1991_evaluation of an inference network based retr

EvaluationofanInferenceNetwork-BasedRetrievalModelHOWARDTURTLEandW.BRUCECROFTUniversityofMassachusettsTheuseofinferencenetworkstosupportdocumentretrievalisintroduced.Anetwork-basedretrievalmodelisdescribedandcomparedtoconventionalprobabilisticandBooleanmodels.Theperformanceofaretrievalsystembasedontheinferencenetworkmodelisevaluatedandcomparedtoperformancewithconventionalretrievalmodels,CategoriesandSubjectDescriptors:H.30[InformationStorageandRetrieval]:General;H.3.3[InformationStorageandRetrieval]:InformationSearchandRetrieval–retrieualmodels;H.3.m[InformationStorageandRetrieval]:Miscellaneous.GeneralTerms:Experimentation,Performance,TheoryAdditionalKeyWordsandPhrases:Documentretrieval,inferencenetworks,networkretrivalmodels.1.INTRODUCTIONNetworkrepresentationshavebeenusedininformationretrievalsinceatleasttheearly1960’s.Networkshavebeenusedtosupportdiverseretrievalfunctions,includingbrowsing[38],documentclustering[7],spreadingactiva-tionsearch[4],supportformultiplesearchstrategies[11],andrepresentationofuserknowledge[27]ordocumentcontent[40].Recentworksuggeststhatsignificantimprovementsinretrievalperfor-mancewillrequiretechniquesthat,insomesense“understand”thecontentofdocumentsandqueries[9,43]andcanbeusedtoinferprobablerelation-shipsbetweendocumentsandqueries.Inthisview,informationretrievalisaninferenceorevidentialreasoningprocessinwhichweestimatetheproba-bilitythatauser’sinformationneed,expressedasoneormorequeries,ismetgivenadocumentas“evidence.”Networkrepresentationsshowpromiseasmechanismsforinferringthesekindsofrelationships[4,12].ThisworkwassupportedinpartbytheAirForceOfficeofScientificResearchundercontract90-0110andbyNSFgrantIRI-8814790.Authors’Address:ComputerandInformationScienceDepartment,UniversityofMassachusetts,Amherst,MA01003.email:[e-mail],eduPermissiontocopywithoutfeeallorpartofthismaterialisgrantedprovidedthatthecopiesarenotmadeordistributedfordirectcommercialadvantage,theACMcopyrightnoticeandthetitleofthepublicationanditsdateappear,andnoticeisgiventhatcopyingisbypermissionoftheAssociationforComputingMachinery.Tocopyotherwise,ortorepublish,requiresafeeand/orspecificpermission.@1991ACM1046-8188/91/0700-0187$01.50ACMTransactionsonInformationSystems,Vol.9,No3,July1991,Pages187-222.188.H.TurtleandW.BCroftTheideathatretrievalisaninferenceorevidentialreasoningprocessisnotnew.Cooper’slogicalrelevance[6]isbasedondeductiverelationshipsbetweenrepresentationsofdocumentsandinformationneeds.Wilson’ssitua-tionalrelevance[441extendsthisnotiontoincorporateinductiveoruncertaininferencebasedonthedegreetowhichdocumentssupportinformationneeds.Thetechniquesrequiredtosupportthesekindsofinferencearesimilartothoseusedinexpertsystemsthatmustreasonwithuncertaininformation.Anumberofcompetinginferencemodelshavebeendevelopedforthesekindsofexpertsystems[19,21]andseveralofthesemodelscanbeadaptedtothedocumentretrievaltask.Intheresearchdescribedhereweadaptaninferencenetworkmodeltotheretrievaltask.Theuseofthemodelisintendedtodothefollowing:—Supporttheuseofmultipledocumentrepresentationschemes.Researchhasshownthatagivenquerywillretrievedifferentdocumentswhenappliedtodifferentrepresentations,evenwhentheaverageretrievalper-formanceachievedwitheachrepresentationisthesame.Katzer,forexam-ple,foundlittleoverlapindocumentsretrievedusingsevendifferentrepresentations,butfoundthatdocumentsretrievedbymultiplerepresen-tationswerelikelytoberelevant[20].Similarresultshavebeenobtainedwhencomparingterm-withcluster-basedrepresentations[2]andterm-withcitation-basedrepresentations[16].—Allowresultsfromdifferentqueriesandquerytypestobecombined.Givenasinglenaturallanguagedescriptionofaninformationneed,differentsearcherswillformulatedifferentqueriestorepresentthatneedandwillretrievedifferentdocuments,evenwhenaverageperformanceisthesameforeachsearcher[20,24].Again,documentsretrievedbymultiplesearchersaremorelikelytoberelevant.Adescriptionofaninformationneedcanbeusedtogenerateseveralqueryrepresentations(e.g.,probabilistic,Boolean),eachusingadifferentquerystrategyandeachcapturingdifferentaspectsoftheinformationneed.Thesedifferentsearchstrategiesareknowntoretrievedifferentdocumentsforthesameunderlyinginformationneed[9].—Facilitateflexiblematchingbetweenthetermsorconceptsmentionedinqueriesandthoseassignedtodocuments.Thepoormatchbetweenthevocabularyusedtoexpressqueriesandthevocabularyusedtorepresentdocumentsappearstobeamajorcauseofpoorrecall[151.Recallcanbeimprovedusingdomainknowledgetomatchqueryandrepresentationconceptswithoutsignificantlydegradingprecision.Theresultingformalretrievalmodelintegratesseveralpreviousmodelsinasingletheoreticalframework;multipledocumentandqueryrepresentationsaretreatedasevidencewhichiscombinedtoestimatetheprobabilitythatadocumentsatisfiesauser’sinformationneed.Inwhatfollowswebrieflyreviewcandidateinferencemodels(Section2),presentaninferencenetwork-basedretrievalmodel(Sections3and5),comparethenetworkmodeltocurrentretrievalmodels(Section4),andACMTransactIonsonInformationSystems,Vol9,No3,July1991EvaluationofanInferenceN