Recursive neural conditional random fields for aspect-based sentiment analysis

《Recursive neural conditional random fields for aspect-based sentiment analysis》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、RecursiveNeuralConditionalRandomFieldsforAspect-basedSentimentAnalysisWenyaWang†‡SinnoJialinPan†DanielDahlmeier‡XiaokuiXiao††NanyangTechnologicalUniversity,Singapore‡SAPInnovationCenterSingapore†{wa0001ya,sinnopan,xkxiao}@ntu.edu.sg,‡{d.dahlmeier}@sap.comAbstractliverytimesinthec

2、ity.”,theaspecttermisdeliverytimes,andtheopiniontermisfastest.Inaspect-basedsentimentanalysis,extract-Amongpreviouswork,oneoftheapproachesingaspecttermsalongwiththeopinionsbe-istoaccumulateaspectandopiniontermsfromaingexpressedfromuser-generatedcontentisseedcollectionwithoutlabel

3、information,byutiliz-oneofthemostimportantsubtasks.Previ-ingsyntacticrulesormodificationrelationsbetweenousstudieshaveshownthatexploitingcon-nectionsbetweenaspectandopiniontermsisthem(Qiuetal.,2011;Liuetal.,2013b).Inthepromisingforthistask.Inthispaper,wepro-aboveexample,ifweknowfa

4、stestisanopinionposeanoveljointmodelthatintegratesrecur-word,thendeliverytimesisprobablyinferredtobesiveneuralnetworksandconditionalrandomanaspectbecausefastestisitsmodifier.However,fieldsintoaunifiedframeworkforexplicitas-thisapproachlargelyreliesonhand-codedrulesandpectandopiniont

5、ermsco-extraction.TheisrestrictedtocertainPart-of-Speech(POS)tags,proposedmodellearnshigh-leveldiscrimina-e.g.,opinionwordsarerestrictedtobeadjectives.tivefeaturesanddoublepropagatesinforma-tionbetweenaspectandopinionterms,simul-Anotherapproachfocusesonfeatureengineeringtaneously

6、.Moreover,itisflexibletoincor-basedonpredefinedlexicons,syntacticanalysis,poratehand-craftedfeaturesintotheproposedetc.(JinandHo,2009;Lietal.,2010).Asequencemodeltofurtherboostitsinformationextrac-labelingclassifieristhenbuilttoextractaspectandtionperformance.Experimentalresultsonth

7、eopinionterms.Thisapproachrequiresextensiveef-datasetfromSemEvalChallenge2014task4fortsfordesigninghand-craftedfeaturesandonlyshowthesuperiorityofourproposedmodelcombinesfeatureslinearlyforclassificationwhichoverseveralbaselinemethodsaswellastheignoreshigherorderinteractions.winni

8、ngsystemsofthechallenge.Toovercomethelim