Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank

《Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank》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、RecursiveDeepModelsforSemanticCompositionalityOveraSentimentTreebankRichardSocher,AlexPerelygin,JeanY.Wu,JasonChuang,ChristopherD.Manning,AndrewY.NgandChristopherPottsStanfordUniversity,Stanford,CA94305,USArichard@socher.org,faperelyg,jcchuang,angg@cs.stanford.edufjeaneis,manning,cgpottsg@stanfo

2、rd.eduAbstract–0–Semanticwordspaceshavebeenveryuse-00–0Thisfilm.fulbutcannotexpressthemeaningoflonger–0phrasesinaprincipledway.Furtherprogress00++doesn’tcaretowardsunderstandingcompositionalityin0+abouttaskssuchassentimentdetectionrequires++richersupervisedtrainingandevaluationre-+00+orsourcesand

3、morepowerfulmodelsofcom-+0000+witanyofposition.Toremedythis,weintroducea+00++++cleverness,otherkindintelligenthumorSentimentTreebank.Itincludesfinegrainedsentimentlabelsfor215,154phrasesintheFigure1:ExampleoftheRecursiveNeuralTensorNet-parsetreesof11,855sentencesandpresentsworkaccuratelypredictin

4、g5sentimentclasses,veryneg-newchallengesforsentimentcomposition-ativetoverypositive(––,–,0,+,++),ateverynodeofaality.Toaddressthem,weintroducetheparsetreeandcapturingthenegationanditsscopeinthisRecursiveNeuralTensorNetwork.Whensentence.trainedonthenewtreebank,thismodelout-performsallpreviousmeth

5、odsonseveralmet-rics.Itpushesthestateoftheartinsinglesentencepositive/negativeclassificationfrommodelstoaccuratelycapturetheunderlyingphe-80%upto85.4%.Theaccuracyofpredictingnomenapresentedinsuchdata.Toaddressthisneed,fine-grainedsentimentlabelsforallphrasesweintroducetheStanfordSentimentTreebanka

6、ndreaches80.7%,animprovementof9.7%overbagoffeaturesbaselines.Lastly,itistheonlyapowerfulRecursiveNeuralTensorNetworkthatmodelthatcanaccuratelycapturetheeffectscanaccuratelypredictthecompositionalsemanticofnegationanditsscopeatvarioustreelevelseffectspresentinthisnewcorpus.forbothpositiveandnegat

7、ivephrases.TheStanfordSentimentTreebankisthefirstcor-puswithfullylabeledparsetreesthatallowsfora1Introductioncompleteanalysisofthecompositionaleffectsofsentimentinlanguage.ThecorpusisbasedonSemanticvectorspacesforsinglewordsh