THE ELECTRICAL ENGINEERING AND APPLIED SIGNAL PROCESSING SERIES Edited by Alexander Poularikas The Advanced Signal Proce...
55 downloads
1061 Views
5MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
THE ELECTRICAL ENGINEERING AND APPLIED SIGNAL PROCESSING SERIES Edited by Alexander Poularikas The Advanced Signal Processing Handbook: Theory and Implementation for Radar, Sonar, and Medical Imaging Real-Time Systems Stergios Stergiopoulos The Transform and Data Compression Handbook K.R. Rao and P.C. Yip Handbook of Multisensor Data Fusion David Hall and James Llinas Handbook of Neural Network Signal Processing Yu Hen Hu and Jenq-Neng Hwang Handbook of Antennas in Wireless Communications Lal Chand Godara Noise Reduction in Speech Applications Gillian M. Davis Signal Processing Noise Vyacheslav P. Tuzlukov Digital Signal Processing with Examples in MATLAB® Samuel Stearns Applications in Time-Frequency Signal Processing Antonia Papandreou-Suppappola The Digital Color Imaging Handbook Gaurav Sharma Pattern Recognition in Speech and Language Processing Wu Chou and Biing Huang Juang
Forthcoming Titles Propagation Data Handbook for Wireless Communication System Design Robert Crane Smart Antennas Lal Chand Godara Nonlinear Signal and Image Processing: Theory, Methods, and Applications Kenneth Barner and Gonzalo R. Arce
E\&5&3UHVV//&
Forthcoming Titles (continued) Soft Computing with MATLAB® Ali Zilouchian Signal and Image Processing Navigational Systems Vyacheslav P. Tuzlukov Wireless Internet: Technologies and Applications Apostolis K. Salkintzis and Alexander Poularikas
E\&5&3UHVV//&
PATTERN RECOGNITION in SPEECH and LANGUAGE PROCESSING Edited by
WU CHOU Avaya Labs Research
BIING HWANG JUANG Georgia Institute of Technology
CRC PR E S S Boca Raton London New York Washington, D.C. E\&5&3UHVV//&
/LEUDU\RI&RQJUHVV&DWDORJLQJLQ3XEOLFDWLRQ'DWD 3DWWHUQUHFRJQLWLRQLQVSHHFKDQGODQJXDJHSURFHVVLQJHGLWHGE\:X&KRXDQG %LLQJ+ZDQJ-XDQJ SFP ,QFOXGHVELEOLRJUDSKLFDOUHIHUHQFHVDQGLQGH[ ,6%1DONSDSHU $XWRPDWLFVSHHFKUHFRJQLWLRQ3DWWHUQUHFRJQLWLRQV\VWHPV,&KRX:X,, -XDQJ%+%LLQJ+ZDQJ 7.63 d³GF
7KLVERRNFRQWDLQVLQIRUPDWLRQREWDLQHGIURPDXWKHQWLFDQGKLJKO\UHJDUGHGVRXUFHV5HSULQWHGPDWHULDO LVTXRWHGZLWKSHUPLVVLRQDQGVRXUFHVDUHLQGLFDWHG$ZLGHYDULHW\RIUHIHUHQFHVDUHOLVWHG5HDVRQDEOH HIIRUWVKDYHEHHQPDGHWRSXEOLVKUHOLDEOHGDWDDQGLQIRUPDWLRQEXWWKHDXWKRUVDQGWKHSXEOLVKHUFDQQRW DVVXPHUHVSRQVLELOLW\IRUWKHYDOLGLW\RIDOOPDWHULDOVRUIRUWKHFRQVHTXHQFHVRIWKHLUXVH 1HLWKHUWKLVERRNQRUDQ\SDUWPD\EHUHSURGXFHGRUWUDQVPLWWHGLQDQ\IRUPRUE\DQ\PHDQVHOHFWURQLF RUPHFKDQLFDOLQFOXGLQJSKRWRFRS\LQJPLFURÀOPLQJDQGUHFRUGLQJRUE\DQ\LQIRUPDWLRQVWRUDJHRU UHWULHYDOV\VWHPZLWKRXWSULRUSHUPLVVLRQLQZULWLQJIURPWKHSXEOLVKHU $OO ULJKWV UHVHUYHG$XWKRUL]DWLRQ WR SKRWRFRS\ LWHPV IRU LQWHUQDO RU SHUVRQDO XVH RU WKH SHUVRQDO RU LQWHUQDO XVH RI VSHFLÀF FOLHQWV PD\ EH JUDQWHG E\ &5& 3UHVV //& SURYLGHG WKDW SHU SDJH SKRWRFRSLHGLVSDLGGLUHFWO\WR&RS\ULJKW&OHDUDQFH&HQWHU5RVHZRRG'ULYH'DQYHUV0$ 86$ 7KH IHH FRGH IRU XVHUV RI WKH 7UDQVDFWLRQDO 5HSRUWLQJ 6HUYLFH LV ,6%1 7KHIHHLVVXEMHFWWRFKDQJHZLWKRXWQRWLFH)RURUJDQL]DWLRQVWKDWKDYHEHHQJUDQWHG DSKRWRFRS\OLFHQVHE\WKH&&&DVHSDUDWHV\VWHPRISD\PHQWKDVEHHQDUUDQJHG 7KHFRQVHQWRI&5&3UHVV//&GRHVQRWH[WHQGWRFRS\LQJIRUJHQHUDOGLVWULEXWLRQIRUSURPRWLRQIRU FUHDWLQJQHZZRUNVRUIRUUHVDOH6SHFLÀFSHUPLVVLRQPXVWEHREWDLQHGLQZULWLQJIURP&5&3UHVV//& IRUVXFKFRS\LQJ 'LUHFWDOOLQTXLULHVWR&5&3UHVV//&1:&RUSRUDWH%OYG%RFD5DWRQ)ORULGD 7UDGHPDUN1RWLFH 3URGXFWRUFRUSRUDWHQDPHVPD\EHWUDGHPDUNVRUUHJLVWHUHGWUDGHPDUNVDQGDUH XVHGRQO\IRULGHQWLÀFDWLRQDQGH[SODQDWLRQZLWKRXWLQWHQWWRLQIULQJH
9LVLWWKH&5&3UHVV:HEVLWHDWZZZFUFSUHVVFRP E\&5&3UHVV//& 1RFODLPWRRULJLQDO86*RYHUQPHQWZRUNV ,QWHUQDWLRQDO6WDQGDUG%RRN1XPEHU /LEUDU\RI&RQJUHVV&DUG1XPEHU 3ULQWHGLQWKH8QLWHG6WDWHVRI$PHULFD 3ULQWHGRQDFLGIUHHSDSHU
E\&5&3UHVV//&
Preface
