Clarifying Intentions in Dialogue_A Corpus Study

《Clarifying Intentions in Dialogue_A Corpus Study》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

ClarifyingIntentionsinDialogue:ACorpusStudy∗JulianJ.SchloderandRaquelFern¨andez´InstituteforLogic,LanguageandComputationUniversityofAmsterdamjulian.schloeder@gmail.com,raquel.fernandez@uva.nlAbstractAspartofourongoingworkongroundingindialogue,wepresentacorpus-basedinvestigationofintention-levelclarificationrequests.Weproposetorefineexistingtheoriesofgroundingbycon-sideringtwodistincttypesofintention-relatedconversationalproblems:intentionrecognitionandintentionadoption.Thisdistinctionisbacked-upbyanannotationexperimentconductedonacorpusassembledwithanovelmethodforautomaticallyretrievingpotentialrequestsforclarification.1IntroductionDialogueiscommonlymodelledasajointactivitywheretheinterlocutorsarenotmerelymakingin-dividualmoves,butactivelycollaborate.Acentralcoordinationdeviceisthecommongroundofthedialogueparticipants,theinformationtheymutuallytakeforgranted(Stalnaker,1978).Thiscommongroundischangedandexpandedoverthecourseofaconversationinaprocesscalledgrounding(Clark,1996).Weareinterestedinthemechanismsusedtoestablishagreement,i.e.,intheconversationalmeanstoestablishabeliefasjoint.Toinvestigatethisissue,inthispaperweexaminecaseswheregrounding(partially)fails,asindicatedbythepresenceofclarificationsrequests(CRs).Incontrasttopreviouswork(i.a.,Gabsdil,2003;Purver,2004;Rodr´ıguezandSchlangen,2004),whichhasmostlyfocusedonCRstriggeredbyacousticandsemanticunderstandingproblems,weareparticularlyconcernedwithproblemsrelatedtointentionrecognition(goingbeyondsemanticinterpretation)andintentionadoption(i.e.,mutualagreement).Thefollowingexamples,fromtheAMIMeetingCorpus(Carletta,2007),arecasesinpoint:(1)A:Ithinkthat’sall.(2)A:Justuhdothatquickly.(3)A:I’dsaytwo.B:Meeting’sover?B:Howdoyoudoit?B:Why?Intheseexamples,itcannotbesaidthatBhasfullygroundedA’sproposal,butalsonotthatBrejectsA’sutterance.Rather,Basksaquestionthatisconducivetothegroundingprocess.In(1),BhasapparentlyunderstoodA’sutterance,butisunsureastowhetherA’sintentionwastoconcludethesession.WethereforeconsiderCRslikeB’squestionin(1)asrelatedtointentionrecognition.Incontrast,in(2)and(3),Bdisplaysunwillingnessorinability(butnooutrightrefusal)togroundA’sproposal,andrequestsfurtherinformationsheneedstoestablishcommonground,i.e.,toadoptA’sintentionasjoint.RequestsforinstructionshavealsobeenrelatedtoclarificationinBenotti’s(2009)workonmultiagentplanning.Inthispaper,wepresentacorpus-basedinvestigationofintention-levelclarification,partofanongo-ingprojectthataimstoanalysethegroundingprocessbeyondsemanticinterpretation.Inthenextsection,weintroducesometheoreticalobservationsandrefineexistingtheoriesofgrounding(Clark,1996;All-wood,1995)bydistinguishingbetweenintentionrecognitionandintentionadoption.Wethenpresentasystematicheuristictoretrievepotentialclarificationrequestsfromdialoguecorporaanddiscusstheresultsofasmall-scaleannotationexperiment.1Weendwithpointersforfuturework.