UCLA Computer Science Department Technical Report - PDF document

Presentation on theme: " UCLA Computer Science Department Technical Report"— Presentation transcript

UCLAComputerScienceDepartmentTechnicalReport#100025 Theremainderofthispaperisorganizedasfollows.WediscussrelatedworkinSection2andprovideanoverviewofourgoalsandapproachinSection3.Section4describesoursystemforextractingkeywordsfromcontent-basedtextsources,andinSection5wepresentmethodsforexpandingthosetermstoaddressvocabularymismatchproblemsbetweenthesourcekeywordsandthosechosenbyadvertisers.WeevaluatethekeywordsgeneratedbythesemethodsforvariousvideotypesandsourcesoftextinSection6,anddiscussourobservations,conclusions,andfutureworkinSection7.2RelatedWorkSponsoredsearch,oradvertisingdisplayedalongsidethesearchresultsofauser-suppliedkeywordquery,typicallyinvolvesacomplexcombinationofadvertisersbiddingonkeywords,reviewofadvertisementsforrelevance,andanauctionprocesstoplaceadsalongsidesearchresults.See[2]foranoverviewofsponsoredsearch.Indisplayorcontent-matchadvertising,however,explicitkeywordsforthecontentarenotpro-vided.Inonlineadvertisingitisimportanttodisplayadsrelevanttoapage'scontent[24].Withoutuser-suppliedkeywords,researchershaveinvestigatednumerouskeywordidenticationtechniquesandapproachestomatchadvertisementswiththecontentofWebpages.Ontologiesortaxonomieshavebeenusedincombinationwithfeatureidenticationforsemanticapproachestomatchingadvertisementswithcontent[5,7].Ontologiesareoftendomain-specicandtedioustoconstruct,andtheindividualtextelementsfromscriptsordialogareoftenterse,makingtheuseofclassicationtechniquesorontologiesmoreerrorprone.Yih,Goodman,andCarvalho[25]establishseveralfeaturesofWebpagesandquerylogs,suchasfrequency,textualcharacteristics(e.g.capitalization),andstructuralcuesforidentifyingadvertisingkeywords.Ribeiro-Netoetal.[18]proposestrategiesformatchingthetextofaWebpagewithtext-basedadvertisementsinaknownadinventory.Theyaddressthevocabularyimpedanceproblembyrepresentingapagewithconceptsfromitsnearest(mostsimilar)neighbors.Ravietal.[17]proposeatwophasegenerativemodelforidentifyingrelevantadvertisingkeywordsforagivenWebpage.TheyuseapopularmachinetranslationmethodtolearnaprobabilisticsetofkeywordmappingsfromatrainingcorpusofWebpagesassociatedwithadsandadvertiserchosenkeywords.TermweightsareassignedbasedonHTMLfeatures.Abigramlanguagemodeltrainedonsearchqueriesisusedtohelprankthegeneratedcandidatekeywords.Findingrelatedbut\lessobvious"(andthereforelessexpensive)keywordsfromanadvertiserspointofview[11,1]hasbeenaddressedaswell.Ourworkdiersfromtheseproblemsinseveralways.OurtextsourcesareplaintextwhichlackexplicitstructuralcuessuchasHTMLmarkup.Wethereforemustresorttostatisticalmethodsforrankingandselectingkeywordsfromthesourcetext.Severaloftheabovetechniquesalsorequiretaggedtrainingdata,languageanddomain-specicontologies,orpre-constructedpoolsofavailableadvertisements.Ourmethodsarelanguage-independentandunsupervised,notrequiringanytrainingdata.Relatedtermidenticationisawellresearchedproblemintheinformationretrievaldomain,wheretaskssuchasqueryrewritingorexpansionarewidelystudied.Voorhees[23]describedtheuseoflexicalrelationshipscontainedinWordNetforqueryexpansion.Buckleyetal.