文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 农林牧渔 > 林学 > Knowledge Supervised Text Classification with
侵权投诉

Knowledge Supervised Text Classification with

分本分类

KnowledgeSupervisedTextClassi cationwith

NoLabeledDocuments

CongleZhang1,2,Gui-RongXue1,2,andYongYu

2ApexLab,ShanghaiJiaotongUniversity,Shanghai,200240StateKeyLabofCAD&CG,ZhejiangUniversity,Hangzhou,310058

{zhangcongle,grxue,yyu}@apex.sjtu.edu.cn1

Abstract.Intraditionaltextclassi cationapproaches,thesemanticmeaningsoftheclassesaredescribedbythelabeleddocuments.Sincela-belingdocumentsisoftentimeconsumingandexpensive,itisapromisingideathataskinguserstoprovidesomekeywordstodepicttheclasses,in-steadoflabelinganydocuments.However,shortpiecesofkeywordsmaynotcontainenoughinformationandthereforemayleadtounreliableclas-si er.Fortunately,therearelargeamountofpublicdataeasilyavailableinwebdirectories,suchasODP,Wikipedia,etc.Weareinterestedinexploringtheenormouscrowdintelligencecontainedinsuchpublicdatatoenhancetextclassi cation.Inthispaper,weproposeanoveltextclassi cationframeworkcalled“KnowledgeSupervisedLearning”(KSL),whichutilizestheknowledgeinkeywordsandthecrowdintelligencetolearntheclassi erwithoutanylabeleddocuments.Wedesignatwo-stageriskminimization(TSRM)approachfortheKSLproblem.Itcanopti-mizetheexpectedpredictionriskandbuildthehighqualityclassi er.Empiricalresultsverifyourclaim:ouralgorithmcanachieveabove0.9onMicro-F1onaverage,whichismuchbetterthanbaselinesandevencomparableagainstSVMclassi ersupervisedbylabeleddocuments.1Introduction

Supervisedlearning[2,5]andsemi-supervisedlearning[18,19]algorithmsarewidelyusedintextclassi cation.Theyrequirehighqualitylabeleddocumentsforeachclasstobuildtheclassi er.Somestudies[13]proposeanotherstrat-egythatuserscanproviderepresentativekeywordstodepicttheclasses.Suchalternativeisverypromisingbecauselabelingdocumentsisoftenmuchmoreex-pensivethangivingsomekeywords.Forexample,wheninternetserviceproviderswanttoclassifywebpagesaccordingtousers’interests,itismucheasiertoaskthemissuesomewordstodescribetheirinterestthantocollectwebpagestheyarereallyinterestedin.

However,onecanhardlyexpectuserstoprovidelonganddetailedkeywords.Itmeansthattheinformationofthekeywordsarenotenoughtolearnareli-ableclassi er.Thegapbetweenthekeywordsandthecorrespondingsemanticmeaningshouldbe lledup.Wenoticethatmanywebdirectoriesgrowsenor-mousandeasilyavailable,e.g.Opendirectoryproject(ODP),Wikipedia,Ya-hoo!Directory.SuchpublicdatasetscontainhugeamountofarticlestogetherT.-B.HoandZ.-H.Zhou(Eds.):PRICAI2008,LNAI5351,pp.509–520,2008.cSpringer-VerlagBerlinHeidelberg2008

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email