LING287CS424PStanfordPottsSentimentlexicons33MicroWNOpCerinietal2007Documentationanddownloadhttpwwwunipvitwnop PosNegSynset 110truea2reala4210illustriousa1famousa13050real ID: 515934
Download Pdf The PPT/PDF document "LING287/CS424P,Stanford(Potts)Sentimentl..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
LING287/CS424P,Stanford(Potts)Sentimentlexicons3.2WordNet(Miller1995;Fellbaum1998)Download/documentation:http://wordnet.princeton.edu/Webinterface:http://wordnetweb.princeton.edu/perl/webwnAPIs:http://wordnet.princeton.edu/wordnet/related-projects/idle,aSynset:idlea01Denition:notinactionoratworkExamples:fanidlelaborer,idledrifters,theidlerich,anidlemindgAlsosees:fineectivea01,unemployeda01gSimilartos:funengageds01,...,bone-idles01g...Lemma:idlea01idle·Antonyms:fbusya01busyg·Derivationallyrelatedforms:ffaineancen01idlenessg·Pertainyms:fg·...Synset:baselesss01Denition:withoutabasisinreasonorfactExamples:fbaselessgossip,unfoundedsuspicions,...,unwarrantedjealousygAlsosees:fgSimilartos:funsupporteda01g...Lemma:baselesss01baseless·Antonyms:fg·Derivationallyrelatedforms:fg·Pertainyms:fg·...Lemma:baselesss01groundless·......Table3:Examplefromstringstolexicalstructure.FieldsmarkedwiththeemptysethappentobeemptyforthisexamplebutcanbelledInWordNet3.0(Miller2009),thestringidlewithcategoryahas7synsets.Collectively,thosesynsetshave18lemmas.Eachsynsetandeachlemmacanreachothersynsets(lemmas)viarelationslikethoseexempliedabove...(seesec.5.1).OpenThesaurus(German):http://www.openthesaurus.de/Moreglobally:http://www.globalwordnet.org/(withfreedownloadsforatleastArabic,Danish,French,Hindi,Russian,Tamil)3 LING287/CS424P,Stanford(Potts)Sentimentlexicons3.3Micro-WNOp(Cerinietal.2007)Documentationanddownload:http://www.unipv.it/wnop PosNegSynset 110truea2reala4210illustriousa1famousa1...30.50reala6tangiblea240.250existenta2reala150.1250.125reala2...11000demandv6 (a)`Common':Fiveevaluatorsworkingtogether,110synsets. Evaluator1Evaluator2Evaluator3Pos1Neg1Pos2Neg2Pos3Neg3Synset 1101010gooda15wella2210100.750sweet-smellinga1perfumeda2...3101010gooda23unspoilta1unspoileda140.500.2500.250hota16...4960.50100.50healv3bring_aroundv2curev1 (b)`Group1':Threeevaluatorsworkingseparately,496synsets.Completeagreementon197(40%).Polarityagreement(sign(Pos1Neg1)=sign(Pos2Neg2)=sign(Pos3Neg3))on387(78%). Evaluator1Evaluator2Pos1Neg1Pos2Neg2Synset 10101forlorna1godforsakena2lorna1desolatea220101rottena231010intimatea2cozya2informala440000federala1...4990000termv1 (c)`Group2':Twoevaluatorsworkingseparately,499synsets.Completeagreementon395(79%).Polarityagreement(sign(Pos1Neg1)=sign(Pos2Neg2))on471(90.4%).Table4:ThethreegroupsoftheMicro-WNOp.ThestringslistedunderSynsetidentifylemmas.4 LING287/CS424P,Stanford(Potts)Sentimentlexicons PosNegObjTotal a(adj)624550157n(noun)3058273361r(adv)402428v(verb)273693156 Total330139440 702 (a)Bysynset. PosNegObjTotal a15113589375n83144663890r905362v87112233432 Total3303911,038 1,759 (b)Bylemma. PosNegObjTotal a10511362280n66107483656r703744v6890178336 Total246310760 1,316 (c)Byhstring,tagipair. (d)Thedistributionofscores,bylemma.Objectivityvaluesare1(Pos+Neg),soascoreof1theremeansthatbothPositiveandNegativewere0.Relativelyfewofthesynsetsreceivednon-0/1ratings.Table5:Micro-WNOplimitedtothe702synsetsonwhichallannotatorsagreedexactlyonallvalues(tab.5(a)).Posmeansthatthepositivescorewashigher,Negthatthenegativescorewashigher,andObjthatthetwoscoreswerethesame(eveniftheyweregreaterthan0).3.4POStags:relatingHarvardInquirerandWordNet/Micro-WNOpmatches`noun'(case-insensitive)7!nmatches`verb'or`supv'(case-insensitive)7!vmatches`modif'(case-insensitive)7!ais`LY'ormatches`LY'(case-sensitive)7!rmatchesnoneoftheabovenochangeTable6:PartialmappingfromInquirercategoriestoWordNetcategories.Themappingispartialfortworeasons.First,Inquirerincludesdeterminers,prepositions,andotherpartsofspeechnotfoundinWordNet(andhencenotrepresentedatallinMicro-WNOp).Second,evensettingthoseclassesaside,Inquirerparts-of-speecharehighlysubcategorizedanddon'tnecessarilyfallintothefour-wayWordNettypology.5 LING287/CS424P,Stanford(Potts)Sentimentlexicons4AfewnotesonassessmentPredictedPosNegObjPos 15 10100ObservedNeg10 15 10Obj10100 1,000 sum_diagonal sum_all_cells (a)Accuracy.Typicallyappropriateonlyifthecategoriesareequalinsize.PredictedPosNegObjPos 15 10 100ObservedNeg 10 15 10Obj 10100 1,000 true_positive true_positive+false_positive (b)Precision:thecorrectguessespenalizedbythenumberofincorrectguesses.YoucanoftengethighprecisionforacategoryCbyrarelyguessingC,butthiswillruinyourrecall.PredictedPosNegObjPos 15 10 100ObservedNeg10 15 10Obj 10 100 1,000 true_positive true_positive+false_negative (c)Recall:thecorrectguessespenalizedbythenum-berofmisseditems.YoucanoftengethighrecallforacategoryCbyalwaysguessingC,butthiswillruinyourprecision.Table9:Confusionmatrices.Thesevaluesshouldallbeinterpretedasboundedby[0,1].(Ifyouaretemptedtodivideby0,return0.)Denition1(F1).2precision*recall precision+recallNote:F1givesequalweighttoprecisionandrecall.However,dependingonwhatyou'restudying,youmightvalueonemorethantheother.Denition2(Macroaverage).Averageprecision/recall/F1overalltheclassesofinterest.Denition3(Microaverage).Sumcorrespondingcellstocreatea22confusionmatrix,andcalculateprecision=recall=F1intermsofthatnewmatrix.PosNegObjPos1510100Neg101510Obj101001,000)264PosOtherPos15110Other201,125NegOtherNeg1520Other1101,125ObjOtherObj1,000110Other11050375)YesNoYes1,030240No2402,300Note:Beforeusingthesesummarymeasures,lookatthedistributionofyourby-categorynumberstomakesurethatyou'renotlosingimportantstructure.Formore:ManningandSchütze1999:§8.1;Manningetal.2009:§8.3,13.6.7 LING287/CS424P,Stanford(Potts)Sentimentlexicons5WordNet-basedapproaches5.1Simplesense/sentimentpropagationHypothesis:Sentimentisconstantthroughoutregionsoflexicallyrelateditems.Thus,sentimentpropertiesofhand-builtseed-setswillbepreservedaswefollowWordNetrelationsoutfromthem.5.1.1Algorithm(Valituttietal.2004;EsuliandSebastiani2006;sec.5.3forprecedents)WORDNETSENSEPROPAGATE(S,iter)InputS:alistofsynsets.Forexample: fbrilliants01,nwin01g,fsadlyr01,grossa01giter:thenumberofiterationsOutputT:LENGTH(S)1+itersynsetmatrix:0BB@S[1]iter-thpropagationofS[1]......S[LENGTH(S)]iter-thpropagationofS[n]1CCA1initializeT:aLENGTH(S)1+itermatrixsuchthatT[i][1]=S[i]for1¶i¶1+iter2fori 1toiter3forj 1toLENGTH(S)4newSame SAMEPOLARITY(T[j][i])5others SLENGTH(S)k=1T[k][i]fork6=jTheotherseed-setsinthiscolumn.6newDiff