Diverse Classiers for NLP Disambiguation Tasks Comparison, Optimization, Combination, and E

《Diverse Classiers for NLP Disambiguation Tasks Comparison, Optimization, Combination, and E》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、DiverseClassiersforDisambiguationTasksComparison,ptimization,Combination,andEvolutionakubZavrel

2、SvenDegroeve}Anneool

3、WalterDaelemans

4、ristiinaokinen}

5、CTS-anguageTe hnologyGroup,UniversityofAntwerpfzavrel

6、kool

7、daelemguia.ua.a .be}CenterforEvolutionary

8、anguageEngineering,
eper,Belgiumfsven.degroeve

9、kristiina.jokinengsail. omAbstra t
nthispaperwereportpreliminaryresultsfromanongoingstudythatinvestigatestheperfor-man eofma hinelearning lassiersonadiversesetofaturalanguagero essing()tasks.First,we omparea

10、numberofpopularexistinglearningmethods(euralnetworks,emory-basedlearning,Ruleindu tion,De isiontrees,aximumEntropy,Winnower- eptrons,aiveBayesandSupportVe tora hines),anddis usstheirpropertiesvisavistypi aldatasets.ext,weturntomethodstooptimizethepara

11、metersofsinglelearningmethodsthrough ross-validationandevolutionaryalgorithms.Thenweinvestigatehowwe angetthebestofallsinglemethodsthrough ombinationofthetestedsystemsin lassierensembles.Finallywedis ussnewandmorethoroughmethodsofautomati ally onstru tingensemb

12、lesof lassiersbasedonthete hniquesusedforparameteroptimization.eywords:odelsandalgorithmsfor omputationalneuralar hite tures1
ntrodu tion
nre entyearstheeldofaturalanguagero essing()hasbeenradi allytransformedbyaswit hfromadedu tivemethodology(i.e.expl

13、ainingdatafromtheoriesormodels onstru tedmanually)toanindu tivemethodology(i.e.derivingmodelsandtheoriesfromdata)(seee.g.Ab-ney(1996)forareview).Animportant omponentofthistransformationistherealizationthatmanytasks anbemodeledassimple lassi ationtasksorasens

14、emblesofsimple lassi-ers(Daelemans,1996;Ratnaparkhi,1997).Thushasbeenableto apitalizeonalargebodyofresear hintheeldofma hinelearningandstatisti almodeling.This,a ompaniedbythe ontinuingexplosionof omputerpower,storagesize,andavailabilityoftraining orpora,h

15、asleadtoin reasinglya uratelanguagemodelsforaqui klygrowingnumberoflanguagemodelingtasks.However,whi hma hinelearningmethodshavethebestperforman eondatasetsisstil