∗TheresearchpresentedinthispaperhasbeenfundedbytheEuropeanCommunity’sSeventhFrameworkProgramme(FP7/2007-2013)undergrantagreementno.567652ESSENCE:EvolutionofSharedSemanticsinComputationalEnvironments(http://www.essence-network.com/).1Wewillmakeourannotateddatafreelyavailable.46Proceedingsofthe11thInternationalConferenceonComputationalSemantics,pages46–51,London,UK,April15-172015.c2015AssociationforComputationalLinguistics LevelJointActionExampleClarification1contactAandBpayattentiontoeachotherAreyoutalkingtome?2perceptionAproducesasignalandBperceivesitWhatdidyousay?3understandingAconveysameaningandBrecognisesitWhatdidyoumean?4.1intentionrecognitionAintendsaprojectandBunderstandsitWhatdoyouwant?uptake4.2intentionadoptionAproposesaprojectandBacceptsitWhyshouldwedothis?Table1:GroundinghierarchyforspeakerAandaddresseeBwithrefineduptakelevel.2TheoreticalObservationsAsextensivelydiscussedbyHulstijnandMaudet(2006),theintentionallevelweareinterestediniscommonlydenotedwiththetermuptake.Inparticular,inClark’s(1996)stratificationofthegroundingprocessintofourdistinctlevels(seeTable1forourtakeonit),thefourthlevel,“proposalandconsid-eration(uptake),”isrelatedtothespeaker’sintentions.Whendiscussingjointprojectsatlevel4,Clarkintroducesthenotionofjointconstruals:thedeterminationandconsiderationofspeakermeaning,in-cludingtheintendedillocutionaryforce(Clark,1996,pp.212–213).However,healsopointsoutthatuptakemayfailduetounwillingnessorinability:“whenrespondentsareunwillingorunabletocomplywiththeprojectasproposed,theycandeclinetotakeitup”(Clark,1996,p.204).Wecontendthatthisdifferencebetweenconstrualandcompliance—betweenintentionrecognitionandintentionadoption—hasbeenobscuredintheliteraturesofar.2Forexample,intheirannotationschemeforCRs,Rodr´ıguezandSchlangen(2004)reproducetheunderspecificationinlabellingtheirlevel4CRsas“recognisingorevaluatingspeakerintention.”Sincewe,withClark(1996),considersuchintentionalcategoriestobepartofthegroundinghier-archy,weexpectproblemsonanintentionalleveltobeevincedinmuchthesamewayasotherconver-sationalmishaps:inparticularbyCRsaimedatfixingthesedifferenttypesofconversationaltrouble.WhenstudyingtheCRsannotatedasintentionrelatedinthecorpusofRodr´ıguezandSchlangen(2004)weindeedfindexamplesrelatedtorecognitionandotherswhichaimatadoption:3(4)K:okay,againfromthetop(5)K:formethatisinfactbelowthisI:fromtheverytop?I:whybelow?K:no,well,[...]K:yes,itbelongsthere,allokay.In(4),speakerIhasevidentlynotfullyunderstoodwhatK’squestionis,despitehavingsuccessfullyparsedandunderstoodthepropositionalcontentofK’sutterance.Ontheotherhand,Idisplaysnosuchproblemin(5),butrathersomereluctancetoadoptK’sassertionascommonground.Weconsider(4)tobeaclarificationquestionrelatedtointentionrecognitionwhereastheonein(5)relatestointentionadop-tion.AparticularlystrikingclassofintentionrecognitionCRsarespeechactdeterminationquestionsasinthefollowingexample:4(6)A:Andwe’regoingtodiscuss[...]who’sgonnadowhatandjustclarifyB:AreyouaskingmewhetherIwannabeinthere?Ourhypothesisisthattheclassesofclarificationrequestsrelatedtointentionrecognitionandintentionadoption,respectively,aredistinctanddiscernible.Inparticular,weproposetoimproveuponClark’s(1996)hierarchybysplittinghisuptake-levelintotwo,separatingrecognitionfromadoption.Table1showsouramendedhierarchyandconstructedexamplesforclarificationrequestsevincingfailureatacertainlevel.Totestthishypothesis,wehavesurveyedexistingcorporaofCRsandassembledanovelcorpusofintention-relatedCRstocheckifannotatorscouldreasonablydiscernthetwoclasses.2WhileDIT++(Bunt,2012)stratifiesthegroundinghierarchyinto“attention/perception/interpretation/evaluation/execution,”itissimilarlyunderspecified:Tous,evaluation(e.g.,checkinganassertedpropositionforconsistency)relatestointentionadoption,whereas(semantic)understandingand(pragmatic)intentionretrieval(e.g.,recognisingonlevel4.1thatanindicativewasintendedasaninformactandhencerequiresaconsistencycheckonlevel4.2)areagaindistinctcategories.3Wethanktheauthorsforprovidinguswiththeirannotatedcorpus;inthedialogues,IisexplainingtoKhowtoassembleapaperairplane.WehadtheGerman-languageexamplestranslatedtoEnglishbyanativespeakerofGerman.4RetrievedfromtheBritishNationalCorpus(BNC)(Burnard,2000)usingSCoRE(Purver,2001).47 