Memory based learning in NLP

《Memory based learning in NLP》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、Te hni alReport:U
UCDCS-R-99-2125,U
UCCS,ar h99.emoryBasedearninginDanRothDepartmentofComputerS ien eUniversityof
llinoisatUrbana-Champaign1304WSpringeldAve.,Urbana,
61801danr s.uiu .eduAbstra tla ingseeminglydierentlearningparadigms{Bononehandandfeature-bas

2、edmethods,in ludingbothWestudymemory-basedlearningmethodsand[ma hinelearningandstatisti s-basedmethodsRoth,showthatthey anbeviewedaslearninglinear℄1998;1999)ontheother{withinthesame omputa-predi torsover ommonlyusedfeaturespa es.tionalparadigmallowsonetodevelopanundersta

3、ndingWesuggestthatthisviewallowsonetostudyoftheadvantagesanddisadvantagesofea happroa h,memorybasedmethodsandothersu essfulandtothe ir umstan esunderwhi hthesemightsur-learningalgorithmsusedinwithinthefa e.Weaddressonlyafewoftheseissues{expressivity,same omputational

4、frameworkandmayre-spa e omplexitytradeos,featuresele tionandweight-sultinimprovedalgorithmsandabetterun-ing{inthisshortabstra t.derstandingfortheroleoflearninginnaturalWenotethatwhileweshowthatBmethods anlanguageinferen es.beviewedasafairlylimited lassoffeature-basedl

5、earn-ingmethods,fromsomeperspe tivesitmaybeillumi-natingtorealizethatruleabstra tion"methodsmay1
ntrodu tionbethoughtofasBmethods,andthatsomefeature-basedmethodsaremoresuitedforthatviewthanothers.emory-basedlearning(B)methodsrepresentsomeThemain ontributionofthepap

6、erisinprovidingaofthesu essfullearningapproa hestriedinre entyearsunieda ountforavarietyofmethodsthatarewidelyinempiri alnaturallanguageAppearin
CA
-99pro-usedinnaturallanguageappli ations.Thehopeis[ essingZavrelandDaelemans,1997;Daelemansetal.,thatprovidinga oherent

7、viewofdierentlearningap-℄1999;Argamonetal.,1998.Advo atesofthesemeth-proa hes ouldhelpindevelopingbetterlearningmeth-odsattributethesu esstothefundamentalhypothesisodsandabetterunderstandingfortheroleoflearningofmemorybasedmethods;namely,theassumptionthatinnaturallangu

8、ageinferen es.performan einnaturallanguagepro essingisbasedonreasoningonthebasisofsimilarityofnewsituat