Feature-rich memory-based classification for shallow nlp and information extraction

《Feature-rich memory-based classification for shallow nlp and information extraction》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、Feature-RichMemory-BasedClassicationforShallowNLPandInformationExtractionJakubZavrel1andWalterDaelemans21TextkernelBV,Nieuwendammerkade28/a17,1022AB,Amsterdam,TheNetherlandszavrel@textkernel.nl2CNTS,UniversityofAntwerp,Universiteitsplein1,BuildingA,B-2610Antwerpen,

2、Belgiumwalter.daelemans@uia.ua.ac.beAbstract.Memory-BasedLearning(MBL)isbasedonthestorageofallavailabletrainingdata,andsimilarity-basedreasoningforhandlingnewcases.Byinter-pretingtaskssuchasPOStaggingandshallowparsingasclassicationtasks,theadvantagesofMBL(implicits

3、moothingofsparsedata,automaticintegrationandrelevanceweightingofinformationsources,handlingexceptionaldata)contributetostate-of-the-artaccuracy.However,HiddenMarkovModels(HMM)typicallyachievehigheraccuracythanMBL(andotherMachineLearningapproaches)fortaskssuchasPOSta

4、ggingandchunking.Inthispaper,weinvestigatehowtheadvantagesofMBL,suchasitspotentialtointegratevarioussourcesofinforma-tion,cometoplaywhenwecompareourapproachtoHMMsontwoInformationExtraction(IE)datasets:thewell-knownSeminarAnnouncementdatasetandanewGermanCurriculumVit

5、aedataset.1Memory-BasedLanguageProcessingMemory-BasedLearning(MBL)isasupervisedclassication-basedlearningmethod.Avectoroffeaturevalues(aninstance)isassociatedwithaclassbyaclassierthatlazilyextrapolatesfromthemostsimilarset(nearestneighbors)selectedfromallstoredtra

6、iningexamples.Thisisincontrasttoeagerlearningmethodslikedecisiontreelearning[26],ruleinduction[9],orInductiveLogicProgramming[7],whichabstractageneralizedstructurefromthetrainingsetbeforehand(forgettingtheexamplesthemselves),andusethattoderiveaclassicationforanewin

7、stance.InMBL,adistancemetriconthefeaturespacedeneswhatarethenearestneighborsofaninstance.Metricswithfeatureweightsbasedoninformation-theoryorotherrelevancestatisticsallowustouserichrepre-sentationsofinstancesandtheircontext,andtobalancethein uencesofdiverseinformat

8、ionsourcesincomputingdistance.NaturalLanguageProcessing(NLP)taskstypicallyconcernthemappingofaninputrepresentation(e.g.,aseriesofwords)int