Computational Approach to Anaphora Resolution in S

JournalofArticialIntelligenceResearch15(2001)263-287Submitted3/01;published10/01ComputationalApp[e-mail]aticosUniversidaddeAlicanteAlicante,[e-mail]aticosUniversidaddeAlicanteAlicante,SPAINAbstractThispaperpresentsanalgorithmforidentifyingnoun-phraseantecedentsofpronounsandadjectivalanaphorsinSpanishdialogues.Webelievethatanaphoraresolutionrequiresnumeroussourcesofinformationinordertondthecorrectantecedentoftheanaphor.Thesesourcescanbeofdierentkinds,e.g.,linguisticinformation,discourse/dialoguestructureinformation,ortopicinformation.Forthisreason,ouralgorithmusesvariousdierentkindsofinformation(hybridinformation).Thealgorithmisbasedonlinguisticconstraintsandpreferencesandusesananaphoricaccessibilityspacewithinwhichtheal-gorithmndsthenounphrase.Wepresentsomeexperimentsrelatedtothisalgorithmandthisspaceusingacorpusof204dialogues.ThealgorithmisimplementedinProlog.Accordingtothisstudy,95.9%ofantecedentswerelocatedintheproposedspace,apreci-sionof81.3%wasobtainedforpronominalanaphoraresolution,and81.5%foradjectivalanaphora.1.IntroductionAnaphoraresolutionisoneofthemostactiveareasofresearchinNaturalLanguagePro-cessing(NLP).ThecomprehensionofanaphoraisanimportantprocessinanyNLPsystem,yetitisamongthetoughestproblemsincomputationallinguisticsandNLP.AccordingtoHirst(1981):Anaphora,indiscourse,isadeviceformakinganabbreviatedreference(con-tainingfewerbitsofdisambiguatinginformation,ratherthanbeinglexicallyorphoneticallyshorter)tosomeentity(orentities)intheexpectationthatthere-ceiverofthediscoursewillbeabletodisabbreviatethereferenceand,thereby,determinetheidentityoftheentity.Thereferencetoanentity(e.g.,apronoun)isgenerallycalledananaphor,theentitytowhichtheanaphorrefersisitsreferent,andthepreviousreferencetothesameentityistheanaphor’santecedent.Forinstance,inthestatement\Johniateanapple.Heiwashungry,thepronounheistheanaphorandthenounJohnistheantecedent.Ananaphoricproblemcanbedescribedaslyingsomewherebetweentheresolutionandthegenerationofanaphora,theformertermbeingthedisabbreviatingofthereferenceandc2001AIAccessFoundationandMorganKaufmannPublishers.Allrightsreserved.Palomar&Martnez-Barcothelatterbeingtheabbreviatingformofthereferencetoanentity.Thispaperfocusesexclusivelyontheresolutionofanaphoraandnotontheirgeneration.Anaphoracanbeclassiedinmanydierentways,dependingupontheparticularcriteriaonechoosestoemploy.Regardingtheelementthatcarriesoutthereference(theanaphor),forexam-ple,cleardistinctionsshouldbemadebetweenpronominalanaphora,adjectivalanaphora,denitedescriptions,one-anaphora,surface-countanaphora,verbal-phraseanaphora,andtimeand/orlocationreferences.Thispaperfocusesontheresolutionofpronominalandadjectivalanaphora.1Itiswidelyagreedthattheprocessofresolvinganaphorainnaturallanguagetextsmaybesupportedbyavarietyofstrategiesthatemploydierentkindsofknowledge.Bydierentkindsofknowledgewemeanthevarioussourcesofinformationusuallyemployedforanaphoraresolution,includingmorphologicalagreement,syntacticparallelism,semanticinformation,discoursestructure,topicalknowledge,andsoon.Naturallanguageprocessing(NLP),and,specically,anaphoraresolution,usesmanyresourcesandsourcesofinformationfortworeasons:(1)numerousresourcesareavailabletothescienticcommunity;and(2)humansemploymanysourcesofinformationinordertoresolvedierentlinguisticphenomena.Wepresentanalgorithmthatcoordinatesdierentformsofknowledgebydistinguishingbetweenlinguisticknowledge(constraintsandpreferences)anddialogue-structureknowl-edge(anaphoricaccessibilityspace).Thealgorithmidentiesthenounphrasetowhichathird-personpersonalordemonstrativepronounoradjectivalanaphor2refersinaSpan-ishdialogue.WecallthisalgorithmARDi(anaphoraresolutionindialogues).ARDiwasimplementedinProlog.InSection2below,wepresentrelatedworkonanaphoraresolutionindialogues.InSection3,wesuggestanannotationschemeforcapturingSpanishdialoguestructure.InSection4,anaccessibilityspacebasedonthisannotationschemeisdened.InSection5,wepresentthealgorithmARDi.Finally,anexperimentalstudyofthealgorithmispresentedinSection6.2.RelatedworkonanaphoraresolutionindialoguesForanaphoraresolutionindialogues,aproliferationofmethodsbasedondialoguestructure(discourse-orientedapproaches)havebeendeveloped.Amongthese,weshouldliketoespe-ciallyacknowledgetheworkofGrosz(1977,1981),inwhichtheinuenceofdialoguestruc-tureinanaphoraresolutionisjustied.Grosz’sworkfocusesspecicallyontask-orienteddialogues.Otherstudies,suchasthosepublishedbyGroszetal.(1983,1995),presentacenteringframeworkasamodeltoexplainthecoherenceoflocaldiscoursesegmentsinwhichthespeaker’sfocusofattentionisrelatedtoreferringexpressions.Thismodelhasachievedsuccessfulresultsinanaphoraresolutioninmonologues,butwouldrequirecertainmodicationstobesuccessfullyappliedtodialogues.Alongthoselines,ByronandStent1.Wehavedealtexclusivelywithpronominalandadjectivalanaphorabecausetheyappearedmostfre-quentlyinthed