exploit site dlevel inform improv web search improv latent concept expans hierarch markov random field learn effici rank recent increas attent focus directli optim rank measur induc sparsiti learn model attempt made relat approach problem learn rank paper spars algorithm directli optim normal discount cumul gain ndcg wide rank measur begin establish reduct framework reduc rank measur ndcg import weight pairwis classif provid sound theoret guarante reduct bound realiz ndcg regret term properli weight pairwis classif regret impli good perform robustli transfer pairwis classif rank base convert pairwis loss function conceiv account sparsiti rank model gradient possess perform guarante sake achiev sparsiti algorithm name rsrank devis perform regular truncat gradient descent final experiment result benchmark collect confirm signific advantag rsrank comparison baselin method rank tempor constraint learn effici rank recent increas attent focus directli optim rank measur induc sparsiti learn model attempt made relat approach problem learn rank paper spars algorithm directli optim normal discount cumul gain ndcg wide rank measur begin establish reduct framework reduc rank measur ndcg import weight pairwis classif provid sound theoret guarante reduct bound realiz ndcg regret term properli weight pairwis classif regret impli good perform robustli transfer pairwis classif rank base convert pairwis loss function conceiv account sparsiti rank model gradient possess perform guarante sake achiev sparsiti algorithm name rsrank devis perform regular truncat gradient descent final experiment result benchmark collect confirm signific advantag rsrank comparison baselin method relev rank onlin date system relev rank onlin date system learn concept import weight depend model model queri concept term depend shown signific posit effect retriev perform task web search relev high rank critic previou work treat concept equal import assumpt hold longer complex queri paper show effect exist term depend model natur extend assign weight concept demonstr weight depend model train exist learn rank techniqu small number train queri studi compar effect endogen collectionbas exogen base extern sourc featur determin concept import test weight depend model perform experi publicli trec corpora proprietari web corpu experiment result model consist significantli outperform standard bag word model unweight term depend model combin endogen exogen featur gener result retriev effect anatomi ad structur index retriev sponsor search search engin adapt feedback control adjust time dsensit queri semi dparametr dparametr term weight inform retriev abstract previou research term weight inform retriev focus develop special parametr term weight function exampl includ idf vector space formul bm languag model weight term weight function take specif parametr form weight function proven highli effect impos strict constraint function form term weight constraint possibl degrad retriev effect paper propos class term weight scheme call semi parametr nonparametr weight weight scheme make fewer assumpt underli term weight data speak argu robust weight scheme potenti significantli effect compar exist parametr scheme grow amount train data semi dparametr dparametr term weight inform retriev abstract previou research term weight inform retriev focus develop special parametr term weight function exampl includ idf vector space formul bm languag model weight term weight function take specif parametr form weight function proven highli effect impos strict constraint function form term weight constraint possibl degrad retriev effect paper propos class term weight scheme call semi parametr nonparametr weight weight scheme make fewer assumpt underli term weight data speak argu robust weight scheme potenti significantli effect compar exist parametr scheme grow amount train data build enrich document represent aggreg anchor text anchor text plai critic role varieti search task perform hypertextu domain includ enterpris search wiki search web search common practic enrich document standard textual represent anchor text incom hyperlink approach match relev page inlink paper propos method overcom anchor text sparsiti enrich document represent anchor text aggreg hyperlink graph aggreg mechan act smooth diffus anchor text domain rigor evalu propos approach larg web search test collect result show approach significantli improv retriev effect longer difficult queri improv search relev implicitli tempor queri web search queri implicit intent detect effect improv search qualiti user enter queri toyota camri find offici web onlin expans rare queri sponsor search sponsor search system task match queri relev advertis current state art match algorithm expand user queri varieti extern resourc web search result expans base algorithm highli effect larg ineffici appli real time practic algorithm appli offlin popular queri result expens oper cach fast access queri time paper describ effici effect approach match ad rare queri process offlin approach build expand queri represent leverag offlin process relat popular queri experiment result show approach significantli improv effect advertis rare queri neglig increas comput cost swing swing learn advertis web textual advertis interpret search problem corpu ad displai context contrast convent inform retriev system return result corpu document lexic relat queri web advertis accept occasion desir show result ad relev user interest show irrelev ad avoid annoi user produc econom benefit paper pose decis problem swing show ad incom request propos method address problem simpl threshold approach machin learn approach collect analyz set candid ad augment extern knowledg experiment evalu base editori judgment show predict high accuraci swing content match sponsor search advertis statist view bin retriev model bag word effect model depend featur inform retriev automat featur select markov random field model inform retriev previou applic markov random field model inform retriev manual chosen featur difficult imposs priori set featur task data set develop automat featur select techniqu paper describ greedi procedur automat select featur markov random field model inform retriev propos robust method describ class textual inform retriev featur experiment result evalu standard trec test collect show featur select algorithm produc model significantli effect equal effect model manual select featur past similar measur short segment text abstract measur similar document queri extens studi inform retriev grow number task requir comput similar short segment text task includ queri reformul sponsor search imag retriev standard text similar measur perform poorli task data spars lack context work studi problem inform retriev perspect focus text represent similar measur examin rang similar measur includ pure lexic measur stem languag model base measur formal evalu analyz method queri queri similar task queri web search log analysi insight