[6]notedthatrelatedtermswilltypicallyco-occurnon-randomlyindocumentsrelevanttoaquery.Morerecently,SahamiandHeilman[19]useasimilarnotiontocomputethesemanticsimilarityofshorttextsnippetsusingWebsearchresultsasanopaquecontext.OurworkfollowsalongtheselinesbyusingWebsearchresultstodiscoverrelatedco-occurringterms.WealsousetheimplicitsemanticrelationshipscapturedinthehyperlinkedstructureofWikipediatoidentifyrelatedterms.Hauptmannsummarizesmany\lessonslearned"regardingspeechrecognitionaccuracyandtheeectsofworderrorrateoninformationretrievalprecision[9].Inparticulartheirresearchshowsthatthebestsystemsachieveworderrorratesaround0.15underidealconditions(suchasin-studioanchorsforbroadcastnews),andthatretrievalperformancedegradesrelativelygracefullywithrespecttoperfecttexttranscriptsuntilworderrorratesapproach0.40.Whiletheirworkonretrievalisorthogonaltoourfocusofadvertisingkeywordselection,weconsiderthesendingsaswecompareourresultsbetweenperfecttexttranscripts(closedcaptioningtracks)andspeechtranscripts.Keywordidenticationformultimediaoftenutilizes,inpart,attributesextractedfromimagesaspartofalargerfeaturespaceformachinelearning.VelivelliandHuang[22]predicttagsforvideosbasedonimagefeaturesandspeechtranscripts.Usingacollectionofspeechtranscripts,theyperformaPLSI-basedclusteringtoformktopicthemes.Eachclusterisusedtogenerateaunigramlanguagemodeli,andasceneisassumedtobeamixtureofthesemodelsandanunderlyingbasemodel.Tagsarethenpredictedusingacombinationofshotfeaturesandkeywordco-occurrencebasedonaconstructedtrainingset.2 UCLAComputerScienceDepartmentTechnicalReport#100025 Figure1:ScriptProcessingWork owthisstage,ourgoalistoincreasethelikelihoodofmatchinganadvertiser'skeywordswhileminimizingdeclineinrelevancyoftheadswhenmatchesdooccur.ThisrelatedtermminingprocessisdescribedinSection5.Inbothsteps,keywordsareidentiedandrankedwithoutconsultinganinventoryofadsoradvertisersuppliedkeywords.4ProcessingSourceTextIntherststageofprocessing,weanalyzetheformatandcomplexitiesofvideo-basedtextsources,suchasscripts,anddescribemethodsoftextanalysisbasedontraditionalstatisticalanalysisandgenerativemodels.Inthisworkweconsiderthreesourcesoftextdataforavideo:MovieScript-ascriptorscreenplayisadocumentthatoutlinesallofthevisual,audio,behavioral,andspokenelementsrequiredtotellastory.Sincelmproductionisahighlycollaborativemedium,thedirector,cast,editors,andproductioncrewwillusevariousformsofthescripttointerprettheunderlyingstoryduringtheproductionlmingprocess.Numerousindividualsareinvolvedinthemakingofalm,thereforeascriptmustconformtospecicstandardsandconventionsthatallinvolvedpartiesunderstandandthuswilluseaspecicformatwithrespecttothelayout,margins,notation,andotherproductionconventions.Thisdocumentisintendedtostructureallofthescriptelementsusedinascreenplay.ClosedCaptioning(CC)track-adocumentwhichcontainsaseriesoftimecodesandtextofthespokendialog.Eachtimecodeindicateswhenandforwhatdurationthecorrespondingtextappearsonscreen.Closedcaptioningtrackslackadditionalcues,suchasvisualinformationorindicatorsofthecurrentspeaker.Speech-To-Text(STT)-isaprocessbywhichaudiodatacontainingdialogornarrativecontentisautomaticallyconvertedtoatexttranscription.Theoutputtypicallyconsistsofaseriesofwords,eachwithanassociatedtimecodeandduration.Thesourceaudiomaybeofpoorqualityorcontainnon-speechsoundssuchasmusicorsoundeectartifacts,whichgenerallycontributetotranscriptionerrors.TranscriptionqualityistypicallymeasuredbytheoverallWordErrorRate(WER).AfrequentgoalofSTTsystemsistoreducetheimpactofahighWER,thougherrorratesonheterogeneouscontentistypicallyquitehigh.Figure1outlinestheprocessingwork owforacompletemoviescript,whichincludesnon-speechelementssuchassceneheadingsandactiondescriptions.Wewilldescribeeachofthesestepsnext.Notethatthework