OTHERPOLARITY(others)Fortheexperiments,Irstcalculateallthepropagationsetsandtheneliminatetheirpairwiseintersectionfromeach,toensurenooverlap.7T[j][i+1] (newSame[newDiff)8returnTSAMEPOLARITY(synsets)1newsynsets fg2fors2synsetsSynset-levelrelations.3newsynsets newsynsets[fsg[AlsoSees(s)[SimilarTos(s)4forlemma2Lemmas(s)Lemma-levelrelations.5foraltLemma2(DerivationallyRelatedForms(lemma)[Pertainyms(lemma))6newsynsets newsynsets[fSynset(altLemma)g7returnnewsynsetsOTHERPOLARITY(synsets)1newsynsets fg2fors2synsets3forlemma2Lemmas(s)Lemma-levelrelations.4foraltLemma2Antonyms(lemma)5newsynsets newsynsets[fSynset(altLemma)g6returnnewsynsets8 LING287/CS424P,Stanford(Potts)Sentimentlexicons FreevariablesTheseed-sets,theWordNetrelationscalledinSAMEPOLARITYandOTHERPOLARITY,thenumberofiterations,thedecisiontoremoveoverlap. 5.1.2ExperimentFormulateseed-setsbyhand,usingintuitions.Usethealgorithmtoderivepositive,negative,andobjectivesetsofhstring,posipairs.positiveexcellent,good,nice,positive,fortunate,correct,superiornegativenasty,bad,poor,negative,unfortunate,wrong,inferiorobjectiveadministrative,nancial,geographic,constitute,analogy,ponder,material,public,department,measurement,visualTable10:Seed-sets. Figure1:Growthinthesizeofthesetsasthenumberofiterationsgrows. Figure2:Percentageofitemsfromthegoldstandardthatareunreachedbypropagation.9 LING287/CS424P,Stanford(Potts)Sentimentlexicons (a)F1usingdef.4,option(i):Scoreonlyhstring,posipairsthatareintheintersectionofInquirerWN(Micro-WNOp)withtheunionofthethreederivedsets. (b)F1usingdef.4,option(ii):Treatallitemsnotreachedbythepropagationalgorithmasmembersofobjective. (c)F1usingdef.4,option(iii):Createacategory`missing'topenalizeeffectivenessmeasuresfortheothercategories.Figure3:UsingF1toassessthepropagationalgorithmagainstInquirerandMicro-WNOp.Thethreescoringoptionsemphasizedifferentaspectsofthealgorithm.11 LING287/CS424P,Stanford(Potts)Sentimentlexicons5.2.2G08resultsInourexperiments,theoriginalseed-setcontained20neg-ativeand47positivewordsthatwereselectedbyhandtomaximizedomaincoverage,aswellas293neutralwordsthatlargelyconsistofstopwords.[...]Runningthealgorithmresultedinanexpandedsentimentlexiconof5,705positiveand6,605negativewords,someofwhichareshowninTable2withtheirnalscores.Adjectivesformnearly90percentoftheinducedvocabulary,followedbyverbs,nounsandnallyadverbs. G08donotstophere.Theygoontousetheirlexiconforclassicationandsummarization.We'llreturntothoseportionsofthepaperinthenexttwoclasses.5.2.3OurexperimentMicro-WNOpprovidesthesetofseed-sets:weusetheentiresetofpositive,negative,andobjectivesetsasdenedintab.5(c).Theideaisthatwewillusethisalgorithm,nottolearnnewsentimentlexicons,butrathertolearnthestrengthrelationshipsbetweenwordsofthesamepolarity. Seed-setAfterpropagation P2461,021N310978M760154,585 Total1,316156,584 Table11:Experimentalresults.Thecombinedseeds-setsaretheentireMicro-WNOpcorpus.ThetotalafterpropagationisthenumberoflemmasinWordNet3.0.ThegaininthepositiveandnegativecategoriesisnotaslargeasG08saw.ThisisalmostcertainlyduetothefactthattheMicro-WNOpcorpusdoesnotcoverasmuchofthedomainastheirseed-setsdid. Figure4:Distributionofnon-0scoresinthenallexicon.Thevastmajorityarecloseto0.Thelargestis185.12(valour)andthesmallestis61.1(rophy,`streetnamesforunitrazepan'(!?)).13 