strength weak method includ import tradeoff effect effici pseudo dalign multilingu corpora machin translat document align refer find correspond document exact translat defin pseudo align task find topic oppos exact correspond document languag appli semisupervis method pseudo align multilingu corpora specif construct topicbas graph languag exact correspond subset document project unalign document share lower dimension space demonstr close document lower dimension space tend share topic applic machin translat cross lingual inform analysi experiment result show pseudo align multilingu corpora feasibl document align produc qualit sound techniqu requir linguist knowledg corpu averag corpu consist exact correspond topic correspond occur top foreign neighbor lowerdimension space exact correspond occur top foreign neighbor space show substanti improv result method incorpor languag independ inform latent concept expans markov random field queri expans form pseudo relev feedback relev feedback common techniqu improv retriev effect previou approach import issu role featur import model term depend paper propos robust queri expans techniqu base markov random field model inform retriev techniqu call latent concept expans mechan model term depend expans arbitrari featur model power framework simpl term occurr featur implicitli expans techniqu evalu techniqu relev model state art languag model queri expans techniqu model demonstr consist signific improv retriev effect trec data set describ techniqu gener meaning multi term concept task queri suggest reformul linear featur dbase model inform retriev abstract number linear featur base model propos inform retriev commun recent model present differ share common underli framework paper explor discuss theoret issu framework includ paramet space detail supervis train algorithm directli maxim evalu metric consider averag precis present result show train model lead significantli test set perform compar train method directli maxim metric final show linear featur base model consist significantli outperform current state art retriev model correct choic featur bag word model implicit user prefer inform retriev paper report recent work field inform retriev attempt overli simplifi approach repres document queri bag word simpl model make difficult accur model user inform estim sensit gener parameter retriev model improv estim relev model larg extern corpora inform retriev algorithm leverag collect statist improv perform statist comput small evalu corpu larger evalu corpora improv perform specif advoc incorpor extern corpora base languag model refer process extern expans compar tradit pseudo relev feedback techniqu extern expans stabl topic ectiv term averag precis result show high qualiti corpu compar evalu corpu ectiv web result show extern expans outperform simul relev feedback addit propos method predict extent extern expans improv retriev perform measur demonstr posit correl improv averag precis indri trec notebook lesson learn terabyt track submiss classif approach document pair relev feedback inform explor textual textual document pair featur classifi unjudg document relev relev predict rank baselin document retriev featur includ citat measur url similar featur machin learn system document rank differ score assign baselin retriev system similar measur track inform flow text similar span spectrum broad topic similar extrem document ident intermedi level similar result summar paraphras copi stronger form topic relev applic inform flow analysi question answer task paper explor mechan measur intermedi kind similar focus task identifi piec inform origin sentenc sentenc document document comparison incorpor algorithm recap prototyp inform flow analysi tool experiment result recap mechan propos exist method identifi intermedi form similar indri trec terabyt track work detail experi carri indri search engin trec terabyt track result present task includ cienci ad hoc name page find cienci run focus queri optim techniqu ad hoc run import term proxim document qualiti name page find run investig document prior document structur umass robust mixtur relev model queri expans paper describ umass trec robust track experi focu approach term proxim pseudo relev feedback extern collect result approach highli ectiv analysi statist question classif fact dbase question question classif system plai import role question answer system wide rang domain goal question classif accur assign label question base expect answer type approach past reli match question hand craft rule rule requir labori ort creat su er specif statist question classif method overcom issu emploi machin learn techniqu empir show statist approach robust achiev good perform divers data set hand tune examin role di erent syntact semant featur perform find semant featur tend increas perform pure syntact featur final analyz common misclassif error provid insight wai overcom infer network approach imag retriev imag retriev system fragment text imag queri user complex inform easili express form paper propos model base infer network framework inform retriev emploi power queri languag structur queri oper term weight combin text imag queri model parametr method estim probabl infer network imag annot retriev result report compar publish system illustr structur weight queri result show power queri languag result system perform robust compar exist approach formal multipl dbernoulli model languag model languag model assum document gener sampl multinomi distribut formal method estim exist multinomi case paper reexamin languag model base multipl bernoulli distribut assumpt explor past formal present smooth languag model estim assumpt formal bayesian standpoint formal multipl dbernoulli model languag model languag model assum document gener sampl multinomi distribut formal method estim exist multinomi case paper reexamin languag model base multipl bernoulli distribut assumpt explor past formal present smooth languag model estim assumpt formal bayesian standpoint indri trec terabyt track paper overview experi carri trec terabyt track indri search engin indri effici effect distribut search engin inqueri base infer network framework support structur queri unlik inqueri languag model probabl network ad flexibl describ approach terabyt track involv automat construct structur queri titl portion trec topic method term proxim inform html document structur addit number optim procedur effici queri process explain combin languag model infer network approach retriev paper describ develop structur document collect user gener text numer metadata explor exploit metadata inform retriev ir collect consist document extract youtub video page basketbal gener nba nation basketbal associ set topic relev judgement addit collect user profil relat nba collect baselin ir experi report effect video metadata retriev effect result surprisingli show search video titl perform significantli search addit metadata text field video tag descript 