资源描述:
《Maximum Entropy Markov Models for Information Extraction and Segmentation》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、MaximumEntropyMarkovModelsforInformationExtractionandSegmentationAndrewMcCallumMCCALLUM@JUSTRESEARCH.COMDayneFreitagDAYNE@JUSTRESEARCH.COMJustResearch,4616HenryStreet,Pittsburgh,PA15213USAFernandoPereiraPEREIRA@RESEARCH.ATT.COMAT&TLabs-Research,180ParkAve,FlorhamPark,N
2、J07932USAAbstractforstate-transitionprobabilitiesandstate-specificobserva-tionprobabilities.GreatlycontributingtotheirpopularityHiddenMarkovmodels(HMMs)areapowerfulistheavailabilityofstraightforwardproceduresfortrain-probabilistictoolformodelingsequentialdata,ingbymaxim
3、umlikelihood(Baum-Welch)andforusingandhavebeenappliedwithsuccesstomanythetrainedmodelstofindthemostlikelyhiddenstatese-text-relatedtasks,suchaspart-of-speechtagging,quencecorrespondingtoanobservationsequence(Viterbi).textsegmentationandinformationextraction.Inthesecases
4、,theobservationsareusuallymod-Intext-relatedtasks,theobservationprobabilitiesaretyp-eledasmultinomialdistributionsoveradiscreteicallyrepresentedasamultinomialdistributionoveradis-vocabulary,andtheHMMparametersaresetcrete,finitevocabularyofwords,andBaum-Welchtrainingtoma
5、ximizethelikelihoodoftheobservations.isusedtolearnparametersthatmaximizetheprobabilityofThispaperpresentsanewMarkoviansequencetheobservationsequencesinthetrainingdata.model,closelyrelatedtoHMMs,thatallowsob-Therearetwoproblemswiththistraditionalapproach.servationstober
6、epresentedasarbitraryoverlap-First,manytaskswouldbenefitfromaricherrepresenta-pingfeatures(suchasword,capitalization,for-tionofobservations—inparticulararepresentationthatde-matting,part-of-speech),anddefinesthecondi-scribesobservationsintermsofmanyoverlappingfeatures,ti
7、onalprobabilityofstatesequencesgivenob-suchascapitalization,wordendings,part-of-speech,for-servationsequences.Itdoesthisbyusingthematting,positiononthepage,andnodemembershipsinmaximumentropyframeworktofitasetofexpo-WordNet,inadditiontothetraditionalwordidentity.Fornenti
8、almodelsthatrepresenttheprobabilityofaexample,whentryingtoextractpreviouslyunseencom-stategivenanobservationandthepre