Literature-Review-英文文献综述模板

TextRecognitionwithMachineLearningbasedonTextStructureLiteratureReviewYifanShiStudentID:27291944Email:[e-mail]ficialIntelligenceFacultyofPhysicalSciences&Eng,UniversityofSouthamptonAbstract—ThefastdevelopingMachineLearningalgorithmsintroducedtosemanticareanowadayshasbroughtvasttechniquesintextrecognition,classification,andprocessing.However,thereisalwaysacontradictionbetweenaccuracyandspeed,ashigheraccuracygenerallyrepresentsmorecomplicatedsystemaswellaslargetrainingdatabase.Inordertoachieveabalancebetweenfastspeedandgoodaccuracy,manybrilliantdesignsareusedintextprocessing.Inthisliteraturereview,theseeffortsareintroducedinthreelayers:Natural-LanguageProcessing,TextClassification,andIBMWatsonSystem.Keywords—MachineLearning,Natural-LanguageProcessing,TextClassification,IBMWatsonI.INTRODUCTIONThegrowingpopularityoftheInternethasbroughtincreasingnumberofusersonline,withavastamountofmessages,blogs,articles,etc.tobedealtwith.Thesetexts,knownasnatural-languagetexts,containpossibleusefulinformationbuttakealongtimeforhumantoread,understandanddealwith.Despitethepopularsearchenginetechnologynowadaysinhelpinguserstofindthesourceswithkeywords,semantictechniquesarealsoneededbymanycompaniestoimprovetheiruser-friendlyworkingenvironment.Inthisliteraturereview,Iwillintroduceseveralimportantsemantictechniques,startingfromthemostbasicNatural-LanguageProcessing,concentratinginthemeaningofwordsandsentences,followedbyTextClassificationwhichisfocusedonparagraphsandarticles.Then,IwillintroducealandmarksystemnamedIBMWatson,whichhasDeepQAasitsworkingpipeline.Finally,aconclusionwillbeincludedtogivesomecommentsonthesetechniques.II.NATURALLANGUAGEPROCESSINGInordertodealwiththehumannatural-language,itisnecessarytotransformtheunstructuredtextintowell-structuredtablesofexplicitsemantics(Ferrucci,2012).AccordingtoLiddy(2001),Natural-LanguageProcessing(NLP)isaseriesofcomputationaltechniquesusedtoanalyzeandrepresentnaturallyorganizedtextinordertoachievecertaintasksandapplications.CollobertandWeston(2008)havecategorizedNLPtasksintosixtypes:Part-Of-SpeechTagging,Chunking,NamedEntityRecognition,SemanticRoleLabeling,LanguageModels,andSemanticallyRelatedWords.Inadditiontothis,theyalsoimplementedMultitaskLearningwithDeepNeuralNetworkstobuildasuccessfulunifiedarchitecturewhichavoidedtraditionallargeamountofempiricalhand-designedfeaturestotrainthesystembyusingbackpropagationtraining(Collobertetal.,2011).III.TEXTCLASSIFICATIONOneofthesimplewaytorepresentanarticleforalearningalgorithmistousethenumberoftimesthatdistinctwordsappearinthedocument(Joachims,2005).However,duetothelargeamountofpossiblewordsusedinarticles,itwouldcreateaveryhighdimensionalspaceoffeatures.Joachims(1999)suggestsaTransductive1TextRecognitionwithMachineLearningbasedonTextStructureSupportVectorMachinestodoclassificationbecauseofitseffectivelearningabilityeveninhighdimensionalfeaturespace.Ratherthanusingnon-linearSupportVectorMachine(SVM),Dumaisetal.(1998)comparedlinearSVMwithanotherfourdifferentlearningalgorithmswhichareFindSimilar,DecisionTrees,NaiveBayes,andBayesNets,whichalsosupportsSVMintextclassificationbecauseofitshighaccuracy,fastspeedaswellasitssimplemodel.Sebastiani(2002)alsorecommendsNeuralNetworkasapotentialselectionintextclassificationinthatitsaccuracyisonlyslightlylowerthanSVMincomparison.Thecross-documentcomparisonofsmallpiecesoftext,usinglinguisticfeaturessuchasnounphrases,andsynonymsisintroducedbyHatzivassiloglouetal.(1999).Thesimilarityoftwoparagraphsisdefinedbythesameactionconductedonthesameobjectbythesameactor.Therefore,drawingfeaturesaccordingtonounsandverbswouldgenerallyconcludeaparagraphintoseveralprimitiveelements.Inadditiontothesimilarprimitiveelements,restrictionssuchasordering,distancesandprimitive(matchingnounandverbpairs)arealsoimplementedtoexcludeweaklyrelatedfeatures.Thefeatureselectionmethodscaneffectivelyreducethedimensionsofdataset(Ikonomakis,2005)whilekeepingtheperformanceofclassification.Tomakesurewhichwordsaretobekept,anEvaluationfunctionhasbeenintroducedbySoucyandMineau(2003)tomeasurehowmuchinformationwecangetbyclassifyingthroughasingleword.AnotherimprovementbyHanetal.(2004)istousePrincipalComponentAnalysis(PCA)toreducethedimensionintransformationoffeatures.NigamandMccallum(2000)combineExpectation-MaximizationandNaiveBayesclassifiertotraintheclassifierwithcertainamountoflabeledtextsfollowedbylargeamountofunlabeleddocuments,whichrealizestheautomatictrainingwithouthugeamountofhand-designedtrainingdata.IV.IBMWATSONTheIBMWatsonprojecthasshownusthatcomputersysteminopen-domainquestion-answering(QA)ispossibletobeathumanchampionsinJeopardy.AsFerrucci(2012)mentioned,thestructureofWatsonismorecomplicatedthananysingleagentasithashundredsofalgorithmsworkingtogether,inthewaythatMinsky(1988)introducedinSocietyofMind.Generally,WatsonconsistsofpartswhichareDeepQA,NaturalLanguageProcessing(NLP),MachineLearning(ML),andSemanticWebandCloudComputing(Gliozzoetal.,2013).TheDeepQAsystemanalyzesthequestionbydifferentalgorithms,givingdifferentinterpretationsofquestionsa