#RRTQCEJGU VQ VJG RTQDNGOU QH FGUKIPKPI URGGEJ CPF NCPIWCIG RTQEGUUKPI CNIQTKVJOU HQT JWOCP OCEJKPG EQOOWPKECVKQP WUGF VQ DG VCMGP HTQO VJG RGTURGEVKXGU QH NKP IWKUVKEU CPF URGGEJ UEKGPEG WPVKN VJG NCVG U &WG VQ VJG CFXCPEGU KP EQORWV KPI CPF UVCVKUVKECN OQFGNKPI FCVC FTKXGP RCVVGTP TGEQIPKVKQP OGVJQFU JCXG DGEQOG C HCUV OQXKPI TGUGCTEJ CTGC FWTKPI VJG RCUV VYQ FGECFGU CPF EQPVTKDWVGF OWEJ VQ VJG RTQITGUU KP VJKU ſGNF #U VJG GTC QH KPHQTOCVKQP CIG EQPVKPWGU VQ FGXGNQR YG YKVPGUU CP GXGT KPETGCUKPI PGGF KP KPVGNNKIGPV JWOCPOCEJKPG EQOOWPKECVKQPU CU YGNN CU VJG ETGCVKQP QH OCEJKPG WPFGTUVCPFCDNG OGVCFCVC HQT 9GD EQPVGPV CPF QVJGT KPHQTOCVKQP UQWTEGU 6JKU JCPFDQQM KU VQ ſNN VJG PGGF QH C U[UVGOCVKE CPF WRVQFCVG RTGUGPVCVKQP QH PGY RCVVGTP TGEQIPKVKQP CRRTQCEJGU KP URGGEJ CPF NCPIWCIG RTQEGUUKPI 6JG DQQM UVCTVU YKVJ HWPFCOGPVCNU CPF TGEGPV VJGQTGVKECN CFXCPEGU KP RCVVGTP TGEQI PKVKQP YKVJ CP GORJCUKU QP ENCUUKſGT FGUKIP ETKVGTKC CPF QRVKOK\CVKQP RTQEGFWTGU +V EQXGTU UGXGTCN TGEGPV TGUGCTEJ CFXCPEGU KP VJKU CTGC UWEJ CU VJG OKPKOWO GTTQT TCVG /%' OGVJQF VJG OKPKOWO $C[GU TKUM CRRTQCEJ CFCRVKXG U[UVGO FGUKIP CPF FGEKUKQP TWNGU PGWTCN PGVYQTMU FKUVTKDWVGF TGEQIPK\GTU CPF FGEKUKQP HWUKQP 6JGUG OGVJQFU FGRCTV HTQO VJG EQPXGPVKQPCN RCTCFKIO YJKEJ NKPMU C ENCUUKſGT FGUKIP VQ VJG ENCUUKECN RTQDNGO QH FKUVTKDWVKQP GUVKOCVKQP +PUVGCF OQTG OGCPKPIHWN ETKVGTKC CTG KPVTQFWEGF YJKEJ UKIPKſECPVN[ KORTQXG VJG FKUETKOKPCVKQP RQYGT QH C ENCUUKſGT RCTVKEWNCTN[ YJGP CRRNKGF VQ URGGEJ RTQDNGOU KP YJKEJ VJG PQVKQP QH FCVC FKUVTKDWVKQP KU FKHſEWNV VQ TGCNK\G 6JG UGEQPF RCTV QH VJG DQQM KU VJGTGHQTG URGEKCNN[ HQEWUGF QP VJG CRRTQCEJGU CPF OGVJQFU CRRNKGF VQ URGGEJ RTQEGUUKPI +V EQXGTU VQRKEU UWEJ CU $C[GU OKPKOWO TKUM CRRTQCEJ VQ URGGEJ TGEQIPKVKQP NCTIG XQECDWNCT[ URGGEJ TGEQIPKVKQP DCUGF QP UVCVKU VKECN OGVJQFU TGEQIPKVKQP QH URQPVCPGQWU URGGEJ KP FKCNQIWG KPVGTCEVKQP URGGEJ CPF URGCMGT XGTKſECVKQP CPF CWFKQ KPHQTOCVKQP TGVTKGXCN CPF KPFGZKPI 6JGUG EJCRVGTU RTQXKFG C EQORTGJGPUKXG EQXGTCIG QH TGEGPV CFXCPEGU KP CRRN[KPI RCVVGTP TGEQIPK VKQP VQ TGCN U[UVGOU KP URGGEJ CPF CWFKQ RTQEGUUKPI 6JG VJKTF RCTV QH VJG DQQM KU FGXQVGF VQ VQRKEU QH RCVVGTP TGEQIPKVKQP KP NCPIWCIG RTQEGUUKPI +V EQPVCKPU EJCRVGTU KP NCPIWCIG OQFGNKPI DCUGF QP NCVGPV UGOCPVKE KP FGZKPI UCNKGPV KPHQTOCVKQP TGRTGUGPVCVKQP CPF RTQEGUUKPI KP PCVWTCN NCPIWCIG FKC NQIWG U[UVGO UVCVKUVKECN OCEJKPG VTCPUNCVKQP OGVJQFU KP VQRKE FGVGEVKQP VTCEMKPI CPF PCOG KFGPVKV[ KFGPVKſECVKQP 6JGUG VQRKEU CTG PGY VTGPFU KP NCPIWCIG RTQEGUU KPI CPF UKIPKſECPV RTQITGUU JCU DGGP OCFG KP TGEGPV [GCTU +V JCU C FKTGEV KORCEV VQ VJG RTCEVKEG CPF KORNGOGPVCVKQP QH KPHQTOCVKQP RTQEGUUKPI U[UVGOU HQT 9GD EQPVGPV DTQCFECUV PGYU CPF QVJGT EQPVGPVTKEJ KPHQTOCVKQP TGUQWTEGU 6JKU DQQM KU C EQNNGEVKXG GHHQTV OQVKXCVGF D[ VJG GZEKVGOGPV QH VJG PGY CFXCPEGU KP E\&5&3UHVV//&
VJKU ſGNF CPF VJG WTIGPV PGGF VQ DTKPI VJGUG CFXCPEGU VQ C IGPGTCN CWFKGPEG 6JG EQP VTKDWVKPI CWVJQTU QH VJKU DQQM CTG NGCFKPI GZRGTVU KP VJG ſGNF QH URGGEJ CPF NCPIWCIG RTQEGUUKPI #VVGORVU CTG OCFG VQ OCMG GCEJ EJCRVGT UGNHEQPVCKPGF CPF EQORTG JGPUKDNG HQT TGCFGTU YKVJ IGPGTCN DCEMITQWPF KP RCVVGTP TGEQIPKVKQP CPF KPHQTOCVKQP RTQEGUUKPI +V KU KPVGPFGF VQ DG C JCPFDQQM QT TGHGTGPEG VGZVDQQM HQT TGUGCTEJGTU ITCFWCVG UVWFGPVU CPF CFXCPEGF WPFGTITCFWCVG UVWFGPVU YJQ YCPV VQ HQNNQY VJG PGY CFXCPEGU KP RCVVGTP TGEQIPKVKQP 5WHſEKGPV TGHGTGPEGU CTG RTQXKFGF CV VJG GPF QH GCEJ EJCRVGT VQ UGTXG CU CP GPVT[ RQKPV HQT CP KPVGTGUVGF TGCFGT VQ RWTUWG HWTVJGT 9G YQWNF NKMG VQ VJCPM CNN EQPVTKDWVQTU QH VJKU DQQM 9KVJQWV VJGKT EQOOKVOGPV CPF SWCNKV[ QH YQTM VJKU DQQM YQWNF PQV DG RQUUKDNG 9G CRRTGEKCVG VJG UWRRQTV CPF GPEQWTCIGOGPV HTQO QWT EQNNGCIWGU CV #XC[C .CDU 4GUGCTEJ FWTKPI VJG RTGRCTCVKQP QH VJKU DQQM +V YCU C RNGCUCPV YQTMKPI GZRGTKGPEG YKVJ %4% 2TGUU VJGKT VGEJPKECN UWRRQTV YCU XGT[ JGNRHWN VQ WU
9W %JQW $KKPI*YCPI ,WCPI Basking Ridge, New Jersey September, 2002
E\&5&3UHVV//&
Contributors
A. Abella 5RGGEJ 4GUGCTEJ #66 .CDQTCVQTKGU (NQTJCO 2CTM 0, James Allan &GRCTVOGPV QH %QORWVGT 5EKGPEG 7PKXGTUKV[ QH /CUUCEJWUGVVU #OJGTUV #OJGTUV /# T. Alonso 5RGGEJ 4GUGCTEJ #66 .CDQTCVQTKGU (NQTJCO 2CTM 0, Jerome R. Bellegarda 5RQMGP .CPIWCIG )TQWR #RRNG %QORWVGT +PE %WRGTVKPQ %# William Byrne %GPVGT HQT .CPIWCIG CPF 5RGGEJ 2TQEGUUKPI ,QJPU *QRMKPU 7PKXGTUKV[ $CNVKOQTG /&
7PKXGTUKVyG FG 2CTKU 5WF 1TUC[ %GFGZ (TCPEG Vaibhava Goel 6, 9CVUQP 4GUGCTEJ %GPVGT +$/ ;QTMVQYP *GKIJVU 0; Allen L. Gorin 5RGGEJ 4GUGCTEJ #66 .CDQTCVQTKGU (NQTJCO 2CTM 0, Qiang Huo &GRCTVOGPV QH %QORWVGT 5EKGPEG CPF +PHQTOCVKQP 5[UVGOU 6JG 7PKXGTUKV[ QH *QPI -QPI *QPI -QPI %JKPC Biing-Hwang Juang #XC[C .CDU 4GUGCTEJ $CUMKPI 4KFIG 0,
Wu Chou #XC[C .CDU 4GUGCTEJ $CUMKPI 4KFIG 0,
