Sentimentanalysis,ahotresearchtopic,presentsnewchallengesforunderstandingusers＇opinionsandjudg-mentsexpressedonline.Theyaimtoclassifythesubjectivetextsbyassigningthemapolaritylabel.Inthispaper,weintroduceanovelmachinelearningframeworkusingauto-encodersnetworktopredictthesentimentpolaritylabelatthewordlevelandthesentencelevel.Inspiredbythedimensionalityreductionandthefeatureextractioncapabilitiesoftheauto-encoders,weproposeanewmodelfordistributedwordvectorrepresentation＂PMI-SA＂usingasinputpointwise-mutual-information＂PMI＂wordvectors.Theresultedcontinuouswordvectorsarecombinedtorepresentasentence.Anunsupervisedsentenceembeddingmethod,calledContextualRecursiveAuto-Encoders＂CoRAE＂,isalsodevelopedforlearningsentencerepresentation.Indeed,CoRAEfollowsthebasicideaoftherecursiveauto-encoderstodeeplycomposethevectorsofwordsconstitutingthesentence,butwithoutrelyingonanysyntacticparsetree.TheCoRAEmodelconsistsincombiningrecursivelyeachwordwithitscontextwords（neighbors＇words：previousandnext）byconsideringthewordorder.Asupportvectormachineclassifierwithfine-tuningtechniqueisalsousedtoshowthatourdeepcompositionalrepresentationmodelCoRAEimprovessignificantlytheaccuracyofsentimentanalysistask.Experimentalresultsdemon-stratethatCoRAEremarkablyoutperformsseveralcompetitivebaselinemethodsontwodatabases,namely,SanderstwittercorpusandFacebookcommentscorpus.TheCoRAEmodelachievesanefficiencyof83.28%withtheFacebookdatasetand97.57%withtheSandersdataset.
