalcontent,somemightonlyrefer to modificationssuchas “im achtzehntenJahrhundert”(“in the eighteenthcentury”) or the unsaturatedphrase“lebte in Frankreich”(“li ved in France”)ofexample(3), repeatedas(11).



(11) is thefirst sentencein PatrickSuskind’snovel “DasParfum”. (“Perfume”)(thedataarefrom BraunschweilerandFitzpatrickandLahiri 1998ff). Intermediatephrasesin intonationalstructurenotalwayscorrespondto propositionsor to simplediscoursereferents. Therefore,we needa morefine-graineddiscoursestructurethatallows to constructcorrespondingdiscoursesegments.

194

3 Discoursestructur e

Thereare two main familiesof approachesto discoursestructure. According toonefamily, discoursestructureis understoodasrealizingrelationsbetweenpropo-sitions.Herethestructureis representedasatreeof propositions(e.g.Hobbs1990,Roberts1996,Buring2000)asillustratedin Figure18.2.In theotherview thedis-courseis incrementally(re)constructed,asillustratedin Figure18.3.Thisprogramis executedin DRT (KampandReyle 1993),whichwill bepresentedbelow.

Figure18.2:Discoursestructureasa tree

Figure18.3: Incrementaldiscoursestructure

The initial problemthatmotivateddiscourserepresentationtheoriesis the in-terpretationof nominalandtemporalanaphorain discourse.Thephenomenonofcross-sententialanaphoraforcesa semanticsto extendits limits from thesentenceto thediscourse.Thekey ideain theapproachto semanticsof discourse,exempli-fiedin (Heim1982)and(Kamp1981),is eachnew sentenceor phraseis interpretedasanadditionor ’update’of thecontext in which it is used.This updateoftenin-volvesconnectionsbetweenelementsfrom the sentenceor phrasewith elementsfrom the context. Informally described,a sequenceof sentencesS1, S2, S3, S4

is interpretedby incrementallyconstructinga discourserepresentationasin Fig-ure18.3above. Anaphoricrelationsanddefiniteexpressionsarecapturedby linksbetweenobjectsin this representation.In orderto derive thetruth conditionof thesentence,therepresentationis embeddedinto a model. TheDRSin (12b)graphi-cally describesa discourserepresentationstructure(in shortDRS)with two parts.

195

Onepartis calledtheuniverseof theDRS,theotherits conditionset.A DRSis anorderedpair consistingof its universeandconditionset, written as � UK, ConK � .TheDRSin (12b)hasasits universeonediscoursereferentx andasits conditionasetof propertiesthatareascribedto thediscoursereferentsin theuniverse.In (12b)thepropertyof beingamanandof walking is ascribedto thediscoursereferentx.

(12) a. A manwalks

b. � x � man(x)andwalk(x)�The sequenceor conjunctionof two sentencesas in (13a), receives a DRS

incrementally. We start with the alreadyestablishedDRS for the first conjunctin (13b), thena new discoursereferentfor thepronounheanda conditionfor thepredicatewhistleisaddedin (13c).Theanaphoriclink of thepronounisgraphicallyrepresentedasy=?, indicatingthat thepronounis still unresolved. Thediscoursereferentwhich standsfor ananaphoricexpressionmustbeidentifiedwith anotheraccessiblediscoursereferentin theuniverse,herethey is identifiedwith thex, asin (13d).

(13) a. A manwalks.He whistles

b. � x � man(x)& walk(x)�c. � x,y � man(x)& walk(x) & y=? whistle(y)�d. � x,y � man(x)& walk(x) & y=x whistle(y)�

Thenew discoursereferentintroducedby thepronounmustbelinkedor iden-tified with an alreadyestablishedandaccessiblediscoursereferent.DRT definesaccessibilityin termsof structuralrelations,i.e., thediscoursereferentmustbeinthe same(or a higher) universe. With this conceptof accessibility, the contrastbetween(14) and(15) canbe describedby the differencein the setof discoursereferentsthatareaccessiblefor thediscoursereferentu of thepronounit. Thecon-structionrule for thenegationin 15 createsanembeddeddiscourseuniversewiththediscoursereferenty andtheconditionsdonkey(y) andx ownsy. Theanaphoricpronounit in thesecondsentencecannotfind a suitablediscoursereferentsinceithasno accessto the embeddeddiscourseuniversewith the only fitting discoursereferenty.