owislargelythesameforclosedcaptioningtracksandspeechtranscripts,whichcanbeformattedto\look"likeascreenplay.Inthosecases,script-specicprocessingstepsaresimplyomitted.4.1ScriptParsingTelevisionandmoviescriptsarefrequentlywritteninplaintextandfollowaconventional\screenplay"formatwhichallowshumanreaderstoeasilydierentiateandinferthepropersemanticsfordierentscriptelements,suchasdialogorsceneheadings.Forexample,sceneheadingsaretypicallywrittenonasinglelineinallcapitalletters,beginningwithINTorEXTtodenotewhetherthesettingisinteriororexterior,andendingwithanindicatoroftimeofdaysuchasMORNINGorNIGHT.Figure2showsabriefsnippetofatypicalscript.Understandingthesemanticsofatextelementishelpfulwhenprocessingit.Forexample,characternamesappearfrequentlyinascriptpriortoeachoftheirlinesofdialog,thoughwegenerallyndthemtobeapoorchoiceforadvertisingkeywords.Weaddamachine-readablehierarchicalstructure4 UCLAComputerScienceDepartmentTechnicalReport#100025 Figure2:ExampleScriptSnippetandsemanticstoeachtextsegmentofascriptusinganitestatemachinebasedparserderivedfromconventionalscreenplaywritingrules.Thisisdepictedasstep(1)inFigure1.Moviescriptdocumentsareconvertedintoastructuredandtaggedrepresentationwhereallscriptelements(sceneheadings,actiondescriptions,dialoglines,etc.)aresystematicallyextracted,tagged,andrecordedasobjectsintoaspecializeddocumentobjectmodel(DOM)forsubsequentprocessing.AllobjectswithintheDOM(e.g.,entiresentencestaggedbytheircorrespondingtypeandscriptsection)arethenprocessedusingbothstatisticalmethodstoidentifykeywordsofinterest,andanaturallanguageprocessing(NLP)enginethatidentiesandtagsthenounitemsidentiedineachsentence.Theseextractedandtaggednounelementsarethencombinedwithtime-alignmentinformationandrecordedintoametadatarepository.Wedescribethisalignmentprocessnext.4.2Speech-to-Text(STT)ProcessingSTTtranscriptscontaintimecodeinformationthatplaysanimportantroleinassociatingscriptkeywordstospecicpointsintimeinthevideocontent.Inthissectionofthework ow,avideooraudiolethatcontainsspokendialogthatcorrespondstothedialogsectionsoftheinputscriptisreadandprocessedusingaSpeech-to-Textenginethatgeneratesatranscriptionofthespokendialog,shownas(2)inFigure1.Forthisprocess,wealsoperformanimportantoptimization.Automaticspeechrecognitionenginestypicallyincorporateaknownvocabularyandprobabilisticmodelsofspeech(oftenbasedonwordN-grams).Whenthedialogdataisavailablefromascript,weconstructacustomlanguagemodeltobiasthetranscriptionenginetowardstheexpectedvocabularyandwordsequences,whichhelpstoincreasethetranscriptionaccuracy.4.3ScriptandSTTTranscriptAlignmentAtthisstage,wehavetaggedandstructuredscriptdata(withoutanytimeinformation)fromstep(1),andanoisy,relativelyinaccurateSTTtranscriptwithveryprecisetimecodeinformationfromstep(2).Tomakeuseofthekeywordsandconceptsgeneratedbythelaterprocessingsteps,thescriptdatamustbetime-alignedwiththeSTTdata.Thisisaccomplishedinstep(3)byusingtheLevenshteinWordEditDistance[12]algorithmtondthebestwordalignmentbetweenscriptdialogandtheSTTtranscript.Theresultofthisphaseofprocessingisatime-alignedsourcescriptthatcanassociatescriptactionanddialogkeywordswithprecisepointsintimewithinthevideocontent.Thisdataisstoredintoametadatarepository.4.4StatisticalGenerationofKeywordTermsInthenalstepforasourcetext(script,CC,ortranscript),thetime-codedtextelementsfromthemetadatarepositoryareusedtobuildasuxwordN-gramtreethatisprunedbyN-gramtermfrequencytodiscoverthemostdominantterms,basedinlargepartontheworkofChimandDeng[8].Thisisshownas(5)inFigure1.BeforeN-gramtermgeneration,weperformedaone-timeprocessofselectingastopwordvocabularyspecictothedomainofmoviescripts.Usingfrequencystatisticscomputedfromalargecorpusofscripts,wemanuallyidentiedasetofstopwordsfromthemostfrequentlyoccurringterms.DuringN-gramtermgeneration,thefollowingstepsarefollowed:1.Corpusstopwordsareremovedfromthesourcetext.2.AnN-gramtermtreewithsequencesuptolengthN=4iscreatedbycollectingandcountingN-gramoccurrencesfromthescript.5 UCLAComputerScienceDepartmentTechnicalReport#100025 