AStudyofPoissonQueryGenerationModelfor
InformationRetrieval
QiaozhuMei,HuiFang,ChengxiangZhai
DepartmentofComputerScienceUniversityofIllinoisatUrbana-Champaign
Urbana,IL61801
{qmei2,hfang,czhai}@uiuc.edu
ABSTRACT
Manyvariantsoflanguagemodelshavebeenproposedforinformationretrieval.Mostexistingmodelsarebasedonmultinomialdistributionandwouldscoredocumentsbasedonquerylikelihoodcomputedbasedonaquerygenerationprobabilisticmodel.Inthispaper,weproposeandstudyanewfamilyofquerygenerationmodelsbasedonPoissondis-tribution.Weshowthatwhileintheirsimplestforms,thenewfamilyofmodelsandtheexistingmultinomialmodelsareequivalent,theybehavedi erentlyformanysmoothingmethods.WeshowthatthePoissonmodelhasseveralad-vantagesoverthemultinomialmodel,includingnaturallyac-commodatingper-termsmoothingandallowingformoreac-curatebackgroundmodeling.Wepresentseveralvariantsofthenewmodelcorrespondingtodi erentsmoothingmeth-ods,andevaluatethemonfourrepresentativeTRECtestcollections.Theresultsshowthatwhiletheirbasicmod-elsperformcomparably,thePoissonmodelcanoutperformmultinomialmodelwithper-termsmoothing.Theperfor-mancecanbefurtherimprovedwithtwo-stagesmoothing.CategoriesandSubjectDescriptors:H.3.3[Informa-tionSearchandRetrieval]:RetrievalModelsGeneralTerms:Algorithms
Keywords:Languagemodels,Poissonprocess,querygen-eration,formalmodels,termdependentsmoothing
1.INTRODUCTION
Asanewtypeofprobabilisticretrievalmodels,languagemodelshavebeenshowntobee ectiveformanyretrievaltasks[21,28,14,4].Amongmanyvariantsoflanguagemod-elsproposed,themostpopularandfundamentaloneisthequery-generationlanguagemodel[21,13],whichleadstothequery-likelihoodscoringmethodforrankingdocuments.Insuchamodel,givenaqueryqandadocumentd,wecom-putethelikelihoodof“generating”queryqwithamodelestimatedbasedondocumentd,i.e.,theconditionalprob-
Permissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforpro torcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthe rstpage.Tocopyotherwise,torepublish,topostonserversortoredistributetolists,requirespriorspeci cpermissionand/orafee.
SIGIR’07,July23–27,2007,Amsterdam,TheNetherlands.Copyright2007ACM978-1-59593-597-7/07/0007...$5.00.
abilityp(q|d).Wecanthenrankdocumentsbasedonthelikelihoodofgeneratingthequery.
Virtuallyalltheexistingquerygenerationlanguagemod-elsarebasedoneithermultinomialdistribution[19,6,28]ormultivariateBernoullidistribution[21,18].Themultino-mialdistributionisespeciallypopularandalsoshowntobequitee ective.Theheavyuseofmultinomialdistributionispartlyduetothefactthatithasbeensuccessfullyusedinspeechrecognition,wheremultinomialdistributionisanat-uralchoiceformodelingtheoccurrenceofaparticularwordinaparticularpositionintext.Comparedwithmultivari-ateBernoulli,multinomialdistributionhastheadvantageofbeingabletomodelthefrequencyoftermsinthequery;incontrast,multivariateBernoullionlymodelsthepresenceandabsenceofqueryterms,thuscannotcapturedi erentfrequenciesofqueryterms.However,multivariateBernoullialsohasonepotentialadvantageovermultinomialfromtheviewpointofretrieval:inamultinomialdistribution,theprobabilitiesofallthetermsmustsumto1,makingithardtoaccommodateper-termsmoothing,whileinamultivari-ateBernoulli,thepresenceprobabilitiesofdi erenttermsarecompletelyindependentofeachother,easilyaccommo-datingper-termsmoothingandweighting.Notethattermabsenceisalsoindirectlycapturedinamultinomialmodelthroughtheconstraintthatallthetermprobabilitiesmustsumto1.
Inthispaper,weproposeandstudyanewfamilyofquerygenerationmodelsbasedonthePoissondistribution.Inthisnewfamilyofmodels,wemodelthefrequencyofeachtermindependentlywithaPoissondistribution.Toscoreadocu-ment,wewould rstestimateamultivariatePoissonmodelbasedonthedocument,andthenscoreitbasedonthelike-lihoodofthequerygivenbytheestimatedPoissonmodel.Insomesense,thePoissonmodelcombinestheadvantageofmultinomialinmodelingtermfrequencyandtheadvantageofthemultivariateBernoulliinaccommodatingper-termsmoothing.Indeed,similartothemultinomialdistribution,thePoissondistributionmodelstermfrequencies,butwith-outtheconstraintthatallthetermprobabilitiesmustsumto1,andsimilartomultivariateBernoulli,itmodelseachtermindependently,thuscaneasilyaccommodateper-termsmoothing.
Asintheexistingworkonmultinomiallanguagemodels,smoothingiscriticalforthisnewfamilyofmodels.Wede-riveseveralsmoothingmethodsforPoissonmodelinparalleltothoseusedformultinomialdistributions,andcomparethecorrespondingretrievalmodelswiththosebasedonmulti-
A query is treated as a generation process. Given a sequence of terms in...3. A NEW LANGUAGE MODEL FOR INFORMATION RETRIEVAL Our solution is to ...
An Ontology-Based Information Retrieval Model for Vegetables E-Commerce_专业资料。22 1:8egu011 00rr11s07 c t 8Ai v ff5.JunlofIerora nga.. S...
Us, query tf 1` ? # 2-Poisson?4_ ?f H H ? f x? p ?...Robertson. A probabilistic models of information retrieval: Development and ...
A Brief Review of Information Retrieval Models ...For example, the 2-Poisson probabilistic indexing ...5 special case of the query generation model. ...
贝氏主题混合资讯检?模型 Bayesian Topic Mixture Model for Information Retrieval 吴孟淞 许轩睿 简仁宗 国?成功大学资讯工程学系 Department of Computer Science ...
Learning to Rank for Information Retrieval and ...Given a query from the user, the system ... candidate generation and re-ranking (see Fig. ...
A Study of the Poisson-Boltzmann Equation and ...Model 2-D view of 3-D protein in solvent ...Generation Example of Discretization Material ...
Probabilistic Information Retrieval Part ISurvey...query Probability Ranking Principle If a reference ...n-Poisson Indexing Next: Probabilistic Inference ...
56页 1财富值 2-Poisson Model for Prob... 暂无评价 6页 免费...Information Retrieval Threshold For query q, retrieve all documents with ...
generation?give?exact?same?results: ???P(funny?...Translation?model:?accounts?for?multiple?query? ...strategies?for?cross?language?information?retrieval....
热门文档