Discourse Paerns in Customer Complaints Decep3on Dataset · • In e-commerce tasks, an analysis of...
Transcript of Discourse Paerns in Customer Complaints Decep3on Dataset · • In e-commerce tasks, an analysis of...
AnAnatomyofaLie:DiscoursePa,ernsinCustomerComplaintsDecep3onDataset
DinaPisarevskayaIns3tuteforSystemsAnalysisFRCCSCRAS,Moscow,Russia,
HigherSchoolofEconomics&OracleCorp{[email protected]}
}
Candecep3onbecomevisiblefromthediscoursestructureoftext?
Decep3onanddiscoursestructure• Alotofformsofhumanintellectualandcommunica3onac3vityareassociatedwith
certaindiscoursestructures.• RhetoricalStructureTheory(RST)isagoodmeanstoexpresscorrela3onbetweensuch
formofac3vityanditsrepresenta3oninhowassociatedthoughtsareorganizedintext.
• RhetoricStructureTheorypresentahierarchical,connectedstructureofatextasaDiscourseTree,withrhetoricalrela3onsbetweenthepartsofit.
• LogicalArgumenta3onneedsacertaincombina3onofrhetoricalrela3onsofElabora'on,Contrast,CauseandA1ribu'on.
• PersuasivenessreliesoncertainstructureslinkingElabora'on,A1ribu'onandCondi'on.
• Explana3onneedstorelyoncertainchainsofElabora'onrela3onsplusExplana'onandCause
• Arhetoricalagreementbetweenaques3onandananswerisbasedonspecificmappingsbetweentherhetoricalrela3onsofContrast,Cause,A1ribu'onandCondi'onbetweentheformerandthela,er.
• Discoursetreesturnedouttobehelpfultoformadialogueandtobuilddialoguefromtext,tobe,erunderstandthestructureoftexts.
Есливасобвесятнабазаре,тоневажно,будутэтояблоки,грушиилибананы.Важнокаксвамивзаимодействуетпродавец,каковаструктура
этоговзаимодействия
Acustomercomplaintanditsdiscoursetree
“Ihaveaccountswiththemforalmost10years,Ihatedittheircustomerservice!Worstoneever.Idon'tknowwhat'stheirproblems,I'mnotrecommendingtheirservicesandbankingtoanybody,Istoppedusingtheircreditcardsalready!TheonlyreasonIcan'tclosemyaccountswiththem,itcoulddropmycreditscore.Iwillnotclosemycreditcards,butI'mnotdefinitelyusingthemsotheycan'tmakemoneyfromonus!Ijusthadconversa3onwithasupervisorfromCaliforniacalledSteveheandhisrepresenta3vedidn'tevenunderstandmysitua3on,whichwasnotcommonatall,basicallydidn'twanttohelpme!”
Theauthorofthiscomplaintdoesnotprovideasingleargumentbackinguphisclaim.Andtheauthor’sstatementthathiscredithistorycanbenega3velyaffectedbyhisclosingan
accountisamisrepresenta3on.
SpeechActsareassociatedwithdecep3on Communica3vediscoursetree
fortheTheranosa,ackonWallStreetJournalacquisi3on“Itisnotunusualfordisgruntledandterminatedemployeesintheheavilyregulatedhealthcareindustrytofilecomplaintsinanefforttoretaliateagainstemployersfortermina'onofemployment.Regulatoryagencieshaveaprocessforevalua'ngcomplaints,manyofwhicharenotsubstan'ated.Theranostrustsitsregulatorstoproperlyinves'gateanycomplaints.”
Toshowthestructureofadecep3on,discourserela3onsarenecessarybutinsufficient,andspeechactsarenecessarybutinsufficientaswell.Fortheparagraphabove,weneedtoknowthediscoursestructureofinterac3onsbetweenagents,andwhatkindsofinterac3onstheyare.file(employee,complaint)iselaboratedbyretaliate(employee,employer),andevaluate(regula'on,complaints)iselaboratedbytrust(Theranos,regulators,complainants).