Shigeru Katagiri +PVGNNKIGPV %QOOWPKECVKQP .CDQTCVQT[ CPF 5RGGEJ 1RGP .CDQTCVQT[ 0KRRQP 6GNGITCRJ CPF 6GNGRJQPG %QTRQTCVKQP 6QM[Q ,CRCP
Sadaoki Furui &GRCTVOGPV QH %QORWVGT 5EKGPEG 6QM[Q +PUVKVWVG QH 6GEJPQNQI[ 6QM[Q ,CRCP
Lori Lamel .+/5+%045 7PKXGTUKVyG FG 2CTKU 5WF 1TUC[ %GFGZ (TCPEG
Jean-Luc Gauvain .+/5+%045
Qi (Peter) Li $GNN .CDQTCVQTKGU
E\&5&3UHVV//&
.WEGPV 6GEJPQNQIKGU /WTTC[ *KNN 0, John Makhoul $$0 6GEJPQNQIKGU %CODTKFIG /# Hermann Ney .GJTUVWJN HWGT +PHQTOCVKM 8+ *WOCP .CPIWCIG 6GEJPQNQI[ CPF 2CVVGTP 4GEQIPKVKQP %QORWVGT 5EKGPEG &GRCTVOGPV 7PKXGTUKV[ QH 6GEJPQNQI[ #CEJGP )GTOCP[ F. J. Och .GJTUVWJN HWGT +PHQTOCVKM 8+ *WOCP .CPIWCIG 6GEJPQNQI[ CPF 2CVVGTP 4GEQIPKVKQP %QORWVGT 5EKGPEG &GRCTVOGPV 7PKXGTUKV[ QH 6GEJPQNQI[ #CEJGP )GTOCP[
E\&5&3UHVV//&
G. Riccardi 5RGGEJ 4GUGCTEJ #66 .CDQTCVQTKGU (NQTJCO 2CTM 0, Richard M. Schwartz $$0 6GEJPQNQIKGU %CODTKFIG /# J. H. Wright 5RGGEJ 4GUGCTEJ #66 .CDQTCVQTKGU (NQTJCO 2CTM 0,
Contents
1 Minimum Classification Error (MCE) Approach in Pattern Recognition Wu Chou #XC[C .CDU 4GUGCTEJ #XC[C +PE 75# +PVTQFWEVKQP 1RVKOCN %NCUUKſGT HTQO $C[GU &GEKUKQP 6JGQT[ &KUETKOKPCPV (WPEVKQP #RRTQCEJ VQ %NCUUKſGT &GUKIP 5RGGEJ 4GEQIPKVKQP CPF *KFFGP /CTMQX /QFGNKPI *KFFGP /CTMQX /QFGNKPI QH 5RGGEJ /%' %NCUUKſGT &GUKIP 7UKPI &KUETKOKPCPV (WPEVKQPU /%' %NCUUKſGT &GUKIP 5VTCVGI[ 1RVKOK\CVKQP /GVJQFU 1VJGT 1RVKOK\CVKQP /GVJQFU *// CU C &KUETKOKPCPV (WPEVKQP 4GNCVKQP DGVYGGP /%' CPF //+ &KUEWUUKQPU CPF %QOOGPVU 'ODGFFGF 5VTKPI /QFGN $CUGF /%' 6TCKPKPI 5VTKPI /QFGN $CUGF /%' #RRTQCEJ %QODKPGF 5VTKPI /QFGN $CUGF /%' #RRTQCEJ &KUETKOKPCVKXG (GCVWTG 'ZVTCEVKQP 8GTKſECVKQP CPF +FGPVKſECVKQP 5RGCMGT 8GTKſECVKQP CPF +FGPVKſECVKQP 7VVGTCPEG 8GTKſECVKQP 5WOOCT[ 2 Minimum Bayes-Risk Methods in Automatic Speech Recognition Vaibhava Goel £ and William ByrneÝ £ +$/ Ý ,QJPU *QRMKPU 7PKXGTUKV[ /KPKOWO $C[GU4KUM %NCUUKſECVKQP (TCOGYQTM .KMGNKJQQF 4CVKQ $CUGF *[RQVJGUKU 6GUVKPI /CZKOWO #2QUVGTKQTK 2TQDCDKNKV[ %NCUUKſECVKQP 2TGXKQWU 5VWFKGU QH #RRNKECVKQP 5GPUKVKXG #54 2TCEVKECN /$4 2TQEGFWTGU HQT #54 5WOOCVKQP QXGT *KFFGP 5VCVG 5GSWGPEGU /$4 4GEQIPKVKQP YKVJ 0DGUV .KUVU /$4 4GEQIPKVKQP YKVJ .CVVKEGU 5GIOGPVCN /$4 2TQEGFWTGU 5GIOGPVCN 8QVKPI 418'4 E\&5&3UHVV//&
G418'4 'ZRGTKOGPVCN 4GUWNVU 2CTCOGVGT 6WPKPI YKVJKP VJG /$4 %NCUUKſECVKQP 4WNG 7VVGTCPEG .GXGN /$4 9QTF CPF -G[YQTF 4GEQIPKVKQP 418'4 CPF G418'4 HQT /WNVKNKPIWCN #54 5WOOCT[ #EMPQYNGFIGOGPVU
3 A Decision Theoretic Formulation for Robust Automatic Speech Recognition Qiang Huo 6JG 7PKXGTUKV[ QH *QPI -QPI *QPI -QPI %JKPC +PVTQFWEVKQP 1RVKOCN $C[GUŏ &GEKUKQP 4WNG HQT #54 #FCRVKXG &GEKUKQP 4WNGU %QPUVTWEVGF HTQO 6TCKPKPI 5CORNGU 2NWIKP $C[GUŏ &GEKUKQP 4WNGU YKVJ /CZKOWONKMGNKJQQF &GPUKV[ 'UVKOCVG /CZKOWO&KUETKOKPCPV &GEKUKQP 4WNGU /KPKOK\KPI VJG 'O RKTKECN %NCUUKſECVKQP 'TTQT &KUEWUUKQP 8KQNCVKQPU QH /QFGNKPI #UUWORVKQPU KP #54 6[RGU QH &KUVQTVKQPU 6QYCTFU #FCRVKXG CPF 4QDWUV #54 +ORTQXKPI #FCRVKXG &GEKUKQP 4WNGU XKC &GEKUKQP 2CTCOGVGT #FCRVC VKQP &GEKUKQP 2CTCOGVGT #FCRVCVKQP HQT 5VCVKQPCT[ 1RGTCVKPI %QP FKVKQPU &GEKUKQP 2CTCOGVGT #FCRVCVKQP HQT 5NQYN[ %JCPIKPI 1RGT CVKPI %QPFKVKQPU &GEKUKQP 2CTCOGVGT #FCRVCVKQP HQT 5YKVEJKPI 1RGTCVKPI %QP FKVKQPU &KUEWUUKQP 4QDWUV &GEKUKQP 4WNGU &GEKUKQP 4WNG 4QDWUVPGUU /KPKOCZ %NCUUKſECVKQP 4WNG $C[GUKCP 2TGFKEVKXG %NCUUKſECVKQP 4WNG &KUEWUUKQP 5WOOCT[ 4 Speech Pattern Recognition using Neural Networks Shigeru Katagiri 066 %QOOWPKECVKQP 5EKGPEG .CDQTCVQTKGU +PVTQFWEVKQP $C[GU &GEKUKQP 6JGQT[ 2TGRCTCVKQPU &GEKUKQP 4WNG /KPKOWO 'TTQTTCVG %NCUUKſECVKQP E\&5&3UHVV//&
2TQDCDKNKV[ (WPEVKQP 'UVKOCVKQP &KUETKOKPCVKXG 6TCKPKPI 5RGGEJ 4GEQIPK\GTU $CUGF QP 0GWTCN 0GVYQTMU 2TGRCTCVKQPU %NCUUKſECVKQP 'TTQT /KPKOK\CVKQP 5SWCTGF 'TTQT /KPKOK\CVKQP %TQUU 'PVTQR[ /KPKOK\CVKQP (WUKQP QH /WNVKRNG %NCUUKſECVKQP &GEKUKQPU 2TKPEKRNGU 'ZCORNGU QH 'ODQFKOGPV %QPENWFKPI 4GOCTMU #RRGPFKZ /CZKOK\KPI /WVWCN +PHQTOCVKQP
