资源描述:
《Recursive Deep Models for Discourse Parsing 》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。
1、RecursiveDeepModelsforDiscourseParsingJiweiLi1,RumengLi2andEduardHovy31ComputerScienceDepartment,StanfordUniversity,Stanford,CA94305,USA2SchoolofEECS,PekingUniversity,Beijing100871,P.R.China3LanguageTechnologyInstitute,CarnegieMellonUniversity,Pittsburgh,PA15213,USAjiweil@stanford.edual
2、icerumeng@foxmail.comehovy@andrew.cmu.eduAbstractHirst,2012;Jotyetal.,2013))arestillconsid-erablyinferiorwhencomparedtohumangold-Text-leveldiscourseparsingremainsastandarddiscourseanalysis.Thechallengestemschallenge:mostapproachesemployfea-fromthefactthatcomparedwithsentence-levelturest
3、hatfailtocapturetheintentional,se-dependencyparsing,thesetofrelationsbetweenmantic,andsyntacticaspectsthatgoverndiscourseunitsislessstraightforwardtodefine.discoursecoherence.Inthispaper,wepro-Becausetherearenoclause-level‘partsofdis-posearecursivemodelfordiscoursepars-course’analogousto
4、word-levelpartsofspeech,ingthatjointlymodelsdistributedrepre-thereisnodiscourse-levelgrammaranalogoustosentationsforclauses,sentences,anden-sentence-levelgrammar.Tounderstandhowdis-tirediscourses.Thelearnedrepresenta-courseunitsareconnected,onehastounderstandtionscantosomeextentlearnthe
5、seman-thecommunicativefunctionofeachunit,andtheticandintentionalimportofwordsandroleitplayswithinthecontextthatencapsulatesit,largerdiscourseunitsautomatically,.Thetakenrecursivelyallthewayupfortheentiretext.proposedframeworkobtainscomparableManuallydevelopedfeaturesrelatingtowordsandpe
6、rformanceregardingstandarddiscours-othersyntax-relatedcues,usedinmostofthere-ingparsingevaluationswhencomparedcentprevailingapproaches(e.g.,(FengandHirst,againstcurrentstate-of-artsystems.2012;Hernaultetal.,2010b)),areinsufficientforcapturingsuchnestedintentionality.1IntroductionRecently
7、,deeplearningarchitectureshavebeenInacoherenttext,units(clauses,sentences,andappliedtovariousnaturallanguageprocessinglargermulti-clausegroupings)aretightlycon-tasks(fordetailsseeSection2)andhaveshownnectedsemantically,syntactically,andlogically.theadvantagestocapturetherelevan