资源描述:
《cikm2010_probabilistic first pass retrieval for search advertising》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、ProbabilisticFirstPassRetrievalforSearchAdvertising:FromTheorytoPractice∗HemaRaghavanRukminiIyerYahoo!IncMicrosoftCorp4402GreatAmericaParkwaySpeechatMicrosoft
2、MicrosoftResearchSantaClara,CA,95054MountainView,CA94041raghavan@yahoo-inc.comrukmini@microsoft.comABSTRACTGeneralTermsIn
3、formationretrievalinsearchadvertising,asinotherad-Algorithms,Experimentationhocretrievaltasks,aimstofindthemostappropriaterank-ingoftheaddocumentsofacorpusforagivenquery.KeywordsInadditiontorankingtheaddocuments,wealsoneedtosponsoredsearch,advertising,languagemodelsfilterorthreshol
4、dirrelevantadsfromparticipatingintheauctiontobedisplayedalongsidesearchresults.Inthiswork,wedescribeourexperienceinimplementingasuc-1.INTRODUCTIONcessfuladretrievalsystemforacommercialsearchengineTheprimarysourceofrevenueformajorsearchenginesbasedontheLanguageModeling(LM)framewor
5、kforre-isfromadvertisementsdisplayedduringusersearches.Atrieval.TheLMdemonstratessignificantperformanceim-searchenginetypicallydisplayssponsoredtextlistingsonprovementsoverthebaselinevectorspacemodel(TF-IDF)thetopandtherighthandsideoftheweb-searchresults,insystemthatwasinproductio
6、natthetime.Fromamodel-responsetoauserquery.Therevenuemodelfortheselist-ingperspective,weproposeanovelapproachtoincorporateingsis“pay-per-click”wheretheadvertiserpaysthesearchquerysegmentationandphrasesintheLMframework,dis-engineonlyiftheuserclicksonthead.Sponsoredsearchcussimpact
7、ofscorenormalizationforrelevancefiltering,offersamoretargetedandlessexpensivewayofmarketingandpresentpreliminaryresultsofincorporatingqueryex-formostadvertisersascomparedtomassmedialikeTVpansionsusingqueryrewritingtechniques.Fromanimple-andnewspapers,andonlineadspendhasgainedsignifi
8、cantmentationperspective,wealsodiscussreal-timelatencycon-momentuminrecentyears[1].straintsofaproductionsearchengineandhowweovercomeMostofthemajorsearchenginestypicallyhavethousandsthembyadaptingtheWANDalgorithmtoworkwithlan-ofadvertiserswithmillionsoftextadsavailabletobedis-guag
9、emodels.Insum,ourLMformulationisconsider