资源描述:
《2009-Named entity recognition in query》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、NamedEntityRecognitioninQueryJiafengGuo†,GuXu‡,XueqiCheng†,HangLi‡†‡InstituteofComputingTechnology,CASMicrosoftResearchAsiaBeijing,P.R.ChinaBeijing,P.R.Chinaguojiafeng@software.ict.ac.cn,cxq@ict.ac.cn{guxu,hangli}@microsoft.comABSTRACTentityandassign“Game”toitasthemostlikelyclass,“Mov
2、ie”and“Book”aslesslikelyclasses,and“Music”asThispaperaddressestheproblemofNamedEntityRecog-unlikelyclass.Thisisbecausethecontext“walkthrough”nitioninQuery(NERQ),whichinvolvesdetectionofthestronglyindicatesthat“harrypotter”hereismorelikelytonamedentityinagivenqueryandclassificationofthe
3、namedmeantheHarryPottergame.(Ifthequeryisonly“harryentityintopredefinedclasses.NERQispotentiallyusefulinpotter”,then“Book”and“Movie”willbemoreplausible.)manyapplicationsinwebsearch.Thepaperproposestak-NERQisessentiallyusefulformanyapplicationsinwebingaprobabilisticapproachtothetaskusin
4、gquerylogdatasearch.Accordingtoouranalysis,about71%ofsearchandLatentDirichletAllocation.Weconsidercontextsofaqueriescontainnamedentities.Identifyingnamedentitiesnamedentity(i.e.,theremaindersofqueriesafterthenamedinquerieswouldhelpustounderstandsearchintentsbetter,entityisremoved)aswo
5、rdsofadocument,andclassesoftheandthereforeprovidebettersearch.Forexample,inrele-namedentityastopics.Thetopicmodelisconstructedbyavancesearch,wecanimproverankingbytreatingnamednovelandgenerallearningmethodreferredtoasWS-LDAentityandcontextseparately;inquerysuggestion,wecan(WeaklySuperv
6、isedLatentDirichletAllocation),whichem-generatemorerelevantsuggestions,e.g.“harrypotterwalk-ploysweaklysupervisedlearning(ratherthanunsupervisedthrough”→“harrypottercheats”(contextinthesameclass)learning)usingpartiallylabeledseedentities.Experimentalor“halo3walkthrough”(entityinthesam
7、eclass).resultsshowthattheproposedmethodbasedonWS-LDAAsfarasweknow,therewasnopreviousworkonNERQ.canaccuratelyperformNERQ,andoutperformthebaselineTraditionallyNamedEntityRecognition(NER)ismainlymethods.performedonnaturallanguagetexts[6,3,8].Usuallyasu-pervisedlearningapproachisexploite
8、danda