5 Large Vocabulary Speech Recognition Based on Statistical Methods Jean-Luc Gauvain and Lori Lamel .+/5+ (TCPEG +PVTQFWEVKQP 1XGTXKGY .CPIWCIG /QFGNKPI 6GZV 2TGRCTCVKQP 8QECDWNCT[ 5GNGEVKQP 0ITCO 'UVKOCVKQP ./ #FCRVCVKQP 2TQPWPEKCVKQP /QFGNKPI #EQWUVKE /QFGNKPI #EQWUVKE (TQPVGPF /QFGNKPI #NNQRJQPGU *// 2CTCOGVGT 'UVKOCVKQP *// #FCRVCVKQP &GEQFKPI 5RGGEJ0QPURGGEJ &GVGEVKQP &GEQFKPI 5VTCVGIKGU 'HſEKGPE[ %QPſFGPEG /GCUWTGU +PFKECVKXG 2GTHQTOCPEG .GXGNU &KEVCVKQP 5RGGEJ 4GEQIPKVKQP HQT &KCNQI 5[UVGOU 6TCPUETKRVKQP HQT #WFKQ +PFGZCVKQP 2QTVCDKNKV[ CPF .CPIWCIG &GRGPFGPEKGU 6 Toward Spontaneous Speech Recognition and Understanding Sadaoki Furui 6QM[Q +PUVKVWVG QH 6GEJPQNQI[ +PVTQFWEVKQP (QWT %CVGIQTKGU QH 5RGGEJ 4GEQIPKVKQP 6CUMU 5RQPVCPGQWU 5RGGEJ 4GEQIPKVKQP CPF 7PFGTUVCPFKPI 4GXKGY %CVGIQT[ + JWOCPVQJWOCP FKCNQIWG E\&5&3UHVV//&
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
7 Speaker Authentication Qi Li£ and Biing-Hwang Juang Ý £ $GNN .CDU Ý #XC[C .CDU 4GUGCTEJ +PVTQFWEVKQP 5RGCMGT 4GEQIPKVKQP CPF 8GTKſECVKQP 8GTDCN +PHQTOCVKQP 8GTKſECVKQP 2CVVGTP 4GEQIPKVKQP KP 5RGCMGT #WVJGPVKECVKQP $C[GUKCP &GEKUKQP 6JGQT[ 5VQEJCUVKE /QFGNU HQT 5VCVKQPCT[ 2TQEGUU 5VQEJCUVKE /QFGNU HQT 0QP5VCVKQPCT[ 2TQEGUU 5RGGEJ 5GIOGPVCVKQP 5VCVKUVKECN 8GTKſECVKQP 5RGCMGT 8GTKſECVKQP 5[UVGO 8GTDCN +PHQTOCVKQP 8GTKſECVKQP 7VVGTCPEG 5GIOGPVCVKQP 5WDYQTF *[RQVJGUKU 6GUVKPI %QPſFGPEG /GCUWTG %CNEWNCVKQP 5GSWGPVKCN 7VVGTCPEG 8GTKſECVKQP 8+8 'ZRGTKOGPVCN 4GUWNVU 5RGCMGT #WVJGPVKECVKQP D[ %QODKPKPI 58 CPF 8+8 E\&5&3UHVV//&
5WOOCT[
8 HMMs for Language Processing Problems Richard M. Schwartz and John Makhoul $$0 6GEJPQNQIKGU 8GTK\QP +PVTQFWEVKQP 7UG QH 2TQDCDKNKVKGU *KFFGP /CTMQX /QFGNU 0COG 5RQVVKPI 6QRKE %NCUUKſECVKQP 6JG /QFGN 'UVKOCVKPI *// 2CTCOGVGTU %NCUUKſECVKQP 'ZRGTKOGPVU +PHQTOCVKQP 4GVTKGXCN # $C[GUKCP /QFGN HQT +4 6TCKPKPI VJG +4 *// 2GTHQTOCPEG 'XGPV 6TCEMKPI 7PUWRGTXKUGF 6QRKE &GVGEVKQP 5WOOCT[ 9 Statistical Language Models With Embedded Latent Semantic Knowledge Jerome R. Bellegarda #RRNG %QORWVGT +PE +PVTQFWEVKQP 5EQRG .QECNKV[ 5[PVCEVKECNN[Ō&TKXGP 5RCP 'ZVGPUKQP 5GOCPVKECNN[Ō&TKXGP 5RCP 'ZVGPUKQP 1TICPK\CVKQP .CVGPV 5GOCPVKE #PCN[UKU (GCVWTG 'ZVTCEVKQP 5KPIWNCT 8CNWG &GEQORQUKVKQP )GPGTCN $GJCXKQT .5# (GCVWTG 5RCEG 9QTF %NWUVGTKPI 9QTF %NWUVGT 'ZCORNG &QEWOGPV %NWUVGTKPI &QEWOGPV %NWUVGT 'ZCORNG 5GOCPVKE %NCUUKſECVKQP (TCOGYQTM 'ZVGPUKQP 5GOCPVKE +PHGTGPEG %CXGCVU 0ITCO .5# .CPIWCIG /QFGNKPI .5# %QORQPGPV +PVGITCVKQP YKVJ 0ITCOU E\&5&3UHVV//&
%QPVGZV 5EQRG 5GNGEVKQP 5OQQVJKPI 9QTF 5OQQVJKPI &QEWOGPV 5OQQVJKPI ,QKPV 5OQQVJKPI 'ZRGTKOGPVU 'ZRGTKOGPVCN %QPFKVKQPU 'ZRGTKOGPVCN 4GUWNVU %QPVGZV 5EQRG 5GNGEVKQP +PJGTGPV 6TCFG1HHU %TQUU&QOCKP 6TCKPKPI &KUEWUUKQP %QPENWUKQP
10 Semantic Information Processing of Spoken Language – How May I sm Help You? A. L. Gorin, A. Abella, T. Alonso, G. Riccardi, and J. H. Wright, #66 .CDQTCVQ TKGU
+PVTQFWEVKQP %CNN%NCUUKſECVKQP .CPIWCIG /QFGNKPI HQT 4GEQIPKVKQP CPF 7PFGTUVCPFKPI &KCNQI %QPENWUKQPU
11 Machine Translation Using Statistical Modeling Herman Ney, and F. J. Och #CEJGP 7PKXGTUKV[ QH 6GEJPQNQI[ )GTOCP[ +PVTQFWEVKQP 5VCVKUVKECN &GEKUKQP 6JGQT[ CPF .KPIWKUVKEU 6JG 5VCVKUVKECN #RRTQCEJ $C[GU &GEKUKQP 4WNG HQT 9TKVVGP .CPIWCIG 6TCPUNCVKQP 4GNCVGF #RRTQCEJGU #NKIPOGPV CPF .GZKEQP /QFGNU %QPEGRV QH #NKIPOGPV /QFGNNKPI *KFFGP /CTMQX /QFGNU /QFGNU +$/ Ō 6TCKPKPI 5GCTEJ #NIQTKVJOKE &KHHGTGPEGU DGVYGGP 5RGGEJ 4GEQIPKVKQP CPF .CPIWCIG 6TCPUNCVKQP #NKIPOGPV 6GORNCVGU (TQO 5KPING 9QTFU VQ 9QTF )TQWRU %QPEGRV 6TCKPKPI 5GCTEJ 'ZRGTKOGPVCN 4GUWNVU 6JG 6CUM CPF VJG %QTRWU E\&5&3UHVV//&
1HƀKPG 4GUWNVU +PVGITCVKQP KPVQ VJG 8 '4$/1$+. 2TQVQV[RG 5[UVGO (KPCN 'XCNWCVKQP 5RGGEJ 6TCPUNCVKQP 6JG +PVGITCVGF #RRTQCEJ 2TKPEKRNG 2TCEVKECN +ORNGOGPVCVKQP 5WOOCT[ 4GHGTGPEGU 12 Modeling Topics for Detection and Tracking James Allan 7PKXGTUKV[ QH /CUUCEJWUGVVU #OJGTUV 6QRKE &GVGEVKQP CPF 6TCEMKPI 6QRKE CPF 'XGPVU 6&6 6CUMU %QTRQTC 'XCNWCVKQP $CUKE 6QRKE /QFGNU 8GEVQT 5RCEG .CPIWCIG /QFGNU +ORNGOGPVKPI VJG /QFGNU 0COGF 'PVKVKGU &QEWOGPV 'ZRCPUKQP %NWUVGTKPI 6KOG &GEC[ %QORCTKPI /QFGNU 0GCTGUV 0GKIJDQTU &GEKUKQP 6TGGU /QFGNVQ/QFGN /KUEGNNCPGQWU +UUWGU &GHGTTCN /WNVKOQFCN +UUWGU /WNVKNKPIWCN +UUWGU 7UKPI 6&6 +PVGTCEVKXGN[ &GOQPUVTCVKQPU 6KOGNKPGU /QFGNKPI 'XGPVU %QPENWUKQP
E\&5&3UHVV//&
1 Minimum Classification Error (MCE) Approach in Pattern Recognition Wu Chou Avaya Labs Research, Avaya Inc., USA
CONTENTS
+PVTQFWEVKQP 1RVKOCN %NCUUKſGT HTQO $C[GU &GEKUKQP 6JGQT[ &KUETKOKPCPV (WPEVKQP #RRTQCEJ VQ %NCUUKſGT &GUKIP 5RGGEJ 4GEQIPKVKQP CPF *KFFGP /CTMQX /QFGNKPI /%' %NCUUKſGT &GUKIP 7UKPI &KUETKOKPCPV (WPEVKQPU 'ODGFFGF 5VTKPI /QFGN $CUGF /%' 6TCKPKPI 8GTKſECVKQP CPF +FGPVKſECVKQP 5WOOCT[ #EMPQYNGFIGOGPV 4GHGTGPEGU
