AspectBasedSentimentAnalysis
AcknowledgmentsWe would like to give special thanks to our mentors Prof. Smaranda Muresan and Data ScientistPeter Deng for providing their insight and expertise that greatly assisted the research.
Introduction&DatasetTherapidgrowthofe-commercehasledtoanupsurgeofcustomerreviewsthatconstituteimportantsourceofinformationforbothcustomersandbusinesses.Tosummarisecustomers’opinionsonproductsfromtheoverwhelmingamountofreviews,theAspectBasedSentimentAnalysis(ASBA)problemisraised.Thisprojectrecommendsamethodtoextractaspectsfromreviews,andtoidentifythesentimentpolarityofeachcustomerreviewexpressedtowardsspecificaspects.
ReferencesPavlopoulos,Ioannis(2014).“Aspectbasedsentimentanalysis”.In:AthensUniversityofEconomicsandBusiness.
Pontiki,Mariaetal.(2016).“SemEval-2016task5:Aspectbasedsentimentanalysis”.In:Proceedingsofthe10thinternationalworkshoponsemanticevaluation(SemEval-2016),pp.19– 30.
SemEval-2016ABSARestaurantReviews-English:TrainData(n.d.).url:http://metashare.ilsp.gr:8080/.Levy,O.,&Goldberg,Y.(2014).Dependency-basedwordembeddings.InProceedingsofthe52ndAnnualMeeting
oftheAssociationforComputationalLinguistics(Volume2:ShortPapers)(Vol.2,pp.302-308).
AspectDetectionAspectdetectionaimstoderiveaspectsthataredirectlydiscussedinreviewsanditonlyfocusesonexplicitaspects.Themethodcoupleswordembeddingwithclusteringtechniques.Dependency-basedwordembeddingischosenbecauseitcapturesbothsemanticandsyntacticsimilarities.Then,agglomerativeclusteringmethodwithcosinesimilarityisappliedtoclusterwordvectors
ConclusionsThisprojectexploredandexperimentedwithbothsupervisedandunsupervisedABSAmethodsextensively.AppliedvariousNLPtechniquesfromlinearSVMtotransferlearning(pre-trainedwordembeddings)anddeeplearning.Comparedadvantagesanddrawbacksofdifferentapproaches.IntegratedaspectdetectionwithsentimentanalysisandfinallyproposedchainingagglomerativeclusteringtogetherwithSVM+DNNasageneralmethodtosummarise feedbacksfromcustomerreviewsbyaspects.
AspectSpecificSentimentAnalysisSentimentanalysisincludestwophases.Phase(i),foreachaspectinthecandidatelist,abinaryclassifieristrainedtodetectwhetherornotareviewisdescribingit.Thenaspectsarepredictedusingonevstherest.Featureweightsofaclassifierlearnedinphase(i)arefedasfeaturestoclassifiersinphase(ii)todetectsentimentpolarities.
Topic:AspectBasedSentimentAnalysis
Givenareviewasinput,detectsentimentsexpressedtowardsaspectsthatareofconcerntomostcustomers
WordCloud:wordfrequencyinallreviews
Histogramofhuman-annotatedaspectsinthedata
Liutong Zhou(lz2484),Jiachen Xu(jx2318),ZihanYe(zy2293),Youyang Liu(yl3767)IndustryMentor:PeterDeng
FacultyMentor:Smaranda Muresan
Histogramofpolaritiesinthedata
Dendrogram
WordsExpansionofCluster1 EvaluationofAspectDetection
Food,ambience,serviceaspectsdetectionresultsareevaluatedbyprecision,recallandf1-score.Addingexpansionwords
additionaltoclusteringresultcansignificantlyincreaseprecisionsandrecallsforallaspectsTheoptimalnumberofclusterischosentobe15becausethere
issteepincreaseinoverallsimilarityatthatpoint.
Dependency-BasedContextExtractionExample
NumberofClustersvsOverallSimilarity
Algorithm
DictionarySizeImpactsonAspectDetection