Aprofessionalmisrepresenta3on
InJanuary2017,asecretdossierwasleakedtothepress.IthadbeencompiledbyaformerBri'shintelligenceofficialandRussiaexpert,ChristopherSteele,whohadbeenpaidtoinves'gateMrTrump's'estoRussia.ThedossierallegedMoscowhadcompromisingmaterialonMrTrump,includingclaimshewasoncerecordedwithpros'tutesataMoscowhotelduringa2013tripforoneofhisMissUniversepageants.MrTrumpempha'callydeniesthis.ThefilepurportedtoshowfinancialandpersonallinksbetweenMrTrump,hisadvisersandMoscow.ItalsosuggestedtheKremlinhadcul'vatedMrTrumpforyearsbeforeheranforpresident.MrTrumpdismissedthedossier,arguingitscontentswerebasedlargelyonunnamedsources.ItwaslaterreportedthatMrSteele'sreportwasfundedasopposi'onresearchbytheClintoncampaignandDemocra'cNa'onalCommi1ee.FusionGPS,theWashington-basedfirmthatwashiredtocommissionthedossier,hadpreviouslybeenpaidviaaconserva'vewebsitetodigupdirtonMrTrump.Trump-Russialinkacquisi3on(BBC2018]).Foralong3meitwasunable
toconfirmtheclaim,sothestoryisrepeatedoverandoveragaintomaintainareaderexpecta3onthatitwouldbeinstan3atedoneday.Thereisneitherconfirma3onnorrejec3onthatthedossierexists,andthegoaloftheauthoristomaketheaudiencebelievethatsuchdossierdoesexistneitherprovidingevidencenormisrepresen3ngevents.Toachievethisgoal,theauthorcana,achanumberofhypothe3calstatementsabouttheexis3ngdossiertoavarietyofmentalstatestoimpressthereaderintheauthen3cityandvalidityofthetopic.
Ul3mateDecep3onDataset• The dataset of 2746 emotionally charged customer complaints includes
descriptions of problems customers experienced with banks. • Provides clear ground truth, based on available factual knowledge
about the financial domain. • Complaints reveal practices of banks during the financial crisis of 2007,
such as manipulating an order of transactions to charge a highest possible amount of non-sufficient fund fees
• 400 complaints are manually tagged with respect to the parameters related to argumentation and validity of text: perceived complaint validity, argumentation validity, presence of specific argumentation patterns, and detectable misrepresentation.
• Krippendorff’s alpha measure (for three annotators) is more that 80%. • The rest 2,346 complains were auto-tagged based on the model trained
on this 400 set. After that they have also been partially manually evaluated, so that the accuracy of auto tagging exceeds 75%.
• In e-commerce tasks, an analysis of online reviews reliability is needed. The dataset could be used in machine learning to detect if reviews or other online texts of similar genres are truthful or deceptive, based on their discourse features.
The dataset is available at: https://github.com/bgalitsky/relevance-based-on-parse-trees/blob/master/examples/ultimateDeceptionAutoTagged.csv.zip
The dataset is available at: https://github.com/bgalitsky/relevance-based-on-parse-trees/blob/master/examples/ultimateDeceptionAutoTagged.csv.zip
Decep3onRecogni3onPipeline• Communicative Discourse Tree Construction. Rhetorical Structure Theory
presents a hierarchical, connected structure of a text as a Discourse Tree (DT) with rhetorical relations between the parts of it. In communicative discourse trees (CDTs), the labels for communicative actions (VerbNet expressions for verbs) are added to the discourse tree edges, to show which speech acts are attached to which rhetoric relations; this structure helps to understand argumentation.
• Nearest Neighbor learning. To predict the label of the text, once the complete DT is built, one needs to compute its similarity with DTs for the positive class and verify that it is lower than similarity to the set of DTs for its negative class. Similarity between CDT's is defined by means of maximal common sub-DTs.
• SVM Tree Kernel learning. A DT can be represented by a vector of integer counts of each sub-tree type (without taking into account its ancestors). Tree Kernel builder tool was used.
For all texts, we use CDT-kernel learning approach. We combined Stanford NLP parsing, coreference resolution tool, entity extraction, CDT construction (based on automated discourse parser and Tree Kernel builder into one system.
FindingFakeNewsdatasets• Availabledatasetsdonotlooksuitable.Itishardforahuman
todifferen3atebetweengenuineandfakenews.• FakeNews:“Acompletelyfabricatedclaimorstorycreated
withaninten3ontodeceive,o�enforasecondarygain.”• Unfortunatelythefakenewsrecogni3ondomainisvery
poli3cal.VerybiasedintheWestbecausethecommunityisnotevenlydistributedbetweenliberalandconserva3ve
• Nevertheless,foragivenhumanexpertandassigntags,itisusuallypossibletofindadiscourse-levelcorrelate
Andofcoursethereisafamilyofmethodsthatclassifythesedatasets
“successfully”
Evalua3on
Fortheini3alandautoma3callyderiveddatasets,weshow(inbold)theaccuracies,averagingthrough5xcross-valida3on.Forthebo,omthreedatasets,wetestedthesameSVMTreeKernelmodeltrainedonourdataset.Na3veSVMLightassessmentisused.Wedemonstrateitsuniversality,showingitsapplicabilitytotextsofvariousnaturesuchasconsumerreviews.Forgenuinereviews,580casesofdecep3onweredetectedwhichwerefalseposi3ves,assumingthatreviewwritersdonotlie
Onecanseethatkeyword-basedandNaiveBayesclassifierperformjustslightlybe,erthanrandom,sincedecep3onmanifestsitselfatthediscourselevel,notthesyntac3cone.ThenweobservethatproceedingtomachinelearningofDTsdelivers8%gaininclassifica3onaccuracy.