1.1 Introduction 2CVVGTP TGEQIPKVKQP KU C HCUV OQXKPI TGUGCTEJ CTGC 6JG CFXGPV QH RQYGTHWN EQORWV KPI FGXKEGU CPF VJG UWEEGUU QH UVCVKUVKECN CRRTQCEJGU UWEJ CU JKFFGP /CTMQX OQFGN HQT URGGEJ CPF NCPIWCIG RTQEGUUKPI VTKIIGTGF C TGPGYGF RWTUWKV HQT OQTG RQYGTHWN UVCVKUVKECN OGVJQFU VQ HWTVJGT TGFWEG VJG RCVVGTP TGEQIPKVKQP GTTQT TCVG CPF KORTQXG VJG TQDWUVPGUU QH VJG RCVVGTP ENCUUKſGT CETQUU XCTKQWU CFXGTUG EQPFKVKQPU #OQPI VJKU PGY RWTUWKV VJG WUG QH FKUETKOKPCPV HWPEVKQP OGVJQFU KP RCVVGTP TGEQIPKVKQP JCU GOGTIGF CU C RTQOKUKPI CRRTQCEJ CPF KV KU CRRNKGF UWEEGUUHWNN[ VQ URGGEJ CPF NCPIWCIG RTQEGUUKPI 6JKU EJCRVGT KU KPVGPFGF VQ RTQXKFG C TGXKUKV VQ VJG UVCVKUVKECN HQTOWNCVKQP QH VJG OKPKOWO ENCUUKſECVKQP GTTQT /%' DCUGF FKUETKOKPCVKXG OGVJ QFU KP URGGEJ CPF NCPIWCIG RTQEGUUKPI VCMG C ETKVKECN XKGY QH VJG CRRTQCEJ RTQXKFG C EQORTGJGPUKXG QXGTXKGY QH VJG ſGNF CPF JQRGHWNN[ KPURKTG QVJGT KPPQXCVKQPU VJCV YQWNF RQVGPVKCNN[ NGCF VQ PGY FKUETKOKPCVKXG OGVJQFU KP RCVVGTP TGEQIPKVKQP #NVJQWIJ VJG UVCVKUVKECN HQTOWNCVKQP QH /%' DCUGF FKUETKOKPCVKXG OGVJQFU JCU KVU TQQV KP VJG ENCUUKECN $C[GU FGEKUKQP VJGQT[ KV FGRCTVU HTQO VJG EQPXGPVKQPCN RCTCFKIO
6JKU EJCRVGT KU FGXGNQRGF DCUGF QP Œ&KUETKOKPCPVHWPEVKQPDCUGF OKPKOWO TGEQIPKVKQP GTTQT TCVG RCVVGTPTGEQIPKVKQP CRRTQCEJ VQ URGGEJ TGEQIPKVKQPŒ D[ 9W %JQW CRRGCTGF KP Proceedings of The IEEE 8QN 0Q E +'''
E\&5&3UHVV//&
YJKEJ NKPMU C TGEQIPKVKQP VCUM VQ VJG RTQDNGO QH FKUVTKDWVKQP GUVKOCVKQP +PUVGCF KV VCMGU C FKUETKOKPCPV HWPEVKQP DCUGF UVCVKUVKECN RCVVGTP ENCUUKſECVKQP CRRTQCEJ CPF HQT C IKXGP HCOKN[ QH FKUETKOKPCPV HWPEVKQP QRVKOCN ENCUUKſGTTGEQIPK\GT FGUKIP KPXQNXGU ſPFKPI C UGV QH RCTCOGVGTU YJKEJ OKPKOK\G VJG GORKTKECN RCVVGTP TGEQIPKVKQP GTTQT TCVG 6JG WUG QH FKUETKOKPCPV HWPEVKQP KP RCVVGTP TGEQIPKVKQP YCU UVCTVGF OCP[ [GCTU CIQ 1PG ENCUUKECN GZCORNG QH WUKPI FKUETKOKPCPV HWPEVKQP HQT ENCUUKſGT FGUKIP KP UVCVKUVKECN NKVGTCVWTG KU VJG VYQ ENCUU ENCUUKſECVKQP RTQDNGO WUKPI NKPGCT FKUETKOKPCPV HWPEVKQPU = ? +P RCTVKEWNCT C YKPFQY DCUGF OGVJQF YCU FGUETKDGF KP =? HQT VJG VYQ ENCUU ENCUUKſECVKQP RTQDNGO WUKPI NKPGCT FKUETKOKPCPV HWPEVKQPU VJCV OKPKOK\G VJG RTQDCDKNKV[ QH ENCUUKſECVKQP GTTQT TCVG 6JG HQEWU QH VJKU EJCRVGT KU QP VJG TGEGPV FGXGNQROGPV QH VJG IGPGTCN /%' DCUGF FKUETKOKPCVKXG OGVJQFU 6JG FKUETKOKPCPV HWPEVKQPU VJCV YG GPEQWPVGT CTG WUWCNN[ PQPNKPGCT CPF QHVGP TGNCVGF VQ VJG UVTWEVWTG QH VJG UVCVKUVKECN HTCOGYQTM WUGF KP URGGEJ CPF NCPIWCIG RTQEGUUKPI UWEJ CU JKFFGP /CTMQX OQFGNU 6JG TGCUQP QH VCMKPI C FKUETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ VQ ENCUUKſGT FGUKIP CU YKNN DG HWTVJGT GNCDQTCVGF KU FWG OCKPN[ VQ VJG HCEV VJCV YG NCEM EQORNGVG MPQYNGFIG QH VJG HQTO QH VJG FCVC FKUVTKDWVKQP CPF VTCKPKPI FCVC CTG KPCFGSWCVG RCTVKEWNCTN[ KP FGCNKPI YKVJ URGGEJ CPF NCPIWCIG RTQDNGOU 6JG RGTHQTOCPEG QH C TGEQIPK\GT KU PQTOCNN[ FGſ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ſGT ECP DG FKTGEVN[ GXCNWCVGF CPF QRVKOK\GF %NCUUKſGT FGUKIP YKVJQWV CUUWOKPI VJG MPQYNGFIG QH ENCUU RQUVGTKQT RTQDCDKNKVKGU YJKEJ CTG VJG DCUKU QH VJG FKUVTKDWVKQP GUVKOCVKQP DCUGF ENCUUKGT FGUKIP JCU DGGP UVWFKGF KP OCP[ CTGCU +P RCTVKEWNCT 6U[RMKP =? CPF #OCTK = ? RKQPGGTGF VJKU CRRTQCEJ HQT UGNHNGCTPKPI CPF UGNHQTICPK\KPI PGVU 6JG[ HQTOWNCVGF VJG RTQDNGO QH UGNHNGCTPKPI KPVQ C ENCUUKſECVKQP RTQDNGO YJKEJ EQPUKUVU QH QRVKOCN RCTVKVKQPKPI QH VJG QDUGTXCVKQP URCEG KPVQ TGIKQPU HQT YJKEJ VJG GZRGEVGF TKUM KU OKPKOK\GF +P CFFKVKQP C OCVJGOCVKECN OKPKOK\CVKQP RTQEGFWTG IGPGTCNK\GF RTQDCDKNKUVKE FG UEGPV )2& CNIQTKVJO QT UVQEJCUVKE CRRTQZKOCVKQP YCU RTQRQUGF CU C OGCPU HQT ENCUUKſGT FGUKIP WPFGT VJKU HTCOGYQTM 5KPEG VJGP XCTKQWU NQUU HWPEVKQPU JCXG DGGP WUGF KP FGUKIPKPI ENCUUKſGTU KPENWFKPI VJQUG RQRWNCT OGCPUSWCTG GTTQT DCUGF NQUU HWPEVKQPU *QYGXGT OCP[ VTCEVCDNG NQUU HWPEVKQPU FQ PQV JCXG C FKTGEV TGNCVKQP VQ VJG TGEQIPKVKQP GTTQT TCVG OKPKOK\CVKQP CPF VJGTGHQTG CNDGKV DCUGF QP FKUETKOKPCPV HWPEVKQPU VJG[ CTG PQV FKTGEVN[ TGNCVGF VQ TGEQIPKVKQP GTTQT TCVG YJKEJ UJQWNF DG VJG OQUV UGPUKDNG EJQKEG HQT ENCUUKſGT FGUKIP 1XGT VJG RCUV FGECFG VJG /%' DCUGF CRRTQCEJ JCU DGGP FGXGNQRGF VQ QXGTEQOG VJG HWPFCOGPVCN NKOKVCVKQPU QH VJG VTCFKVKQPCN CRRTQCEJ CPF VQ FKTGEVN[ NKPM VJG ENCUUKſGT FGUKIP RTQDNGO VQ ENCUUKſECVKQP GTTQT TCVG OKPKOK\CVKQP +P QTFGT VQ CNNGXKCVG VJG FGRGPFGPE[ QP VJG ENCUU RQUVGTKQT FKUVTKDWVKQPU C FKUETKOKPCPV HWPEVKQP DCUGF /%' CRRTQCEJ YCU RTQRQUGF D[ ,WCPI GV CN =? CU CP CNVGTPCVKXG VQ QRVKOCN ENCUUKſGT FG