(14) a. Pedroownsadonkey. He beatsit.

b. � x,y,z,u � Pedro(x)& donkey(y) & x ownsy z=x & u=y & z beatsu �(15) a. Johndoesnotown adonkey. *He beatsit.

b. � x,z,u � John(x)& � y � & x ownsy � z=x & u=?& z beatsu�4 SegmentedDRT

Asher(1993;1999)developstheextension“segmentedDRT” (=SDRT) thatis notconfinedto the incrementalcompositionof DRSs,but alsocapturesdiscoursere-lationsbetweenthe sentencesin the discourse.He revisesthe classicalDRT of

196

Kamp(1981)andKampandReyle (1993).Sincethemeaningof eachsentenceisconstruedasa function from truth conditionsto truth conditions,the truth condi-tionalcontentof thewholediscourseis reconstructedby thesequentialapplicationof thesefunctions.Asher(1993,256)notesthat “the notionof semanticupdatingin theoriginal DRT fragmentof Kamp(1981)(...) is extremelysimple,exceptfortheproceduresfor resolvingpronounsandtemporalelements,which theoriginaltheorydid not spell out. To build a DRS for the discourseasa whole and thusto determineits truth conditions,onesimply addsthe DRS constructedfor eachconstituentsentenceto what onealreadyhad. (...) This procedureis hopelesslyinadequate,if onewantsto build a theoryof discoursestructureanddiscourseseg-mentation.” In SDRT, eachsentenceSi is first representedasa particularSDRSfor that sentence.The SDRScanthaninteractwith the alreadyestablishedDRSreconstructinga discourserelationsR, suchascausation,explanation,coherence,elaboration,continuation,asinformally sketchedin Figure18.4.Only in a secondsteptherepresentationis integratedinto thealreadyestablishedrepresentation.

Figure18.4:Constructionof a segmentedDRSin SDRT

To summarizethis very shortpresentationof DRT: ThediscoursestructureofDRT providesnot only a new structurebut alsointroducesnew semanticobjects,thediscoursereferents,theconditionsandthediscoursedomains(”boxes”). DRTexplainssemanticcategoriessuchasdefinitenessandanaphorain termsof inter-actionbetweentheserepresentations.Furthermore,theextensionto SDRT allowsexpressionsof discourserelationsbetweenwholepropositionsaswell.

5 Phrasingand Segmentation

Thebasicassumptionof this paperis that intonationalstructurereflectsdiscoursestructure.Thuseachelementof theintonationalstructuremustbeassigneda func-tion in constructingthediscourse.Pitchaccentsintroduceor referto discourseref-erents,boundarytonesof intonationalphrasesrelatethecontentof thesephrases,aproposition,to otherpropositions.However, therehasbeennofunctionfor phrasesthat do not expressa propositions.The investigationof HayesandLahiri (1991)have shown that suchboundarydo function in discoursestructure,e.g. as rightedgeof a focusedphrase(cf. (5) and( 6)). However, thereis no generalapproachto theseboundarytonesin termsof discoursesemantics.HereI sketchanew analy-