Validating claims: trying to achieve the best of both worlds
10
Argumentminingcommunityextractargumentsfromtextbutdoesnotdoanythingwiththem
Logicalargumenta3oncommunityreasonsaboutargumentsbutdonotextractthemfromtext
Wearebuildingapipelinewhichextractsargumentsfromtextandvalidates(reasonsabout)them
+
Arguments validation pipeline architecture
11
Forming Defeasible Logic Program
12
WebuildanexampleofaDeLPforlegalreasoningaboutfactsextractedfromtext.Ajudgehearsanevic3oncaseandwantstomakeajudgmentonwhetherrentwasprovablypaid(deposited)ornot(denotedasrent_receipt).Aninputisatextwhereadefendantisexpressinghispoint.UnderlinedwordsformtheclauseinDeLP,andtheotherexpressionsformedthefacts:Thelandlordcontactedme,thetenant,andtherentwasrequested.However,IrefusedtherentsinceIdemandedrepairtobedone.Iremindedthelandlordaboutnecessaryrepairs,butthelandlordissuedthethree-dayno'ceconfirmingthattherentwasoverdue.Regre_ully,thepropertys'llstayedunrepaired.
Defeasible Logic Program
13
DefeasibleRulesPreparedInAdvancerent_receipt-<rent_deposit_transac'on.rent_deposit_transac'on-<contact_tenant.┐rent_deposit_transac'on-<contact_tenant,three_days_no'ce_is_issued.┐rent_deposit_transac'on-<rent_is_overdue.┐repair_is_done-<rent_refused,repair_is_done.repair_is_done-<rent_is_requested.┐rent_deposit_transac'on-<tenant_short_on_money,repair_is_done.┐repair_is_done-<repair_is_requested.┐repair_is_done-<rent_is_requested.┐repair_is_requested-<stay_unrepaired.┐repair_is_done-<stay_unrepaired.TargetClaimtobeAssessed?-rent_receiptClausesExtractedfromtextrepair_is_done-<rent_refused.Factsfromtextcontact_tenant.rent_is_requested.rent_refused.remind_about_repair.three_days_no'ce_is_issued.rent_is_overdue.stay_unrepaired.
Communicative Discourse Tree
for the complaint
14!
Dialectical Tree for the claim rent_receipt
15
Evaluation of the overall pipeline: Claim Validation
16
17
• Recallislowbecauseinthemajorityofcasestheinvalidityofclaimsisduetofactorsotherthanbeingself-defeated.
• Precisionisrela3velyhighsinceifalogicalflawinanargumentisestablished,mostlikelythewholeclaimisinvalidbecauseotherfactorsbesidesargumenta3on(suchasfalsefacts)contributeaswell.
• Ascomplexityofacomplaintanditsdiscoursetreegrows,F1firstimprovessincemorelogicaltermsareavailableandthengoesbackdownasthereisahigherchanceofareasoningerrorduetoanoisierinput.
Evaluation Results of the overall pipeline: Claim Validation
Conclusions• Weexploredwhetherfakeopinionatedtexthaveasimilarrhetoricalstructuretotextwith
decep3on,andgenuinereviewshavesimilarrhetoricstructuretotextswithoutdecep3on• Wefocusedoninterpreta3onofdiscoursetreesandexploredwaystorepresentthemina
formindica3veofadecep3onratherthanneutralenumera3onoffacts.• Wecameupwithdiscourse-levelcuesfordecep3on• Webuiltandevaluatedamachinelearningframeworkfordetec3ngdecep3onand
demonstrated70%performanceontheUl3mateDecep3ondataset• Nextsteps:manuallyannotatethewholedataset;runfurtherexperimentstoimprove
precisionindecep3ondetec3ontask;forevalua3onpurposes,applyourdetectortootherbusinesscommunica3ondatasetsfromtherealword
Mainapplica3onareaofdecep3on-freeandfakenews-freecontent:chatbots
Certaincategoryofpeopleisunabletolie.Discoursestructureofwhattheysayisveryrudimentary.Childrenwithau3smcanbespecificallytrainedtolie
20+patentsonapplica3onsofdiscoursetreesarefiled
OpenNLP.Similarity
OpenNLP.SVMTreeKernelLearner