LING287/CS424P,Stanford(Potts)Sentimentlexicons5.3SimilarapproachesandideasHuandLiu2004Oneofthersttousethiskindofpropagationforsentiment.Theiralgorithmisverysimilartotheoneinsec.5.1.1,thoughitsinputsandoutputsarestringsratherthansynsetsanditrestrictsattentiontoadjectives(positiveandnegative).KimandHovy2006SensepropagationoverWordNetfromsmallseed-sets,dealingwithoverlapviamaximumlikelihood,whichalsodeliverssentimentscores/rankingsforwords(strings).Godboleetal.2007WordNetpropagationrelativetogeneralweblogtopiccategories,withscoresbasedondistancefromseed-setwords,takingintoaccountthedangersofmixedpaths. AndreevskaiaandBergler2006PropagationusingalimitedsetofbasicWordNetrelations,butwithcre-ativeuseoftheglossestofurtherexpandthelexicon.ThewordlistsofHatzivassiloglouandMcKeown1997(seesec.6.2)areusedtoformulatetheseed-sets,andtheHarvardInquireristhegoldstandard. RaoandRavichandran2009LabelpropagationusingWordNetsupplementedwithotherhand-buildresources.EvaluationispartlywithHarvardInquirer.TheapproachisextendedtoHindiandFrenchwithstrongresults.SentiWordNet(EsuliandSebastiani2006;Baccianellaetal.2010)Builtbyrstpropagatingoutfromthepositiveandnegativeseed-setsintab.10andthenbuildingaclassierfromtheresultingsynsets.Thefeaturesintheclassierarenotthesynsetsthemselves,butratherthesynsetsoftheirdenitions,asprovidedbythePrincetonAnnotatedGlossCorpus:2verygood;ofthehighestquality;madeanexcellentspeech[...]very3very4good1good3good4high3high4quality1quality3make2made3excellent3speech1http://sentiwordnet.isti.cnr.it/ 2http://wordnet.princeton.edu/glosstag.shtml15 LING287/CS424P,Stanford(Potts)Sentimentlexicons6Distributionalapproaches6.1Web-basedpropagationHypothesis:Thebasichypothesisofsec.5.2iscorrect,andwecanmovepasttheconnesofWordNetbyrelatingdistributionalsimilaritytosentiment.6.1.1Algorithm(Velikovichetal.2010) GisacosinesimilaritygraphPandNareseed-sets isathreshold:sentimentscoresbelowitwillberoundedto0Tisthenumberofiterations.Denition6(Cosinesimilarity).LetAandBbevec-torsofco-occurrencescountsforwordswAandwBdenedoveraxed,orderedvocabulary(rowsintab.13(b))oflengthn.Theircosinesimilarity:cosim(wA,wB)def=nPi=1(AiBi) r nPi=1A2ir nPi=1B2i6.1.2Examplesuperbamazingsuperbmoviesuperbmoviesuperbmoviesuperbmoviesuperbmovieamazingmovieamazingmoviecoolsuperb(a)Corpus.0BBBBB@amazingcoolmoviesuperbamazing0021cool0001movie2005superb11501CCCCCA(b)Co-occurrencematrix.0BBBBB@amazingcoolmoviesuperbamazing1.00.450.420.86cool0.451.00.930.0movie0.420.931.00.07superb0.860.00.071.01CCCCCA(c)Cosinesimilaritymatrix.Table13:Examplecorpusandsimilaritymeasures.Fig.5continuestheexample16 LING287/CS424P,Stanford(Potts)Sentimentlexicons (a)Cosinesimilaritygraphforthepositivesub-graphofasmallcorpus.0-valuededgesnotrepresented. (b)Initialmatrix0withseed-setP=fsuperbg. (c)Matrix1.AtthispointF=fsuperbg. (d)Matrix2.Fhasexpandedtoallthenodesing.5(a)reachablefromsuperbinonestep.coolandmoviebenetfromtheirclosenesstoamazing,whichisclosetotheseed-word.Yourvaluesmightbedifferentifyoumodifyinplaceasinline6ofthealgorithm,whichinvolvessomenvaluesinncalculations.Figure5:Theinnerloop(lines3-6)ofthewebpropagationalgorithmwithPfsuperbg.17 