E\&5&3UHVV//&
UKIP #NVJQWIJ VJKU CRRTQCEJ CRRNKGU VQ VJG RCVVGTP TGEQIPKVKQP RTQDNGO KP IGPGTCN KV ſPFU XCTKQWU CRRNKECVKQPU KP URGGEJ CPF NCPIWCIG RTQEGUUKPI +V YCU ſTUV CRRNKGF VQ F[PCOKE VKOG YCTRKPI DCUGF RCVVGTP TGEQIPKVKQP U[UVGOU = ? #RRNKECVKQP VQ JKFFGP /CTMQX OQFGN DCUGF EQPVKPWQWU URGGEJ TGEQIPKVKQP U[UVGOU YCU HQTOWNCVGF CU C UGIOGPVCN CPF UVTKPI OQFGN DCUGF /%' CRRTQCEJ = ? CPF UWEEGUUHWN CR RNKECVKQPU QH VJKU CRRTQCEJ YGTG TGRQTVGF KP = ? 6JKU CRRTQCEJ YCU HWTVJGT GZVGPFGF VQ HQTO C EQODKPGF UVTKPI OQFGN KP YJKEJ VTCKPKPI QH QVJGT OQFGN EQORQPGPVU KP URGGEJ CPF NCPIWCIG RTQEGUUKPI ECP DG CEJKGXGF WPFGT C WPK ſGF /%' HTCOGYQTM = ? +V YCU CRRNKGF VQ FKUETKOKPCVKXG OQFGN EQODKPCVKQP = ? CPF VQ CRRNKECVKQPU KP URGCMGT KFGPVKſECVKQP CPF XGTKſECVKQP = ? 6JG DCUKE KFGC QH VJG /%' CRRTQCEJ YCU HWTVJGT FGXGNQRGF HQT CRRNKECVKQPU KP WVVGT CPEG XGTKſECVKQP RTQDNGOU = ? # IGPGTCN HTCOGYQTM QH EQODKPKPI FG VGEVKQP CPF XGTKſECVKQP KP URGGEJ TGEQIPKVKQP CPF WPFGTUVCPFKPI YCU CNUQ RTQRQUGF KP YJKEJ VJG FKUETKOKPCPV HWPEVKQP DCUGF RCVVGTP TGEQIPKVKQP CRRTQCEJ YCU CRRNKGF KP DQVJ FGVGEVKQP CPF XGTKſECVKQP RTQEGUUGU = ? 9G DGIKP KP VJG PGZV UGEVKQP YKVJ C DTKGH TGXKGY QH VJG $C[GU FGEKUKQP VJGQT[ CPF KVU CRRNKECVKQP VQ VJG HQTOWNCVKQP QH UVCVKUVKECN RCVVGTP TGEQIPKVKQP RTQDNGO 9G KP VTQFWEG VJG FKUETKOKPCPV HWPEVKQP DCUGF UVCVKUVKECN RCVVGTP TGEQIPKVKQP CRRTQCEJ KP 5GEVKQP +P 5GEVKQP YG RTQXKFG C DTKGH KPVTQFWEVKQP VQ URGGEJ TGEQIPKVKQP CPF JKFFGP /CTMQX OQFGNKPI 6JG FKUETKOKPCPV HWPEVKQP DCUGF /%' RCVVGTP TGEQIPKVKQP CRRTQCEJ CPF KVU CRRNKECVKQP VQ *// DCUGF URGGEJ TGEQIPKVKQP U[UVGOU CTG KPVTQ FWEGF KP 5GEVKQP %QORCTKUQPU CTG OCFG VQ QVJGT ETKVGTKC KP URGGEJ TGEQIPKVKQP CPF KP RCTVKEWNCT YG UVWF[ VJG TGNCVKQP DGVYGGP /%' CPF //+ OCZKOWO OWVWCN KPHQTOCVKQP ETKVGTKC KP ENCUUKſGT FGUKIP KP VJG UGEQPF JCNH QH 5GEVKQP +P 5GEVKQP YG UVWF[ VJG GODGFFGF UVTKPI OQFGN DCUGF /%' CRRTQCEJ CPF KVU GZVGPUKQP VQ VJG JKIJGT NGXGN EQODKPGF UVTKPI OQFGN 9G FKUEWUU KUUWGU CPF CRRNKECVKQPU KP FKUETKO KPCVKXG OQFGN EQODKPCVKQP FKUETKOKPCVKXG NCPIWCIG OQFGN GUVKOCVKQP CPF FKUETKO KPCVKXG HGCVWTG GZVTCEVKQP WPFGT VJG IGPGTCN VJGQTGVKECN HTCOGYQTM QH VJG EQODKPGF UVTKPI OQFGN 5GEVKQP KU FGXQVGF VQ CRRNKECVKQPU QH FKUETKOKPCPV HWPEVKQP DCUGF RCVVGTP TGEQIPKVKQP CRRTQCEJ KP XGTKſECVKQP CPF KFGPVKſECVKQP 6JG FKUETKOKPCPV HWPEVKQP CRRTQCEJ KU UVWFKGF HQT XCTKQWU CRRNKECVKQPU KP URGGEJ CPF NCPIWCIG RTQ EGUUKPI UWEJ CU URGCMGT KFGPVKſECVKQP CPF XGTKſECVKQP WVVGTCPEG XGTKſECVKQP TGEQI PKVKQP DCUGF QP IGPGTCNK\GF EQPſFGPEG OGCUWTGU FGVGEVKQP CPF XGTKſECVKQP DCUGF CRRTQCEJ KP URGGEJ TGEQIPKVKQP CPF WPFGTUVCPFKPI 6JG EJCRVGT KU UWOOCTK\GF YKVJ FKUEWUUKQPU KP 5GEVKQP
1.2 Optimal Classifier from Bayes Decision Theory
(QT CP ENCUU ENCUUKſECVKQP RTQDNGO C ENCUUKſGT KU VQ ENCUUKH[ GCEJ TCPFQO UCORNG 6JG Z KPVQ QPG QH VJG / ENCUUGU 9G FGPQVG VJGUG ENCUUGU D[ ENCUUKſGT FGſPGU C OCRRKPI HTQO VJG UCORNG URCEG ¾ VQ VJG FKUETGVG
E\&5&3UHVV//&
ECVGIQTKECN UGV ¾ .GV DG VJG LQKPV RTQDCDKNKV[ FKUVTKDWVKQP QH CPF C SWCPVKV[ YJKEJ KU CUUWOGF VQ DG MPQYP VQ VJG FGUKIPGT QH VJG ENCUUKſGT +P QVJGT YQTFU VJG FGUKIPGT JCU HWNN MPQYNGFIG QH VJG TCPFQO PCVWTG QH VJG UQWTEG (TQO VJG UGV QH LQKPV RTQDCDKNKV[ FKUVTKDWVKQPU VJG OCTIKPCN CPF VJG EQPFKVKQPCN RTQDCDKNKV[ FKUVTKDWVKQPU ECP DG GCUKN[ ECNEWNCVGF +P QTFGT VQ EJCTCEVGTK\G VJG RGTHQTOCPEG QH VJG ENCUUKſGT GXGT[ ENCUU RCKT ECP DG CUUQEKCVGF YKVJ C EQUV QT NQUU HWPEVKQP YJKEJ UKIPKſGU VJG EQUV QH ENCUUKH[KPI QT TGEQIPK\KPI C ENCUU QDUGTXCVKQP KPVQ C ENCUU GXGPV 6JG NQUU HWPEVKQP KU IGPGTCNN[ PQPPGICVKXG YKVJ TGRTGUGPVKPI EQTTGEV ENCUUKſECVKQP 6JG NQUU HWPEVKQP KU C HWPEVKQP HTQO YJGTG KU VJG UGV QH TGCN PWODGTU +P ENCUUKſECVKQP YG OCMG C FGEKUKQP HQT QDUGTXKPI C TCPFQO UCORNG 5KPEG KU VJG ENCUU RQUVGTKQT RTQDCDKNKV[ VJCV VJG TCPFQO KPRWV KU HTQO VJG CXGTCIG NQUU CUUQEKCVGF YKVJ OCMKPI C FGEKUKQP ECP DG FGſPGF CU =?