3.TheresultingsuxtreeisthenprunedbytraversingthetreetocollectandrankthetopmostMfrequentterms.Inourexperiments,weselectthemostfrequentM=20keywords.4.5GenerativeModelsForNoisyDataThestatisticalN-grammethodsworkwellwhenkeywordsandphrasesarerepeatedmultipletimes.Whilethisisoftenthecaseforlongerorwell-formedtextinput,shortornoisytextoftenresultsinthemajorityof(non-stopword)keywordsonlybeingmentionedonce.Withthistypeofinput,statisticalmodelsareunabletodecipherwhichkeywordsaremostimportant.Tobetterhandleshortornoisytextinput,weuseakeywordselectionmethodbasedongenerativetopicmodeling.Inthismodel,weassumethatavideocomprisesasmallnumberofhiddentopics,whichcanberepresentedaskeywordprobabilities,andthatavideo'stextisgeneratedfromsomedistributionoverthosetopics.Thehighlyprobablekeywordsinthosetopicsarelikelytobemostrepresentativeofthevideocontent.WeuseLatentDirichletAllocation(LDA)[4]tolearnthetopicsandcorrespondingtopic-keywordprobabilitydistributionfromtheinputtext.Wethencombinethesetopicstoformarankedkeywordlist.4.5.1GeneratingTopicsTodiscovertheunderlyingtopicsinavideo,wesegmenttheinputtextintosentencesandperformtopicmodelingwithLDA.Theresultingtopic-termdistributionisaKxVmatrix,whereKisthenumberoftopics,Visthesizeoftheinputvocabulary,and[i][j]istheprobabilityofkeywordjintopici.Weformanorderedlistofkeywordskiforeachtopic,sortedbytheirprobabilityin~[i].ThisresultsinKrankedlistsofkeywords,onepertopic,whichmustthenbemergedintoasinglelisttoselectthetopM.WhilesimplyselectingthetopM Kkeywordsfromeachtopicisoneoption,wedescribeamoregeneralsolutionformergingmultiplerankedlistswhenwediscussourapproachtondingrelatedkeywords.ThismethodisdescribedinSection5.3.4.6Statistical-GenerativeHybridMethodTheLDAmodellearnskeywordprobabilitiesfortermswhichareseparatedbywhitespace.Whenpos-sible,however,itispreferabletoidentifymulti-termkeywordsforadvertising.Forexample,thephrase\relationaldatabase"ismorespecicthaneitheroftheindividualwords\relational"or\database",andthushashighervaluetoadvertisers.Tohelpidentifythesemulti-tokenkeywordsinshortornoisytextsources,weuseahybridofstatisticalandgenerativetechniques.WerstprocessthesourcetextusingtheN-grammethodtoidentifyanysignicantmulti-tokenkeywords.Wetheneditthesourcetextbyremovingthewhitespacebetweenthetermsofthesemulti-tokenphrasessotheyappearasasingletoken.Thismodiedsourcetextisthenprocessedusingthegenerativemodel.4.7FilteringtheKeywordsWeapplytwolters,whenpossible,toremovefrequentlyoccurringwordswhichareoftennotusefulinthecontextofmatchingadvertisements.Fromallinputsources,keywordsmatchingalistofEnglishprofanityareremoved.Wealsondthatmaincharacternamesareoftenamongstthetopkeywords,butgenerallydonotretrieverelevantadvertisements.Whengivenacompletescript,weremovecharacternamesfromthekeywordlistusingadictionaryconstructedduringtheparsingandtaggingstage.Forclosedcaptioningandspeechtranscripts,however,thesenamesareunknownandthusmaystillappearinthetopkeywords.Thisismorecommonforclosedcaptioningthanspeechtranscripts,however,aspropernamesarelesslikelytobecorrectlytranscribedbytheSTTengine.Atthispointthemostdominant(possiblymulti-term)keywordswhichoccurinthesourcetext,alongwithassociatedtimecodeinformation,havebeenidentiedandcanbesuggestedasadvertisingkeywordsrelevanttoaparticulartimepointofavideo.AsRibeiro-Netoetal.