LING287/CS424P,Stanford(Potts)Sentimentlexicons6.2.4ThebutruleH&M97conjunctionsusingbutexhibittheoppositepattern,usuallyinvolvingadjectivesofdif-ferentorientations.Thus,arevisedbutstillsimplerulepredictsadifferent-orientationlinkifthetwoadjectiveshavebeenseeninabutconjunction,andasame-orientationlinkotherwise,assumingthetwoadjectiveswereseenconnectedbyatleastoneconjunction. Accuracy Theirdata81.81%Ourdata74% (a)Accuracy.Inaddition,Ipickedran-domlybalancedsubgroupswith20,000coordinationtokensineachcategory.Thebaselineisthen50%,buttheaver-ageaccuracyofthebut-ruleover10runswas67%.PredictedDifferentSameDifferent12,20116,170Same10,78766,265(b)Confusionmatrix. PrecisionRecallF1 Different53%43%47%Same47%86%61% (c)Effectiveness(ourdata)Table17:Assessingthebut-rule.6.2.5LogisticregressionCoefcientEstimateStandarderror 0Intercept1.560.01 1Coord=butthecoordinatorisbut0.850.04 2but-ruleasdescribedinsec.6.2.41.370.02 3Conj1Negatedtherstconjuncthasanexternalnegation1.040.15 4Conj1Negatedthesecondconjuncthasanexternalnegation0.620.07 5StemsMatchatleastoneadjiscomplexandthestemsmatch4.050.37 6Conj1Negated:Conj2Negatedinteraction:bothconjunctsnegatedexternally3.350.29 Table18:Coefcientestimateswithstandarderrors.Allofthefeaturesarebinary:1ifthedescriptionistrue,else0.Anegativecoefcentindicatesthatthefeaturebiasesinfavorofdifferentorientationfortheadjectives,whereasapositivecoefcientindicatesabiasforsameorientation.Thereareclearlyotherfeaturesandinteractionsthatcouldbetrieddatahomework1,problem3,asksyoutostartexploringotheroptions.20 LING287/CS424P,Stanford(Potts)SentimentlexiconsPr(same)=logit10BBBBBBBBB@1.56+0.851+1.371+1.041+0.621+4.050+3.35(11)+1CCCCCCCCCA=0.09(a)notout.stand.ingbutnotaw.fulPr(same)=logit10BBBBBBBBB@1.51+0.850+1.370+1.040+0.620+4.050+3.35(00)+1CCCCCCCCCA=0.82(b)healthyandrobustTable19:ExamplesDenition7(Logisticregressionpredictions).ThepredictioniscorrectifthepolaritydiffersandPr(same)¶0.5orifthepolarityisthesameandPr(same)0.5PredictedDifferentSameEmpiricalDifferent3,08219,702Same1,62770,376(a)Confusionmatrix. PrecisionRecallF1 Different65%14%22%Same78%98%87% (b)Effectiveness.Table20:Interimresults:Theabovemodelcorrectlyclassies77%ofthesamples,usingdef.7.6.2.6DissimilaritymeasureDenition8(Dissimilarity).disapartialfunctionfrom(wordswords)into[0,1]d(x,y)=8:1.0themedianttedvalueof(x,y)ifavailable,else1.0themedianttedvalueof(y,x)ifavailable,else0.5Note:Thettedmodelwilldelivermultiplevaluesforaadjectivepairsthatappearindifferentenvironments.AsfarasIcantell,H&M97don'tspecifyhowtomovefromthesevaluestoauniqueone.Ichosethemedianbecauseitseemstohaveagoodchanceofbeingintherightareaincaseswheretherearemultiplevalues,somesmallandsomelarge.Note:Thisisn'tadistancemeasurebecauseitfailstosatisfythetriangleinequality(thesumofthelengthsofanytwosidesofatriangleisgreaterthanthelengthofthethird).Asaresult,itisn'tappropriateforuseintheobjectivefunctionsforcentroidclusteringalgorithmslikek-means(HatzivassiloglouandMcKeown1993).21 LING287/CS424P,Stanford(Potts)Sentimentlexicons (10.4)+(10.9)+(10.3) 3+(10.1)+(10.9)+(10.01) 3=1.13 (10.4)+(10.9)+(10.3)+(10.8)+(10.9)+(10.2) 4+10.9 2=0.675Figure6:Exchangeofexcellent.6.2.8ClusterlabelingH&M97Theclusteringalgorithmseparateseachcomponentofthegraphintotwogroupsofadjectives,butdoesnotactuallylabeltheadjectivesaspositiveornegative.