6JKU NGCFU VQ C TGCUQPCDNG RGTHQTOCPEG OGCUWTG HQT VJG ENCUUKſGT KG VJG GZRGEVGF NQUU FGſPGF CU
YJGTG TGRTGUGPVU VJG ENCUUKſGTŏU FGEKUKQP CUUWOKPI QPG QH VJG őXCNWGUŒ DCUGF QP C TCPFQO QDUGTXCVKQP FTCYP HTQO C RTQDCDKNKV[ FKUVTKDWVKQP 6JG FGEKUKQP HWPEVKQP FGRGPFU QP VJG ENCUUKſGT FGUKIP 1DXKQWUN[ KH VJG ENCUUKſGT KU UQ FGUKIPGF VJCV HQT GXGT[
VJG GZRGEVGF NQUU KP GSWCVKQP YKNN DG OKPKOK\GF (QT OCP[ CRRNKECVKQPU KPENWFKPI URGGEJ TGEQIPKVKQP VJG NQUU HWPEVKQP KU WUWCNN[ EJQUGP VQ DG VJG \GTQQPG NQUU HWPEVKQP FGſPGF D[
YJKEJ CUUKIPU PQ NQUU VQ EQTTGEV ENCUUKſECVKQP CPF C WPKV NQUU VQ CP[ GTTQT TGICTFNGUU QH VJG ENCUU 9KVJ VJKU V[RG QH NQUU HWPEVKQP VJG GZRGEVGF NQUU KU VJWU VJG GTTQT RTQDCDKNKV[ QH ENCUUKſECVKQP QT TGEQIPKVKQP 6JG EQPFKVKQPCN NQUU DGEQOGU
6JG QRVKOCN ENCUUKſGT VJCV CEJKGXGU OKPKOWO KU VJWU VJG QPG VJCV KORNGOGPVU VJG HQNNQYKPI KH
E\&5&3UHVV//&
(QT OKPKOWO GTTQT TCVG ENCUUKſECVKQP VJG ENCUUKſGT GORNQ[U VJG FGEKUKQP TWNG QH YJKEJ KU ECNNGF VJG őOCZKOWO C RQUVGTKQTŒ /#2 FGEKUKQP 6JG OKPKOWO GTTQT TCVG CEJKGXGF D[ /#2 FGEKUKQP KU ECNNGF ő$C[GU TKUMŒ 9JGP CNN RQUVGTKQT RTQDCDKNKVKGU CTG MPQYP VJG ENCUUKſGT DCUGF QP /#2 TWNG KU CP QRVKOCN ENCUUKſGT DCUGF QP VJG $C[GU FGEKUKQP VJGQT[ *QYGXGT KH VJGUG RTQDCDKNKVKGU CTG PQV MPQYP QT VJG FGEKUKQP TWNG KU PQV DCUGF QP VJG ENCUU RQUVGTKQT RTQDCDKNKV[ VJGP YG ECPPQV WUG VJKU TGUWNV FKTGEVN[ +P RTCEVKEG VJGUG RTQDCDKNKVKGU JCXG VQ DG GUVKOCVGF HTQO C VTCKPKPI FCVC UGV YKVJ MPQYP ENCUU NCDGNU 6JG ENCUUKECN $C[GU FGEKUKQP VJGQT[ VJWU GHHGEVKXGN[ VTCPUHQTOU VJG ENCUUKſGT FGUKIP RTQDNGO KPVQ C FKUVTKDWVKQP GUVKOCVKQP RTQDNGO 6JKU KU VJG DCUKU QH VJG $C[GUKCP UVCVKUVKECN CRRTQCEJ VQ RCVVGTP TGEQIPKVKQP YJKEJ ECP DG UVCVGF CU IKXGP QT EQNNGEV C UGV QH VTCKPKPI FCVC QDUGTXCVKQPU ½ ¾ YKVJ MPQYP ENCUU NCDGNU GUVKOCVG VJG C RQUVGTKQT RTQDCDKNKVKGU HQT CP[ VQ KORNGOGPV VJG OCZKOWO C RQUVGTKQT FGEKUKQP HQT OKPKOWO $C[GU TKUM 6JG C RQUVGTKQT RTQDCDKNKV[ ECP DG TGYTKVVGP CU
5KPEG KU PQV C HWPEVKQP QH VJG ENCUU KPFGZ CPF VJWU JCU PQ GHHGEV KP VJG /#2 FGEKUKQP VJG PGGFGF RTQDCDKNKUVKE MPQYNGFIG ECP DG TGRTGUGPVGF D[ VJG ENCUU RTKQT CPF VJG EQPFKVKQPCN RTQDCDKNKV[ 6JGTG CTG UGXGTCN KUUWGU CUUQEKCVGF YKVJ VJKU ENCUUKECN CRRTQCEJ (KTUV VJG FKUVTKDW VKQPU WUWCNN[ JCXG VQ DG RCTCOGVGTK\GF KP QTFGT HQT VJGO VQ DG RTCEVKECNN[ WUGHWN HQT VJG KORNGOGPVCVKQP QH VJG /#2 TWNG QH 6JG ENCUUKſGT FGUKIPGT VJGTGHQTG JCU VQ FGVGTOKPG VJG TKIJV RCTCOGVTKE HQTO QH VJG FKUVTKDWVKQPU (QT OQUV QH VJG TGCN YQTNF RTQDNGOU VJKU KU C FKHſEWNV VCUM 1WT EJQKEG QH VJG FKUVTKDWVKQP HQTO KU QHVGP NKO KVGF D[ VJG OCVJGOCVKECN VTCEVCDKNKV[ QH VJG RCTVKEWNCT FKUVTKDWVKQP HWPEVKQPU CPF KU XGT[ NKMGN[ VQ DG KPEQPUKUVGPV YKVJ VJG CEVWCN FKUVTKDWVKQP 6JKU OGCPU VJCV VJG VTWG /#2 FGEKUKQP ECP TCTGN[ DG KORNGOGPVGF CPF VJG OKPKOWO $C[GU TKUM IGPGTCNN[ TGOCKPU CP WPCEJKGXCDNG NQYGT DQWPF 5GEQPF IKXGP C RCTCOGVGTK\GF FKUVTKDWVKQP HQTO VJG WPMPQYP RCTCOGVGTU FGſPKPI VJG FKUVTKDWVKQP JCXG VQ DG GUVKOCVGF HTQO C ſPKVG COQWPV QH NCDGNGF VTCKPKPI FCVC TGSWKTKPI VJCV VJG GUVKOCVKQP OGVJQF JCU VQ DG CDNG VQ RTQFWEG EQPUKUVGPV RCTCOGVGT XCNWGU YJGP VJG UK\G QH VJG VTCKPKPI UCORNGU XCTKGU 6JKTF KV TGSWKTGU C VTCKPKPI FCVC UGV QH UWHſEKGPV UK\G KP QTFGT VQ JCXG TGNK CDNG RCTCOGVGT GUVKOCVGU $WV KP RTCEVKEG CPF HQT URGGEJ CPF NCPIWCIG RTQEGUUKPI KP RCTVKEWNCT VTCKPKPI FCVC CTG CNYC[U URCTUG EQORCTGF VQ CNN RQUUKDNG TGCNK\CVKQPU CPF XCTKCVKQPU KP JWOCP URGGEJ CPF NCPIWCIG 6JGUG VJTGG DCUKE KUUWGU RQKPV QWV C HWPFCOGPVCN HCEV VJCV KU FGURKVG VJG EQPEGRVWCN QRVKOCNKV[ QH VJG $C[GU FGEKUKQP VJGQT[ CPF KVU CRRNKECVKQPU VQ RCVVGTP TGEQIPKVKQP KV ECPPQV CNYC[U DG CEEQORNKUJGF KP RTCEVKEG DGECWUG OQUV RTCEVKECN ő/#2Œ FGEKUKQPU KP URGGEJ CPF NCPIWCIG RTQ EGUUKPI CTG PQV VTWG /#2 FGEKUKQPU 6JKU WPFGTUVCPFKPI KU ETKVKECN HQT VJG FKUEWUUKQP VJCV HQNNQYU
E\&5&3UHVV//&
1.3 Discriminant Function Approach to Classifier Design &KUETKOKPCPV HWPEVKQPU QP VJG QVJGT JCPF CTG VJQUG HWPEVKQPU YJKEJ EJCTCEVGTK\G VJG FGEKUKQP TWNG QH VJG ENCUUKſ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½ ¾ ENCUUKſECVKQP RTQDNGO 6JG ENCUUKſGT WUGU C FKUETKOKPCPV HWPEVKQP UWEJ VJCV
KH KH
VJGP Z KU ENCUUKſGF VQ ½ VJGP Z KU ENCUUKſGF VQ ¾
.KPGCT FKUETKOKPCPV HWPEVKQPU CTG VJQUG HWPEVKQPU QH VJG HQTO
¼ YJGTG
¾ CPF ¼ C TGCN PWODGT 1T OQTG IGPGTCNN[
½
¼ ½ ½
YJGTG
½ ¼ ½ ¼
YJGTG KU VJG VTCPURQUKVKQP PQVCVKQP CPF VJG CTG MPQYP NKPGCTN[ KPFGRGPFGPV HWPEVKQPU QH (QT / ENCUU ENCUUKſECVKQP RTQDNGO WUKPI FKUETKOKPCPV HWPEVKQPU C UGV QH FKUETKOKPCPV HWPEVKQPU CTG WUGF CPF VJG ENCUUKſGT KU FGſPGF UWEJ VJCV KHH
9JGP VJG NQUU HWPEVKQP KU URGEKſGF VJG RTQDNGO QH QRVKOCN ENCUUKſGT FGUKIP WUKPI FKUETKOKPCPV HWPEVKQPU DGEQOGU C OKPKOK\CVKQP RTQDNGO QH ſPFKPI C DGUV UGV QH FKUETKOKPCPV HWPEVKQPU HTQO C ENCUU QH FKUETKOK PCPV HWPEVKQPU YJKEJ OKPKOK\GU VJG GZRGEVGF NQUU CU FGſPGF KP 'S +P QVJGT YQTFU VJG ENCUUKſGT FGUKIP RTQDNGO KU VQ ſPF
E\&5&3UHVV//&
´µ ´ µ