[18]describe,however,thekeywordschosendirectlyfromasourceandthekeywordsbidonbyadvertisersmaysueravocabularyimpedanceproblem.Inthenextsectionwedescribenoveltermminingtechniqueswhichcanprovidearicher,morecompletesetofrelevantadvertisingkeywords.6 UCLAComputerScienceDepartmentTechnicalReport#100025 Figure3:RelatedTermsFromSearchResultsAftertheselteringsteps,weconstructavectorspacemodelMforthissmallcorpusof\documents"relevanttoT.BasedonthepopularTF-IDF[20]termweighting,wecomputethecorpusfrequency(CF)andinverse-document-frequency(IDF)weightforeachterminM,andrankthekeywordsaccordingtotheirCF*IDFscore.Thisstepisshowninthework owas(3)inFigure3,producingthenallistofrankedrelatedkeywordsfromsearchresults.5.2MiningwithWikipediaTheseconddatasourceweanalyzeforrelatedtermsisWikipedia,anextensiveknowledgebasewithover3.1millionEnglisharticlesavailableatthetimeofthiswriting.WhereasintheWebcorpuswefocusedonsearchresultsinresponsetoaquery,withWikipediawedirectourattentiontohyperlinks.WithinthetextofaWikipediaarticle,numerousinter-wikilinkspointtootherWikipediapages,whichallowsustomodelWikipediaasadirectedgraphG=fV;Eg.WeconstructtheWikipediagraphwherenodesVrepresentpagesinthemainarticlenamespace,andedgesEdenoteinter-wikilinksbetweenthosepages.Whenbuildingthegraph,twoarticletypesinthemainnamespaceareprocessedspecially.Forambiguoustermssuchas\coach",adisambiguationpageinWikipedialiststheavailablearticlesfordierentsensesoftheterm.Thesepagesserveprimarilyasnavigationalaidesforusers,ratherthanconveyingasemanticrelationshipbetweenterms,andwethere-foreexcludetheminthegraph.Thesecondcategoryofpagesweprocessspeciallyareredirectionpages,whichprovideatranslationforalternateormisspelledwords,inconsistentcapitalization,acronyms,andsoon,intoacanonicalform.InourWikipediagraph,anarticleandallofthepageswhichredirecttoitaremergedintoasinglenode.Weusethelinkstructureofthegraphtobothidentifyandrankcandidaterelatedterms.Thesestepsaredescribedindetailinthefollowingsections.5.2.1IdentifyingCandidateRelatedTermsWithoutclearlydeneddirectedlinksbetweenindividualtermsintheWebcorpus,theapproachesusingWebsearchresultsdescribedabovedependontheassumptionsthatdocumentsretrievedbythesearchenginearerelevanttotheinputterms,andthatothertagsorkeywordsforthosepagesarepotentiallyrelated.Thatis,werelyonco-occurrencebasedmeasurestoidentifywhichtermsaremostlikelyrelated.WithWikipedia,however,wehaveanexplicitlinkstructurebetweenarticleswhichcanbeusedasanindicatorofrelatedness.Werequiretherelatednessbetweentwoarticlenodesaandbtobeasymmetricrelationship:aisrelatedtobifandonlyifbisrelatedtoa.TranslatingthisrequirementtotheWikipediagraphisrelativelystraightforward.Toidentifycandi-daterelatedtermsfortermT,werstlocatetheWikipediapagewithTasthetitle.3GiventhenodetforT,weidentifyanynodesinthegraphwhichformadirectcyclewithtascandidaterelatedterms.Thatis,weselectthesubsetofnodesNVsuchthat:8n2V;n2N=)ft;ng^fn;tg2E(1)Figure4showsasimpleexample,wherefortermt,termsn1andn2arecandidaterelatedterms,butXandYarenot.5.2.2RankingCandidateTermsAgivensetofcandidaterelatedtermsmaybequitelarge.Wenowlookathowtorankthecandidateterms.Tobeagoodsuggestionasanadvertisingkeyword,atermshouldberelativelypopular.Whilewecouldmeasurepopularitythroughexternalsources,suchasquerylogfrequency,wechosetoutilizethegraphstructureofWikipedia.Weapproximatetherelativeimportanceoftermsbycomputing 3WecanrelaxthisrequirementandsearchthetextofWikipediaarticlestoidentifythetoppageorpagesforanyparticularinputterm,albeitatalikelyreductioninqualityofthegeneratedrelatedterms.8 UCLAComputerScienceDepartmentTechnicalReport#100025 SR(camera) WP(camera) Combined SR(advertising) WP(advertising) Combined digitalcamera photography digitalcamera product internet internet lens pornography photography marketing newspaper product canon visualarts canon advertiser videogame marketing nikon photograph nikon business americanfootball newspaper zoom digitalcamera pornography campaign magazine advertiser lmcamera photojournalism lens advertisingagency worldwideweb magazine digitalslr photographiclm digitalphotography internet marketing advertisingagency megapixels aperture photograph