文档介绍:该【Conditional Random Fields - Department of Computer Science 条件随机域-计算机科学系 】是由【核辐射】上传分享,文档一共【14】页,该文档可以免费在线阅读,需要了解更多关于【Conditional Random Fields - Department of Computer Science 条件随机域-计算机科学系 】的内容,可以使用淘豆网的站内搜索功能,选择自己适合的文档,以下文字是截取该文章内的部分文字,如需要获得完整电子版,请下载此文档到您的设备,方便您编辑和打印。ConditionalRandomFields-DepartmentofComputerScience条件随机域-计算机科学系
Experiment1-Results
Model
PhoneAccuracy
PhoneCorrect
Tandem[monophone]
%
%
Tandem[triphone]
%
%
CRF[monophone]
%
%
CRFsystemtrainedonmonophoneswiththesefeaturesachievesaccuracysuperiortoHMMonmonophones
CRFcomesclosetoachievingHMMtriphoneaccuracy
Experiment2
Goals:
ApplyCRFmodeltophoneclassifierdata
ApplyCRFmodeltocombinedphonologicalfeatureclassifierdataandphoneclassifierdata
Performphonerecognition
CompareresultstothoseobtainedviaaTandemHMMsystem
Experiment2-Results
Model
PhoneAcc
PhoneCorrect
Tandem[mono](phones)
%
%
Tandem[tri](phones)
%
%
CRF[mono](phones)
%
%
Tandem[mono](phones/feas)
%
%
Tandem[tri](phones/feas)
%
%
CRF[mono](phones/feas)
%
%
NotethatTandemHMMresultisbestresultwithonlytop39featuresfollowingaprincipalcomponentsanalysis
Experiment3
Goal:
PreviousCRFexperimentsusedphoneposteriorsforCRF,andlinearoutputstransformedviaaKarhunen-Loeve(KL)transformfortheHMMsytem
ThistransformationisneededtoimprovetheHMMperformancethroughdecorellationofinputs
UsingthesamelinearoutputsastheHMMsystem,doourresultschange?
Experiment3-Results
Model
PhoneAccuracy
PhoneCorrect
CRF(phones)posteriors
%
%
CRF(phones)linearKL
%
%
CRF(phones)post.+linear
%
%
CRF(features)posteriors
%
%
CRF(features)linearKL
%
%
CRF(features)post+linear
%
%
CRF(features)linear(noKL)
%
%
Alsoshown–Addingbothfeaturesetstogetherandgivingthesystemsupposedlyredundantinformationleadstoagaininaccuracy
Experiment4
Goal:
PreviousCRFexperimentsdidnotallowforrealignmentofthetraininglabels
BoundariesforlabelsprovidedbyTIMIThandtranscribersusedthroughouttraining
HMMsystemsallowedtoshiftboundariesduringEMlearning
Ifweallowforrealignmentinourtrainingprocess,canweimprovetheCRFresults?
谢谢!