3CorpusStudy3.1PreviousStudiesOurworkbuildsonpreviouscorpusstudiesofCRs(Purveretal.,2003;Rodr´ıguezandSchlangen,2004;RieserandMoore,2005).However,existentstudiesarenotperfectlysuitedforinvestigatingground-ingatthelevelofintentions.5Firstly,theannotationschemeofPurveretal.(2003;2004),whichtheauthorsapplytoasectionoftheBNC(Burnard,2000),makesuseofsemanticcategoriesthatcannoteasilybemappedtotheintention-leveldistinctionsintroducedintheprevioussection.Secondly,whiletheschemesemployedbyRodr´ıguezandSchlangen(2004)andRieserandMoore(2005)(bothbasedonSchlangen,2004)doincludeacategoryforintention-levelCRs,thecorporatheyannotate—theBiele-feldCorpusandtheCarnegieMellonCommunicatorCorpus,respectively—arehighlytask-orientedandhencetheintentionsoftheinterlocutorsaretoalargedegreepresupposed:theparticipantsintendtofulfilthetask.Finally,inallcases,thefocusoftheauthorsdidnotliewithintentionalclarificationandthereforetheymighthaveleftoutquestionsintheirannotationsthatareinterestingtous,inparticularmorecomplexintentionadoptionCRs(whichmaynothavebeenconsideredCRstobeginwith,giventhelackofwellestablishedtheoreticaldistinctionsdiscussedintheprevioussection).Forourstudy,wehavechosentoextractquestionsfromtheAMIMeetingCorpus(Carletta,2007),acollectionofdialoguesamongstfourparticipantsrole-playingadesignteamforaTVremotecontrol.Thedialoguesarelooselytask-andgoal-oriented,buttheconversationismostlyunconstrained.Duetothissetting,weexpectalargeramountofdiscussionanddecisionmaking,whichshouldgiverisetomoreintention-levelCRs.Inaddition,therichannotationsdistributedwiththeAMICorpusenabledustoapplyasophisticatedheuristictoautomaticallyextractpotentialCRs,whichwedescribenext.3.2DataTheAMICorpusisannotatedwithdialogueacts,includingaclassof‘Elicit-*’actsdenotingdifferentkindsofinformationrequests/questions,butwithoutspecificallydistinguishingCRs.However,thecor-pusisalsoannotatedwithrelationsbetweenutterances,looselycalledadjacencypairannotation,6whichindicateswhetherornotanutteranceisconsideredadirectreplytoanotherone.WeutiliseobservationsonthesequentialnatureofCRs(“other-initiatedrepair”)ingroupsettingsmadebySchegloff(2000)toassembleasetofpossibleclarificationrequestsasfollows.TakeallutterancesQwhere:a.Qisturn-initialandannotatedasan‘Elicit-’typeofdialogueact,spokenbyaspeakerB.b.Qisthesecondpartofanadjacencypair;thefirstpart(thesource)isspokenbyanotherspeakerA.c.Qisthefirstpartofanotheradjacencypair;thesecondpart(theanswer)isspokenbyAaswell.ThisheuristicisbasedontheintuitionthatCRsareproperquestions(i.e.,utterancesthatdemandananswer)withabackward-lookingfunction(i.e.,relatedtoanearliersourceutterance)thataretypicallyansweredbythespeakerofthesource.Weexpectthisheuristictohaveasufficientlyhighrecalltobequantitativelyapplicable,butareawarethatitcannotfindeachandeveryCR.7Thereare338utterancesQintheAMICorpussatisfyingthecriteriaabove.Wenotethatthean-notationmanualfortheAMICorpusstatesthatCRsareusuallyannotatedas‘Elicit-’acts,butthatsomeverysimpleCRs(e.g.,‘huh?’)caninsteadbetaggedas‘Comment-about-Understanding(und).’However,thisclassalsocontainssomebackchannelutterances:positivecommentsaboutunderstanding.Ifweapplythesameheuristictotheutterancesannotatedas‘und,’wefind195additionalpossibleCRs.WeconfirmedthatourheuristicsuccessfullyseparatesCRsfrombackchannels,andthattheseCRsare5WehavecarefullystudiedtheannotateddatadescribedinPurveretal.