consumer mtv publicrelations digitalphotography canon aperture job blog google compact photographiclens shutterspeed newspaper publicbroadcastingservice billboard camcorder aerialphotography visualarts agency massmedia videogame slrcamera holography exposure publicrelations google publicity lense single-lensre excamera viewnder company brand productplacement digitalslrcamera focallength moviecamera service broadcasting graphicdesign olympus nikon cameraphone budget musicvideo promotion Table2:ExampleRelatedTermsbyMethod6.1EvaluationDesignWeidentiedthetop20keywordsfromeachavailabletextsourceusingboththestatisticalandhybridapproachesdescribedinSection4.ForeachofthesekeywordsweusetherelatedtermminingtechniquesofSection5toidentifythetop10relatedterms.Thesekeywordswerethenevaluatedwithausersurvey.Forthetopicmodelingphaseofthehybridtechnique,wesetthenumberoftopicsK=5withtheLDAparameters=0:3and=0:1.Userswereshownavideoclip,typicallyaround3minutesinlength,andasetofkeywords.Tokeepthesizeofthekeywordsetmanageable,weshow5ofthetop20keywordsforeachmethodfromeachavailabletextsource,and1ofthetop10relatedtermsforeachofthosekeywords,allchosenandorderedatrandom.Userswereaskedtomakeabinaryassessmentontherelevanceofeachdisplayedkeyword.ForthenewsandeducationalvideosandtheamateurclipsavailableonYouTube,usersareshownthecompletevideo.Forfulllengthlms,usersareshownthetheatricaltrailerandaskedtomakejudgementsbasedonthetrailerandtheirpriorknowledgeofthemovie.Over23peopleparticipatedinthesurvey(personallyidentiableinformationwasnotrequired),withaminimumof9andanaverageof13usersevaluatingeachvideo.6.2EvaluationMetricsWeevaluatethekeywordsgeneratedbyourmethodsusingfourmetrics.Theaveragerelevancyofthekeywordsdisplayedtouserswecalltheprecision.Multipleusersviewingthesamesetofkeywordsmaynotcompletelyagreeonwhichkeywordsarerelevant.Wethereforecomputethepotentialofasource,whichmeasuresthefractionofthekeywordsjudgedrelevantbyatleastoneuser.Moreformally,wedenetheprecisionandpotentialoftextsourceSas:Precision(S)=1 iXijKi(S)\Rij jKi(S)jPotential(S)=jR(S)j jK(S)jRiisthesetofkeywordsjudgedrelevantinevaluationiandKi(S)arethekeywordsdisplayedtotheuserforevaluationiwhichcomefromsourceS.K(S)arethekeywordsfromsourceSdisplayedinatleastoneevaluation,andR(S)arethekeywordsfromsourceSjudgedrelevantbyatleastoneuser,denedas:R(S)=[iKi(S)\RiK(S)=[iKi(S)Theothermetricswedeneareappealandpopularity,whichserveasindicatorsofhowpertinentthekeywordsaretoadvertisers.Appealestimatesthelikelihoodthatakeyworddeemedrelevanttothecontentwillalsobemeaningfultoanadvertiser.Popularitymeasurestheaveragenumberofadvertisersinterestedinarelevantkeyword.WedenetheappealandpopularityofasourceSas:Appeal(S)=jR(S)\Aj jR(S)j10 UCLAComputerScienceDepartmentTechnicalReport#100025 VideoType Precision Potential Statistical Hybrid Statistical Hybrid StudioFilms 0.268 0.252 0.479 0.480 News/Educational 0.442 0.473 0.548 0.717 UserGenerated 0.268 0.368 0.390 0.473 Table4:PrecisionandPotentialforSTT VideoType WER Statistical Hybrid StudioFilms 0.857 0.723 0.690 News/Educational 0.406 0.731 0.961 Table5:RelativePrecisionandWordErrorRateHauptmann'sworkindicatesthatspeech-to-textworderrorratesunder0.4resultinretrievalper-formancecomperabletoaperfecttranscript[9].Atthe0.4threshold,relativeretrievalprecisionisapproximately80%.Wecomputetheaverageworderrorrateforstudiolmsandnews/educationalvideos(usingthedefault\general"languagemodelsforourSTTengine),andcomparetherelativepre-cisionofSTTwithrespecttoclosedcaptioningforthestatisticalandhybridmethods,showninTable5.Usergeneratedvideosarenotincludedbecauseno\correct"transcriptsareavailableforthecontent.Asexpected,theaverageworderrorratesfornewsandeducationalvideosaresubstantiallylower,thoughstillaround0.4.Forthistypeofcontent,therelativeprecisionofSTTis96%oftheclosedcaptioning.Forthehigherworderrorrateoflmswecanstillachieveover70%averagerelativeprecision.Theseresultsfurthersupportuseofthestatisticalselectionmethodsonlongertextinputsandthegenerativemethodsonshortertext,andsuggestthatspeechtranscriptsalonemaybesucienttondmeaningfuladvertisingkeywordsforvideossuchasnewsbroadcasts.6.5PrecisionandPotentialofRelatedTermsWenextlookattheprecisionandpotentialoftherelatedterms.Table6showstheprecisionandpotentialscoresforthetop10relatedtermsfromboththestatistical(S-Related)andhybrid(H-Related)methods.TheseresultsaremostlyconsistentwithTable3,withthemostpreciseinputsource(closedcaptioning)producingthemostrelevantrelatedkeywords.Foreachmethodandsource,theprecisionandpotentialofthesourcekeywordsarehigherthantherelatedterms.InourexperimentswerandomlyselectedfromthetopN=10relatedtermsforeachsourcekeyword.Wenowinvestigatehowtheaverageprecisionoftherelatedtermsisaectedaswevarythisrangefor1N10.Figure5plotstheprecisionoftherelatedkeywordsforeachtextsourceusingthestatisticalselectionmethod.Forclosedcaptioningthetop2relatedtermsgivethehighestprecision,whichislowerthantheprecisionofthesourcetermsbutsignicantlyhigher(p=0:003)thanchoosingfromthetop10.Bothscriptandspeechtranscriptinputsshowanincreaseinprecisionwhenselectingfromthetop3-6terms.Whiletheprecisionisagainlowerthanthesourcekeywords,thereisnoticeableimprovementbetweenselectingfromthetopN=6andN=10forbothscript(p=0:06)andSTT(p=0:03)input.Thisresultsuggeststhatthenumberofrelatedtermstoconsidertoachievethemaximumoverallprecisiondependsontheinputtexttype,withhigherprecisioninputlikeclosedcaptioningachievingitsbestprecisionwithasmallernumberrelatedtermsthanscriptsorspeechtranscripts.Resultsforthehybridselectionmethodexhibitsimilarbehavior.Anotherfactortoconsiderwhenevaluatingtheprecisionoftherelatedkeywordsistherelevancyofthesourcetermbeingexpanded.Anirrelevantsourcetermislesslikelytoresultinrelevantrelated Source Precision Potential S-Related H-Related S-Related H-Related Script 0.254 0.215 0.253 0.222 CC 0.260 0.221 0.262 0.221 STT 0.208 0.186 0.200 0.191 Table6:PrecisionandPotentialofRelatedTerms12 UCLAComputerScienceDepartmentTechnicalReport#100025 Source Statistical S-Related Hybrid H-Related Script 0.726 0.788 0.607 0.792 CC 0.578 0.785 0.543 0.796 STT 0.681 0.827 0.594 0.820 Table7:AppealofKeywordsbySource Source Statistical S-Related Hybrid H-Related Script 3.59 3.96 3.00 4.18 CC 2.11 3.81 2.00 3.77 STT 2.54 4.39 2.56 4.30 Table8:PopularityofKeywordsbySourceForbothappealandpopularitywenoticethat,whileclosedcaptioningwasgenerallyconsideredthemostprecisesourceofkeywords,italsoproducestheleastmeaningfulkeywordsforadvertisers.Thismaybearesultofcharacternamesappearingintheclosedcaptioningkeywords,whichwenotedearlierarelteredoutfromscriptinputtextandarelesslikelytoretrieverelevantads.Finally,welookcloseratthepopularityofkeywordsforspeechtranscripts.Table9comparesthepopularityforsourceandrelatedkeywordsforvariousvideotypes.Inallcasestherelatedkeywordshavehigherpopularitythanthesourcekeywordsbyastatisticallysignicantmargin.Italsoshowsthatnewsandeducationalcontentcontainslesspopularkeywordsforadvertisers.6.7Precision-PopularityTradeosTheresultsabovedemonstratethat,whenrelevant,relatedkeywordsaresignicantlymoreattractivetoadvertisersthansourcekeywords.Theoverallprecisionoftherelatedterms,however,islowerthansourceterms.Weexploretheinherenttradeobetweenkeywordrelevanceandpopularitybycomputingaprecision-weightedpopularitymetric:PWP(S)=Pk2K(S)AkP(S;k) jK(S)j(4)WhereP(S;k)istheprecisionofkeywordkfromsourceS,denedas:P(S;k)=Pijfkg\Ri(S)j