197

sisof intermediateintonationalphrasesandthecorrespondingdiscoursesegments:I introducea mappingrelationfrom intermediateintonationalphrasingto the

string(surfacestructure)of thesentenceunderdiscussion.For example,theshortpausesindicatedby “ � ”, and the long ones(“ � ”) in (3), repeatedas (16), yieldthesegmentationin Figure18.5. Second,eachsegmentreceivesa discourserep-resentation.Third, wecanestablishrelationsbetweenthediscourserepresentationof the sub-clausalunits andthe of the establisheddiscoursestructure,as in Fig-ure18.6. I assumethe following two relationfor modifications:(i) modifying analreadyexisting discoursereferent;(ii) settingthestagefor a discoursereferenttobeintroduced.Modificationalrelationallowsthattheinformationis integratedintothealreadyestablisheddiscourse.Fourth,afterhaving establishedtheserelations,therestof thesentencecanbeanalyzed,andfinally therepresentationof thewholesentenceis integratedinto thediscoursestructure,asin Figure18.7:



Figure18.5:DiscourseSegmentation

Figure18.6:DiscourseRelation“settingthestage”

I have proposedan extensionof Asher’s SDRT with smallerdiscourserepre-sentationsandnew relationsbetweensub-clausaldiscourserepresentations.Thisallows to assigndiscoursefunctionsto intonationalphrasing,includingboundarytonesfor intermediatephrases.

198

Figure18.7:Adding therepresentationfor thewholesentence

Bibliography

Asher, N. (1993). Referenceto Abstract Objectsin Discourse. Kluwer Aca-demic,Dordrecht.

Asher, N. (1999).Discourseandfocus/backgrounddistinction.In P. BoschandR. van der Sandt(Eds.),Focus.Linguistic, Cognitive, and ComputationalPerspectives., pp.247–267.Cambridge:CambridgeUniversityPress.

Braunschweiler, N., J. Fitzpatrick,andA. Lahiri (1998ff.). The konstanzinto-nationdatabase:German,swissgerman,americanenglish,british english,eastbengali,westbengali.Universityof Konstanz.

Buring,D. (2000).Focusandtopic in acomplex modelof discourse,ms.ucsc.

Hayes,B. andA. Lahiri (1991).Bengaliintonationalphonology. Natural Lan-guage andLinguisticTheory(9), 47–96.

Heim, I. (1982).TheSemanticsof Definite and IndefiniteNoun Phrases. Ph.D. thesis,University of Massachusetts.Amherst.Ann Arbor: UniversityMicrofilms.

Hobbs,J.(1990).Thepierrehumbert-hirschberg theoryof intonationalmeaningmadesimple.commentson pierrehumbertandhirschbert.In P. R. Cohen,J.Morgan,andM. E.Pollack(Eds.),Intentionsin Communication., pp.313–323.Cambridge/Mass.:MIT Press.

Kamp, H. (1981).A theory of truth and semanticinterpretation.In J. Groe-nendijk,T. Janssen,andM. Stokhof(Eds.),Formal Methodsin theStudyofLanguage., pp.277–322.Amsterdam:AmsterdamCenter.

Kamp,H. andU. Reyle (1993).FromDiscourseto Logic. Introductionto Mod-eltheoretic Semanticsof Natural Language, Formal Logic and DiscourseRepresentationTheory. Dordrecht:Kluwer.

Liberman,M. (1975).TheIntonationalSystemof English.Ph.D. thesis,MIT.Cambridge/Mass.

Pierrehumbert,J. (1980).ThePhonology andPhoneticsof EnglishIntonation.Ph.D. thesis,MIT. Cambridge/Mass.

199

Pierrehumbert,J. andJ. Hirschberg (1990).The meaningof intonationalcon-tours in the interpretationof discourse.In P. R. Cohen,J. Morgan, andM. E. Pollack (Eds.), Intentionsin Communication., pp. 271–311.Cam-bridge/Mass.:MIT Press.

Roberts,C. (1996). Information structurein discourse.towardsan integratedformal theoryof pragmatics.In J.-H.YoonandA. Kathol (Eds.),OhioStateUniversity [=OSU] WorkingPapers in Linguistics, Volume49,pp.91–136.

Steedman,M. (1991).Structureandintonation.Language 67, 260–296.

Steedman,M. (2000).TheSyntacticProcess.Cambridge:MIT Press.

200

information structure, discourse -

Documents