Topic segmentation with an aspect hidden Markov model(10)

We present a novel probabilistic method for partially unsupervised topic segmentation on unstructured text. Previous approaches to this problem utilize the hidden Markov model framework (HMM). The HMM treats a document as mutually independent sets of words

5.1AspectmodelEMtraining7

Figure3:TemperedEMconvergenceintheATCandNYTcorpora

Withintheseshowsthereare4,917segmentswithavocabularyof35,777uniqueterms.Theshowsconstituteabout4millionwords.Weestimatedtheworderrorrateinthiscorporatobeinthe30%to40%range.Notethattheseareonlyestimatescomputedfromsamplingthecorporaasperfecttranscriptsareunavailabletous.

Additionally,weanalyzedacorpusof3,830articlesfromtheNewYorkTimes(NYT)tocomparetheASRperformancewitherror-freetext.Thiscorpusconstitutesabout4millionwordswithavocabularyof70,792uniqueterms.Inallreportedexper-iments,welearnanaspectmodelwith20hiddenfactors.

5.1AspectmodelEMtraining

Figure3illustratestheperformanceonheldoutdataduringthetemperedEMtrainingoftheaspectmodel(seesection4.1).ThoughtheNYTcorpustakeslongertoconverge(duetothehighervocabularysize),itlearnsmorequicklythantheATCcorpussincethetextcontainsnoerrors.TheATCconvergesfaster(duetothesmallervocabularysize)butstaysatalowCoAP(seesection5.3)forseveraliterationsbeforeperformanceimproves.

5.2Sampleresultsandtopiclabels

Inourexperiments,weusethreevariantsofourtwocorpora.First,wecreaterandomsequencesofsegmentsfromtheATCcorpus.Second,wecreaterandomsequencesfromtheNYTcorpustocomparecleanversusnoisysegmentation.Finally,weusetheactualairedsequencesofATCsegmentssincethisisdomainoftheprimaryproblemwhichwearetryingtotackle.

Intherandomsequencesofsegments,weattainalmostperfectsegmentationonbothcorpora.However,theresultsaremixedwiththeoriginalbroadcastsoftheATC.