Pijfkg\Ki(S)jTable10showstheprecision-weightedpopularityforthestatisticalmethodforeachtextsourceusingthetop5relatedkeywordsfromeachsourcekeyword.Theresultssuggeststhatforscriptinput,theminorimprovementinpopularityofrelatedkeywords(showninTable8)maynotosetthedecreaseinprecision.Forspeechtranscriptinput,however,thereappearstobesomebenetfromrelatedterms.WeexamineSTTinputfurtherinTable11,whichshowsthatoverall,evenwiththedropinprecision,relatedkeywordsarebenecialtoadvertisersfornewsandusergeneratedvideoswhenonlyspeechtranscriptsareavailable.Althoughtherelatedkeywordsforstudiolmspeechtranscriptshavehigherpopularitythansourcekeywords(Table9),therelativeincreaseisnoticeablylowerthanforCCorSTT,andtheresultingprecision-weightedpopularitydoesnotoerimprovement. Source Statistical S-Related Hybrid H-Related StudioFilms 2.97 4.35 2.67 4.39 News/Educational 1.69 4.11 2.21 3.50 UserGenerated 1.89 4.83 2.63 4.75 Table9:PopularityforSpeechTranscripts14 UCLAComputerScienceDepartmentTechnicalReport#100025 [4]D.M.Blei,A.Y.Ng,M.I.Jordan,andJ.Laerty.Latentdirichletallocation.JournalofMachineLearningResearch,3,2003.[5]A.Broder,M.Fontoura,V.Josifovski,andL.Riedel.Asemanticapproachtocontextualadvertising.InSIGIR'07,pages559{566,2007.[6]C.Buckley,G.Salton,J.Allan,andA.Singhal.AutomaticqueryexpansionusingSMART:TREC3.InTextREtrievalConference,1994.[7]Y.Chen,G.-R.Xue,andY.Yu.Advertisingkeywordsuggestionbasedonconcepthierarchy.InWSDM'08,pages251{260,2008.[8]H.ChimandX.Deng.Anewsuxtreesimilaritymeasurefordocumentclustering.InWWW'07,pages121{130,2007.[9]A.Hauptmann.Lessonsforthefuturefromadecadeofinformediavideoanalysisresearch.InCIVR'05,pages1{10,2005.[10]G.JehandJ.Widom.Simrank:ameasureofstructural-contextsimilarity.InKDD'02,pages538{543,2002.[11]A.JoshiandR.Motwani.Keywordgenerationforsearchengineadvertising.InICDMW'06,pages490{496,2006.[12]V.Levenshtein.Binarycodescapableofcorrectingdeletions,insertionsandreversals.InSovietPhysicsDoklady,1966.[13]H.Ma,H.Yang,I.King,andM.R.Lyu.Learninglatentsemanticrelationsfromclickthroughdataforquerysuggestion.InCIKM'08,pages709{718,2008.[14]C.D.Manning,P.Raghavan,andH.Schtze.IntroductiontoInformationRetrieval.CambridgeUniversityPress,NewYork,NY,USA,2008.[15]E.Moxley,T.Mei,X.-S.Hua,W.-Y.Ma,andB.Manjunath.Automaticvideoannotationthroughsearchandmining.InICME'08,pages685{688,2008.[16]L.Page,S.Brin,R.Motwani,andT.Winograd.Thepagerankcitationranking:Bringingordertotheweb.Technicalreport,StanfordDigitalLibraryTechnologiesProject,1998.[17]S.Ravi,A.Broder,E.Gabrilovich,V.Josifovski,S.Pandey,andB.Pang.Automaticgenerationofbidphrasesforonlineadvertising.InWSDM'10,pages341{350,2010.[18]B.Ribeiro-Neto,M.Cristo,P.B.Golgher,andE.SilvadeMoura.Impedancecouplingincontent-targetedadvertising.InSIGIR'05,pages496{503,2005.[19]M.SahamiandT.D.Heilman.Aweb-basedkernelfunctionformeasuringthesimilarityofshorttextsnippets.InWWW'06,pages377{386,2006.[20]G.SaltonandC.Buckley.Term-weightingapproachesinautomatictextretrieval.InInformationProcessingandManagement,pages513{523,1988.[21]S.Siersdorfer,J.SanPedro,andM.Sanderson.Automaticvideotaggingusingcontentredundancy.InSIGIR'09,pages395{402,2009.[22]A.VelivelliandT.S.Huang.Automaticvideoannotationbyminingspeechtranscripts.InCVPRW'06,page115,2006.[23]E.M.Voorhees.Queryexpansionusinglexical-semanticrelations.InSIGIR'94,pages61{69,1994.[24]C.Wang,P.Zhang,R.Choi,andM.D.Eredita.Understandingconsumersattitudetowardadver-tising.InEighthAmericasConf.onInformationSystem,pages1143{1148,2002.[25]W.T.Yih,J.Goodman,andV.R.Carvalho.Findingadvertisingkeywordsonwebpages.InWWW'06,pages213{222,2006.[26]S.Zanetti,L.Zelnik-Manor,andP.Perona.Awalkthroughtheweb'svideoclips.CVPRW'08,pages1{8,2008.16