[...]Wecomputetheaveragefrequencyofthewordsineachgroup,expectingthegroupwithhigheraveragefrequencytocontainthepositiveterms.HereThefrequenciesforunigramsintheGoogleN-gramscorpus(BrantsandFranz2006).23 LING287/CS424P,Stanford(Potts)Sentimentlexicons FreevariablesThechoiceofclusteringalgorithms,theinherentrandomnessintheini-tialpartition,theorderinwhichthewordsaregonethrough(whichcanleadtodifferentlocaloptima),thenatureofthedictionaryd,whichisdeterminedbythefeaturesinthelogisticregression,thefrequencymeasurethatdeterminesclasslabels. 6.2.9Assessment Table21:H&M97'sresults. Figure7:Assessmentofthecoordinationexperimentwithtwoslightlydifferentmodelts.Theclusterlabelingprocedurewaseffective,inthatnotrialhadunder50%accuracy,whichwouldhavebeenasignthatwegotthelabelsbackwards.Dropsineffectivenessseemcorrelatedwithimbalancesintheinitialrandomsplit,asweseewithtrial6ontheleftandtrial9ontheright.24 LING287/CS424P,Stanford(Potts)SentimentlexiconsReferencesAndreevskaia,AlinaandSabineBergler.2006.MiningWordNetforafuzzysentiment:SentimenttagextractionfromWordNetglosses.InProceedingsoftheEuropeanChapteroftheAssociationforComputationalLinguistics(EACL).Baccianella,Stefano;AndreaEsuli;andFabrizioSebastiani.2010.SentiWordNet3.0:Anen-hancedlexicalresourceforsentimentanalysisandopinionmining.InProceedingsoftheSeventhConferenceonInternationalLanguageResourcesandEvaluation,22002204.EuropeanLanguageResourcesAssociation.Blair-Goldensohn,Sasha;KerryHannan;RyanMcDonald;TylerNeylon;GeorgeA.Reis;andJeffReynar.2008.Buildingasentimentsummarizerforlocalservicereviews.InWWWWorkshoponNLPintheInformationExplosionEra(NLPIX).Beijing,China.Brants,ThorstenandAlexFranz.2006.Web1T5-gramversion1.LinguisticDataConsortium,Philadelphia.Cerini,S.;V.Compagnoni;A.Demontis;M.Formentelli;andG.Gandini.2007.Micro-WNOp:Agoldstandardfortheevaluationofautomaticallycompiledlexicalresourcesforopinionmin-ing.InAndreaSansò,ed.,LanguageResourcesandLinguisticTheory:Typology,SecondLanguageAcquisition,EnglishLinguistics.Milan:FrancoAngeliEditore.Esuli,AndreaandFabrizioSebastiani.2006.SentiWordNet:Apubliclyavailablelexicalresourceforopinionmining.InProceedingsofthe5thConferenceonLanguageResourcesandEvaluation,417422.Genova.Fellbaum,Christiane,ed.1998.WordNet:AnElectronicDatabase.Cambridge,MA:MITPress.Godbole,Namrata;ManjunathSrinivasaiah;andStevenSkiena.2007.Large-scalesentimentanal-ysisfornewsandblogs.InProceedingsoftheInternationalConferenceonWeblogsandSocialMedia.Hatzivassiloglou,VasileiosandKathleenR.McKeown.1993.Towardstheautomaticidenticationofadjectivalscales:Clusteringadjectivesaccordingtomeaning.InProceedingsofthe31stAnnualMeetingoftheAssociationforComputationalLinguistics,172182.ACL.Hatzivassiloglou,VasileiosandKathleenR.McKeown.1997.Predictingthesemanticorientationofadjectives.InProceedingsofthe35thAnnualMeetingoftheACLandthe8thConferenceoftheEuropeanChapteroftheACL,174181.ACL.Hu,MinqingandBingLiu.2004.Miningandsummarizingcustomerreviews.InProceedingsofthe10thACMSIGKDDInternationalConferenceonKnowledgeDiscoveryandDataMining,168177.ACL.Kim,Soo-MinandEduardHHovy.2006.Identifyingandanalyzingjudgmentopinions.InRobertC.Moore;JeffA.Bilmes;JenniferChu-Carroll;andMarkSanderson,eds.,ProceedingsoftheHu-manLanguageTechnologyConferenceoftheNorthAmericanChapteroftheAssociationofCompu-tationalLinguistics,200207.ACL.26