(2003)andRodr´ıguezandSchlangen(2004),whichwaskindlyprovidedtousbytheauthorsuponrequest.6Seehttp://mmm.idiap.ch/private/ami/annotation/dialogueactsmanual1.0.pdf.7Inparticular,previousworkindicatesthatsomeCRsaresimplynotanswered;Rodr´ıguezandSchlangen(2004)report8.7%unansweredCRsintheircorpus.Ourheuristicdoesnotfindthese.48 indeedrelatedtolevels1–3ofClark’s(1996)hierarchy.However,theseutterancesarenottheprimarysubjectofourstudy.WehenceforthrefertoCRsonlevels1–3collectivelyaslow-level.3.3AnnotationProcedureAsindicatedabove,weareprimarilyinterestedinthe338possibleCRsannotatedas‘Elicit-’dialogueactsandthereforeincludedonlytheseinourannotation.Sinceourmaininterestisinintention-levelCRsandourprimaryambitionistheinvestigationofintentionadoptionvs.intentionrecognition,weusedthefollowingsimpleannotationscheme:Eachquestionfoundbyourheuristicisannotatedasoneof{not,low,int-rec,int-ad,ambig},wherethecategoriesaredefinedasfollows.•notCR.Selectthiscategoryifyouaresurethatthequestionisnotaclarificationrequest.Thatis,ifitdoesnotservetobettertheaskersunderstandingoftheprevioushighlightedutterance.Forinstanceifthequestionisrequestingnovelinformation,movingthedialogueforward.•lowCR.Selectthiscategoryifthequestionindicatesthattheaskerhasnotfullyunderstoodthesemantic/propositionalcontentoftheprevioushighlightedutterance.Thisincludes,forexample,wordmeaningproblems,acousticproblems,orreferenceresolution.•intentionrecognitionCR.Selectthiscategoryifthequestionindicatessemanticunderstanding,butthattheCRuttererhasnotfullyunderstood(oristryingtoguess)thespeaker’sgoal/intention(theintendedfunctionoftheprevioushighlightedutterance).Theprototypicalcaseisspeechactdetermination.•intentionadoptionCR.SelectthiscategoryifthequestionindicatestheCRuttererhasunder-stood/recognisedthespeaker’smaingoal(theirintention),butdoesnotyetacceptitbecausehewants/needsmoreinformationorhehasincompatiblebeliefs.Forinstance,iftheCRuttererasksaboutthereasonbehindthespeaker’sutterancebeforeacceptingit,orrequestsinformationneededtocarryoutherproposal.•ambiguous.SometimesitmaynotbepossibletodecidewhatfunctionaCRhasprecisely,maybeduetoalackofcontext.Inthosecases,annotatethequestionasambiguous.WeinstructedourannotatorstofollowadecisiontreewheretheyfirstdecidewhetheraquestionisclearlynotaCR,andonlyotherwiseconsiderthedifferentcategoriesofCRs.Thisisbecauseinapilotstudywefoundthatthedistinctionbetween‘notCR’and‘intentionadoptionCR’wasdifficultforsomeannotators.Toreducetheconfusion,wedefinedthe‘notCR’classasonlyclear-cutcasesofnot-CRquestions,attheriskofincurringahigheramountofambiguitywhenthedecisiontreebottomsout,i.e.,whenaquestionthatwasnotdefinitelynotaCRcouldnotbematchedtoaCR-categoryafterall.Ourannotationschemeonlyrefinesonedimension(namely,‘source’)ofthemulti-dimensionalschemesappliedbyRodr´ıguezandSchlangen(2004)andRieserandMoore(2005).Sinceourmainambitioninthisworkistoestablishthetwolevelsofintentionality,weleaveafullerannotationwithfurtherdimensions—suchassyntacticcategorieslikeSchlangen’s(2004)‘form’—forfuturework.Nevertheless,thisisadifficultannotationtask:Annotatorscanonlyplaytheroleofoverhearerandthereforehaveamoreindirectaccesstotheintentionsoftheinterlocutors.Inaddition,CRsinparticularcanbefragmentedandambiguous.Therefore,annotatorswereshownasubstantialdialogueexcerptstarting10utterancesbeforethesourceandendingwitheitherthe10thutteranceaftertheanswertotheCRorwiththeCR-asker’snextreply(thefollow-up).Wefoundthatanswerandfollow-upareparticularlyhelpfulindeterminingthefunctionofaCR:theanswergiveshintstowardsthespeaker’sinterpretationoftheCR,andthefollow-upcanshowwhethertheaskeragreeswiththatconstrual.8Inthefullstudy,thecorpuswasannotatedby2expertannotators,sincewedeemedthetasktobetoocomplexandfine-grainedforna¨ıveannotators.Onethirdofthecorpuswasannotatedbybothannotators,theremainingtwothirdsbyoneannotatoreach.Tocreateagold-standardontheoverlappingsegment,theannotatorsdiscussedtheutteranceswheretheirinitialjudgementdifferedandmutuallyagreedontheappropriateannotation.8Rodr´ıguezandSchlangen(2004)includetheCRasker’s‘happiness’(asevincedbythefollow-up)intheirannotation.49 CategoryCountincluding‘und’ExamplenotCR90(27%)-A:‘YoucancallmePeter.’–B:‘Andyouare?Intheproject?’low-level78(23%)273(62%)A:‘Seventy-fivepercentofusersfinditugly.’–B:‘TheLCD?’intent.recognition53(16%)53(12%)A:‘Ithinkthat’sall.’–B:‘Meeting’sover?’intent.adoption77(23%)77(17%)A:‘That’saveryunnaturalmotion.’–B:‘Doyouthink?’ambiguous40(12%)40(9%)Total338(100%)443(100%)Table2:Distributionofclarificationrequestsinourcorpuswithexamplesforeachcategory.3.4ResultsInthefive-wayclassificationtaskdescribedabove,ourannotatorshadanagreement(Cohen’sκ,1960)ofκ=0.76ontheoverlappingthirdofthecorpus;9ofκ=0.85inthebooleantaskofdeterminingwhetheranutteranceisaCR;andofκ=0.82inthebooleantaskofretrievingintention-relatedCRsfromallotherquestions.ThedistributionofcategoriesisshowninTable2.Inordertocompareourdistributiontopreviouswork,wehavealsorecordedthedistributionweobtainwhendroppingtheitemsannotatedas‘notCR’andaddingthequestionsannotatedas‘Comment-about-Understanding(und)’aslow-levelCRs.ThenthetotalnumberofCRsinourcorpusis443.TheAMICorpuscontainsabout42,000turns,sowefoundthatroughly1.1%ofturnsreceiveclar-ificationaccordingtoourheuristic.Previousstudieshaveindicatedahighernumber:Purver(2004)reportsabout4%andRodr´ıguezandSchlangen(2004)about5.8%.Rodr´ıguezandSchlangen(2004)themselvesconjecturethattheircorpusmightcontainanunusuallyhighamountofCRsduetotheset-ting(aninstructorguidingabuilder).Forcomparison,wehavemanuallyextractedCRsfroma2500-turnsubsetoftheAMICorpus:Wefound52CRsinthatsegment,indicatingthatabout2%ofturnspromptaCR.ItistobeexpectedthatourheuristicmissessomeCRs,e.g.,onesthatdonotreceiveananswer,anditscoverageisdependentonhowsystematictheadjacencypairannotationintheAMICorpusis.Whileourheuristiconlyretrievesanestimated50%ofCRs,10thedistributionofclasseswefoundiscomparabletotheresultsdescribedbyRodr´ıguezandSchlangen(2004)andRieserandMoore(2005):Theyreport63.5%and75%,respectively,oflow-levelCRsand22.2%/20%onintention-level.Rodr´ıguezandSchlangen(2004)marktheremaining14.3%asambiguous,whereasRieserandMoore(2005)report5%“other/several”anddonotmentionanambiguityclass.11Byandlarge,thisiscomparabletothedis-tributionwefound.Wehavelowambiguity(9%)comparedtoRodr´ıguezandSchlangen(2004)becauseweconflateddifferentcategoriesoflower-levelCRsintoone‘lowCR’category.Aswehadhoped,wefindalargeramount(29%)ofintention-levelCRsthanthepreviousstudies.Wetakethesimilarityindistributionsastacitlyconfirmingtheviabilityofourheuristicforquantitativeevaluation.4ConclusionWehavetheoreticallymotivatedadistinctionwithingroundinghierarchiesbetweenintentionrecognitionandintentionadoptionandhavecreatedanovelcorpusofintention-levelCRstoinvestigateitstenability.Ourcorpusisnotonlynovelinitscontents,butalsoinitsconstruction:unlikepreviousstudies,wehavedevelopedandappliedasuitableheuristicthatexploitsrichexistingannotationstoautomaticallyfindpossibleclarificationrequests.Asmall-scaleannotationexperimentonourcorpusshowedthatthetheoreticaldistinctionweproposeisviable.Ourimmediatenextstepinthisprojectisadeeperinvestigationintotheformandproblemsourcesoftheintention-levelCRsinourcorpus,includingamorefine-grainedannotation.9Rodr´ıguezandSchlangen(2004)reportκ=0.7inthetaskofdeterminingthelevelofunderstandingthattheCRaddresses.However,theircategorisationisdifferentfromours.Inparticular,theydonotincludea‘notCR’category.10WesurveyedtheCRsnotfoundbyourheuristicandattributethismostlytotheadjacencypairannotation;however,inadditiontoCRsthatarenotansweredatall,therearealsoCRsthatareansweredbyadifferentpersonthanthesourcespeaker.11Theircategory“ambiguity”referstoaclassofCRsdubbed“ambiguityrefinement”andnottouncertaintyintheannotation.50 ReferencesAllwood,J.(1995).Anactivitybasedapproachtopragmatics.Gothenburgpapersintheoreticallinguis-tics(76),1–38.Benotti,L.(2009).Clarificationpotentialofinstructions.InProceedingsofthe10thAnnualMeetingoftheSpecialInterestGrouponDiscourseandDialogue(SIGdial).Bunt,H.(2012).Thesemanticsoffeedback.InProceedingsofthe16thSemDialWorkshopontheSemanticsandPragmaticsofDialogue(SeineDial).Burnard,L.(2000).ReferenceGuidefortheBritishNationalCorpus(WorldEdition).OxfordUniversityComputingServices.Carletta,J.(2007).Unleashingthekillercorpus:experiencesincreatingthemulti-everythingamimeet-ingcorpus.LanguageResourcesandEvaluation41(2),181–190.Clark,H.H.(1996).Usinglanguage.CambridgeUniversityPress.Cohen,J.(1960).Acoefficientofagreementfornominalscales.EducationalandPsychologicalMea-surement1(20),37–46.Gabsdil,M.(2003).Clarificationinspokendialoguesystems.InProceedingsofthe2003AAAISpringSymposium.WorkshoponNaturalLanguageGenerationinSpokenandWrittenDialogue,Stanford,CA,pp.28–35.Hulstijn,J.andN.Maudet(2006,June).Uptakeandjointaction.CognitiveSystemsResearch7(2-3),175–191.Purver,M.(2001,October).SCoRE:AtoolforsearchingtheBNC.TechnicalReportTR-01-07,De-partmentofComputerScience,King’sCollegeLondon.Purver,M.(2004).TheTheoryandUseofClarificationRequestsinDialogue.Ph.D.thesis,King’sCollege,UniversityofLondon.Purver,M.,J.Ginzburg,andP.Healey(2003).Onthemeansforclarificationindialogue.InCurrentandnewdirectionsindiscourseanddialogue,pp.235–255.Springer.Rieser,V.andJ.D.Moore(2005).Implicationsforgeneratingclarificationrequestsintask-orienteddi-alogues.InProceedingsofthe43rdAnnualMeetingoftheAssociationforComputationalLinguistics(ACL).Rodr´ıguez,K.J.andD.Schlangen(2004).Form,intonationandfunctionofclarificationrequestsinGer-mantask-orientedspokendialogues.InProceedingsofthe8thSemDialWorkshopontheSemanticsandPragmaticsofDialogue(Catalog).Schegloff,E.A.(2000).When‘others’initiaterepair.AppliedLinguistics21(2),205–243.Schlangen,D.(2004).Causesandstrategiesforrequestingclarificationindialogue.InProceedingsofthe5thSIGdialWorkshoponDiscourseandDialogue.Stalnaker,R.(1978).Assertion.InP.Cole(Ed.),Pragmatics,Volume9ofSyntaxandSemantics,pp.315–332.NewYorkAcademicPress.51