欢迎来到天天文库
浏览记录
ID:57125167
大小:953.50 KB
页数:51页
时间:2020-08-01
《搜索引擎实践课件.ppt》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、SearchEnginesInformationRetrievalinPracticeAllslides©AddisonWesley,2008EvaluationEvaluationiskeytobuildingeffectiveandefficientsearchenginesmeasurementusuallycarriedoutincontrolledlaboratoryexperimentsonlinetestingcanalsobedoneEffectiveness,efficiencyandcostarerelatede.g.
2、,ifwewantaparticularlevelofeffectivenessandefficiency,thiswilldeterminethecostofthesystemconfigurationefficiencyandcosttargetsmayimpacteffectivenessEvaluationCorpusTestcollectionsconsistingofdocuments,queries,andrelevancejudgments,e.g.,TestCollectionsTRECTopicExampleRelev
3、anceJudgmentsObtainingrelevancejudgmentsisanexpensive,time-consumingprocesswhodoesit?whataretheinstructions?whatisthelevelofagreement?TRECjudgmentsdependontaskbeingevaluatedgenerallybinaryagreementgoodbecauseof“narrative”PoolingExhaustivejudgmentsforalldocumentsinacollect
4、ionisnotpracticalPoolingtechniqueisusedinTRECtopkresults(forTREC,kvariedbetween50and200)fromtherankingsobtainedbydifferentsearchengines(orretrievalalgorithms)aremergedintoapoolduplicatesareremoveddocumentsarepresentedinsomerandomordertotherelevancejudgesProducesalargenumb
5、erofrelevancejudgmentsforeachquery,althoughstillincompleteQueryLogsUsedforbothtuningandevaluatingsearchenginesalsoforvarioustechniquessuchasquerysuggestionTypicalcontentsUseridentifierorusersessionidentifierQueryterms-storedexactlyasuserenteredListofURLsofresults,theirran
6、ksontheresultlist,andwhethertheywereclickedonTimestamp(s)-recordsthetimeofusereventssuchasquerysubmission,clicksQueryLogsClicksarenotrelevancejudgmentsalthoughtheyarecorrelatedbiasedbyanumberoffactorssuchasrankonresultlistCanuseclickthoughdatatopredictpreferencesbetweenpa
7、irsofdocumentsappropriatefortaskswithmultiplelevelsofrelevance,focusedonuserrelevancevarious“policies”usedtogeneratepreferencesExampleClickPolicySkipAboveandSkipNextclickdatageneratedpreferencesQueryLogsClickdatacanalsobeaggregatedtoremovenoiseClickdistributioninformation
8、canbeusedtoidentifyclicksthathaveahigherfrequencythanwouldbeexpectedhighcorrelationwithrelevance
此文档下载收益归作者所有