基于超图分割的共指消解分析-analysis of coreference resolution based on hypergraph segmentation.docxVIP

AbstractCoreferenceresolutionisthetaskofgroupingmentionsofentitiesintosetsso thatallmentionsinonesetrefertothesameentity.Coreferenceresolutionisan importancesubtaskofinformationextraction,hasimportantapplicationinvarvious fieldsofnaturallanguageprocessingandinformationretrieval.Mostrecentapproachestocoreferenceresolutiondividethistaskintotwostep- s:1)aclassificationprocesstodetermineswhetherapairmentionpairiscoreference ortheconfidenceofcoreferencepair;2)aclusteringprocesstogroupsmentionsinto entitiesbasedontheoutputofstep1.Inthispaper,weusethehypergraphpartition cleverlysolvethecoreferenceresolutionproblem,toavoidthethedivisionintotwo steps,butfromaglobaloptimizationperspective,resolvecoreferenceresolution.We representadocumentasahypergraph,wheretheverticesdenotementionsandthe edgesdenoterelationalfeaturesbetweenmentions.Coreferenceresolutionfromthe perspectiveofglobaloptimization,willexceedthediagramisdividedintomultipleindependentsubgraph，eachsubgraphrepresentsanentitychain.Thispaperfocusesonthreeparts:thehypergraphmodelingforcoreferencere- solution;hypergraphpartitionforcoreferenceresolution;hyperedgeweightlearning basedonunlabelcorpus.Coreferenceresolutionhypergraphmodelingismainlyconcentratedinthehyp- ergraphmodelingforcoreferenceresolution,includethehyperedgeandhyperedgeweight. We have introduced a negative connection,strongly connected，twosuperedgeconcepts,depictingthelinksbetweenentities.Weproposedtwohyperedge weightlearningmethods.Hypergraphpartitionforcoreferenceresolutionrefertoconvertcoreferencer- esolutiontohypergraphpartition.Fromthekwayhypergraphpartitionanditerator 2wayhypergraphpartitioningtosolvethecoreferenceresolutionproblem.kway hypergraphpartitionoptimizationhypergraphcuttinglossed,bymakingsurethe numberofsubgraphtodeterminethenumberofentitychain.Iterator2waypartitionII-tooptimizethetoleranceofthepartition,andgraduallysplitsubgraphtodetermine theentitychain.IntheACE05Chinesecorpus,andtraditionalmethodsprovedthe effecttivenessofourwork.Hyperedg