资源描述：
《Tasks in NLP domain POS Tagging》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、AmoghAsgekar(06329006)JeevanChalke(06329011)VinayDeshpande(06305001)JubinChheda(06305003)OutlineNLPtasksTypesofdomainadaptationSampleselectionbiasStructuralcorrespondinglearningAdaptationbyfeatureaugmentationConclusionTasksinNLPdomainPOSTaggingAssignPOStagstothewordsinagiventes
2、tcorpus.ParsingConstructastructureoutofthegivensentenceformations.WordsensedisambiguationSelectaparticularmeaningofthewordfromvariouspossibilities.NamedentityrecognitionIdentifyingnamedentities(names,addressetc)fromagivencorpus.DomainAdaptationThepre-mentionedtasksareperformedby
3、“learning”fromacorpusandthenapplyingtheknowledgetoclassifythetestinstances.Incasethetrainingdistributionsandtestdistributionsaredifferent,thentheclassifiertendstoperformerroneously.Insuchcases,classifierneedstobedomainadaptedtoperformaccuratelyonboththedomains.APOStaggingtaskConside
4、rthefollowingexample:-LearnerhasaccesstoLabeleddataSrandomlysampledfromthetrainingdistributionPS.UnlabelledsampleTsampledfromanunknowntestdistributionPT.TaskofthelearneristopredictslabelsofpointsgeneratedandlabeledaccordingtoP.TTypesofDomainAdaptationAnalysethecausesfordomaindiver
5、genceandmodelthemintothelearnerSampleselectionbiasDiscoverthedivergenceofthedistributionsduringtrainingStructuralCorrespondenceLearningFeatureAugmentationModelSampleSelectionBiasWhatisSampleSelectionBias?Samples(x,y,s)aredrawnindependentlyfromadomain(X×Y×S)withdistributionD.Sisab
6、inaryspace.Ifs=1,thatinstanceisselected.Fourcasesofdependenceof(x,y)ons:1.s⊥xands⊥y2.s⊥y
7、x3.s⊥x
8、y4.sdependsonbothxandySampleSelectionBiasCorrectionCase:s⊥y
9、xi.e.Thesub-domainselectiondependsonlyonthewordsandnotontheirPOS-tags.NowifDistheoriginaldistributionofdomainandD’isthedistribu
10、tionofselectedsub-domainthen,wecanconvertfromonedomaintootherusingamultiplierPr(s=)1β(X)=Pr(s=
11、1X)Thus,D(x,y,s)=β(X)*D’(x,y,s)•ThepriorprobabilitiesPr(s=1)andPr(s=1
12、x)mustbeknown.•Pr(s=1
13、x)shouldbenon-zeroforeachxi.e.atleastoneinstanceofeachwordshouldbeselected.SampleSelectionBiasinP
