资源描述：
《BoosTexter a boosting-based system for text categorization》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、BoosTexter:ABoosting-basedSystemforTextCategorizationRBERTE.SCHAREschapire@research.att.comAT&TLabs,ShannonLaboratory,180ParkAvenue,RoomA279,FlorhamPark,NJ07932-0971YRASGERsinger@research.att.comAT&TLabs,ShannonLaboratory,180ParkAvenue,RoomA277,FlorhamPark,NJ07932-0971Abst
2、ract.Thisworkfocusesonalgorithmswhichlearnfromexamplestoperformmulticlasstextandspeechcategorizationtasks.Ourapproachisbasedonanewandimprovedfamilyofboostingalgorithms.Wedescribeindetailanimplementation,calledBoosTexter,ofthenewboostingalgorithmsfortextcategorizationtasks.Weprese
3、ntresultscomparingtheperformanceofBoosTexterandanumberofothertext-categorizationalgorithmsonavarietyoftasks.Weconcludebydescribingtheapplicationofoursystemtoautomaticcall-typeidentificationfromunconstrainedspokencustomerresponses.1.IntroductionTextcategorizationistheproblemofclass
4、ifyingtextdocumentsintocategoriesorclasses.Forinstance,atypicalproblemisthatofclassifyingnewsarticlesbytopicbasedontheirtextualcontent.Anotherproblemistoautomaticallyidentifythetypeofcallrequestedbyacustomer;forinstance,ifthecustomersays,“Yes,IwouldliketochargethiscalltomyVisa,”w
5、ewantthesystemtorecognizethatthisisacalling-cardcallandtoprocessthecallaccordingly.(Althoughthisisactuallyaspeech-categorizationproblem,wecanneverthelessapplyatext-basedsystembypassingthespokenresponsesthroughaspeechrecognizer.)Inthispaper,weintroducetheuseofamachine-learningtech
6、niquecalledboostingtotheproblemoftextcategorization.Themainideaofboostingistocombinemanysimpleandmoderatelyinaccuratecategorizationrulesintoasingle,highlyaccuratecategorizationrule.Thesimplerulesaretrainedsequentially;conceptually,eachruleistrainedontheexampleswhichweremostdifficu
7、lttoclassifybytheprecedingrules.Ourapproachisbasedonanewandimprovedfamilyofboostingalgorithmswhichwehavedescribedandanalyzedindetailinacompanionpaper(Schapire&Singer,1998).ThisnewfamilyextendsandgeneralizesFreundandSchapire'sAdaBoostalgorithm(Fre-und&Schapire,1997),whichhasbeenst
8、udiedextensivelyandwhichhasbeenshowntope