YJGTG KU VJG IKXGP HCOKN[ QH FKUETKOKPCPV HWPEVKQPU +H VJG NQUU HWPEVKQP KU IKXGP CU KP 'S CPF VJG RCTVKEWNCT UGV QH FKUETKOKPCPV HWPEVKQPU WUGF KP VJG ENCUUKſGT CTG VJG őVTWGŒ ENCUU RQUVGTKQT RTQDCDKNKV[ VJGP 'S KORNG OGPVU VJG UCOG /#2 FGEKUKQP TWNG CU FGſPGF KP 'S *QYGXGT KV KU KORQTVCPV VQ RQKPV QWV VJCV VJG FKUETKOKPCPV HWPEVKQP CRRTQCEJ VQ VJG QRVKOCN ENCUUKſGT FGUKIP CU URGEKſGF KP 'S QHVGP JCU CP KPſPKVG PWODGT QH UQNWVKQPU GXGP HQT VJG UCOG KU CP QRVKOCN UQNWVKQP VQ 'S ENCUUKſGT +V KU GCU[ VQ UGG VJCV KH
VJGP HQT CP[ KU CPQVJGT QRVKOCN UQNWVKQP CPF FGſPGU VJG UCOG ENCUUKſGT #ICKP VJKU KU SWKVG FKHHGTGPV HTQO VJG FKUVTKDWVKQP GUVKOCVKQP DCUGF CRRTQCEJ KP RCVVGTP ENCUUKſECVKQP +H VJG FKUETKOK PCPV HWPEVKQPU CTG NKOKVGF VQ VJG ENCUU RQUVGTKQT RTQDCDKNKVKGU CP[ FGXKCVKQP HTQO VJG őVTWGŒ ENCUU RQUVGTKQT RTQDCDKNKV[ YKNN TGUWNV KP C FKHHGTGPV ENCUUKſGT CPF YKNN DG KPHGTKQT VQ VJG QRVKOCN /#2 ENCUUKſGT VJCV CEJKGXGU VJG OKPKOWO $C[GU TKUM
6JG WUG QH FKUETKOKPCPV HWPEVKQPU KP UVCVKUVKECN RCVVGTP TGEQIPKVKQP KU VQ UQNXG VJG ENCUUKſGT FGUKIP RTQDNGO YJGP VJG GZCEV HQTO CPF XCNWG QH VJG ENCUU RQUVGTKQT RTQDC DKNKVKGU CTG PQV MPQYP GXGP YKVJ VJG JGNR QH VTCKPKPI FCVC QT VJG ENCUUKſGT JCU VQ DG DCUGF QP C RCTVKEWNCT ENCUU QH FKUETKOKPCPV HWPEVKQPU 6JGUG FKUETKOKPCPV HWPEVKQPU KP VJG ENCUUKſ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
1.4 Speech Recognition and Hidden Markov Modeling 5RGGEJ TGEQIPKVKQP KU C RTQDNGO QH TGEQIPK\KPI C YQTF UGSWGPEG HTQO JWOCP URGGEJ +V ECP DG XKGYGF CU C EQOOWPKECVKQP RTQDNGO 6JG JWOCP DTCKP UGTXGU CU VJG VGZV IGPGTCVQT YJKEJ IGPGTCVGU VJG YQTF UVTKPI 6JG YQTF UVTKPI IQGU VQ VJG CEQWUVKE EJCPPGN YJKEJ EQPUKUVU QH C URGCMGTŏU CTVKEWNCVQT[ CRRCTCVWU CPF QVJGT CEQWUVKE RTQ EGUUGU VJCV EQPXGTV VJG VGZV UVTKPI KPVQ CP CWFKDNG CEQWUVKE YCXGHQTO 6JG CEQWUVKE EJCPPGN KP XGTDCN EQOOWPKECVKQP CEVU CU C FCVC VTCPUFWEGT CPF EQORQUGT 6JG URGGEJ TGEQIPK\GT KU C FGEQFGT YJKEJ RGTHQTOU CP KPXGTUG QRGTCVKQP VQ FGEQFG VJG OGUUCIG HTQO VJG URGGEJ YCXGHQTO 6JGTGHQTG C FGEQFGT RGTHQTOU C OCZKOWO C RQUVGTKQT UWEJ VJCV FGEKUKQP VJCV FGVGTOKPGU VJG YQTF UGSWGPEG
Ï Ï E\&5&3UHVV//&
YJGTG KU VJG UEQTG HTQO CEQWUVKE OQFGNKPI CPF KU VJG UEQTG HTQO VJG NCPIWCIG OQFGN # V[RKECN URGGEJ TGEQIPKVKQP U[UVGO EQPUKUVU QH VJG HQNNQYKPI DCUKE EQORQPGPVU
#EQWUVKE HGCVWTG GZVTCEVKQP #EQWUVKE HGCVWTG GZVTCEVKQP KU VQ GZVTCEV VJG HGC VWTGU HQT URGGEJ TGEQIPKVKQP HTQO VJG URGGEJ YCXGHQTO +V V[RKECNN[ KPENWFGU C UJQTVVKOG EGRUVTCN CPCN[UKU YJKEJ IGPGTCVGU C HGCVWTG XGEVQT QH NQY HTG SWGPE[ EGRUVTCN EQGHſEKGPVU HQT GXGT[ 8CTKQWU UKIPCN RTQEGUU KPI RTQEGFWTGU CTG RGTHQTOGF VQ UGRCTCVG VJG UCNKGPV CEQWUVKE KPHQTOCVKQP HQT URGGEJ TGEQIPKVKQP RWTRQUGU (TQO PQY QP YG YKNN WUG VJG PQVCVKQP ½ Ì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Ò ½ Ò ½ YJKEJ KU VJG RTQDCDKNKV[ QH QDUGTXKPI YQTF Ò IKXGP VJG YQTF JKUVQT[ ½ Ò ½ &GEQFKPI GPIKPG 6JG FGEQFKPI GPIKPG UGCTEJGU HQT VJG DGUV YQTF UGSWGPEG IKXGP VJG HGCVWTG CPF VJG OQFGN (QT URGGEJ TGEQIPKVKQP DCUGF QP *// OQF GNKPI VJKU KU CEJKGXGF VJTQWIJ 8KVGTDK FGEQFKPI (QT C FKUETGVG QDUGTXCVKQP RTQDCDKNKV[ DCUGF U[UVGO YQTF UVTKPI KU IKXGP D[
Ï
Ï
YJGTG Å KU VJG DGUV UVCVG UGSWGPEG IKXGP EQPVKPWQWU FGPUKV[ *//U
CPF VJG OQFGN
Ï
Ï
(QT
YJKEJ KU DCUGF QP VJG NQINKMGNKJQQF UEQTG CNQPI VJG DGUV UVCVG UGSWGPEG É
1.4.1 Hidden Markov Modeling of Speech 5RGGEJ KU IGPGTCVGF HTQO JWOCP CTVKEWNCVQT CPF KV KU WPKSWG KP OCP[ YC[U 9JGP YG URGCM QWT CTVKEWNCVQT[ CRRCTCVWU VJG NKRU LCY VQPIWG CPF XGNWO OQFWNCVGU VJG CKT RTGUUWTG CPF ƀQY VQ RTQFWEG CP CWFKDNG UGSWGPEG QH UQWPFU &WG VQ VJG RJ[UKECN EQPUVTCKPVU VJG CTVKEWNCVQT EQPſIWTCVKQP ECPPQV WPFGTIQ XGT[ FTCUVKE EJCPIGU CPF FWTKPI VJG UJQTV KPVGTXCN YJGTG VJG CTVKEWNCVQT[ EQPſIWTCVKQP UVC[U TGNCVKXGN[ EQP UVCPV C TGIKQP QH őSWCUKUVCVKQPCTKV[Œ KP VJG RTQFWEGF URGGEJ ECP QHVGP DG QDUGTXGF
E\&5&3UHVV//&
*KFFGP /CTMQX OQFGNKPI KU C RQYGTHWN UVCVKUVKECN HTCOGYQTM HQT VKOG XCT[KPI SWCUK UVCVKQPCT[ RTQEGUU CPF C RQRWNCT EJQKEG HQT UVCVKUVKECN OQFGNKPI QH URGGEJ UKIPCN )KXGP C URGGEJ WVVGTCPEG NGV Ü ½ Ü Ü DG C HGCVWTG XGEVQT UGSWGPEG GZVTCEVGF HTQO VJG URGGEJ YCXGHQTO YJGTG Ü FGPQVGU C UJQTVVKOG XGEVQT OGCUWTG OGPV CPF KV KU EQPXGPVKQPCNN[ C EGRUVTCN XGEVQT (WTVJGT EQPUKFGT C ſTUVQTFGT UVCVG /CTMQX EJCKP IQXGTPGF D[ C UVCVG VTCPUKVKQP RTQDCDKNKV[ OCVTKZ YJGTG KU VJG RTQDCDKNKV[ QH OCMKPI C VTCPUKVKQP HTQO UVCVG VQ UVCVG #UUWOG VJCV CV VJG UVCVG QH VJG U[UVGO KU URGEKſGF D[ CP KPKVKCN UVCVG RTQDCDKNKV[ 6JGP HQT CP[ UVCVG UGSWGPEG Õ VJG RTQDCDKNKV[ QH Õ DGKPI IGPGTCVGF D[ VJG /CTMQX EJCKP KU
Õ ¼ ¼ ½ ½¾ Ì ½Ì
5WRRQUG VJG U[UVGO YJGP CV UVCVG RWVU QWV CP QDUGTXCVKQP Ü CEEQTFKPI VQ C FKUVTKDWVKQP Ø Ü Ü 6JG JKFFGP /CTMQX OQFGN WUGF CU C FKUVTKDWVKQP HQT VJG URGGEJ WVVGTCPEG KU VJGP FGſPGF CU Õ Õ Õ ¼ Ø ½ Ø Ø Ü YJGTG KU VJG RCTCOGVGT UGV HQT VJG OQFGN #U ECP DG UGGP KP Ø FGſPGU VJG FKUVTKDWVKQP HQT UJQTVVKOG QDUGTXCVKQPU CPF EJCTCEVGTK\GU VJG DGJCXKQT CPF KPVGTTGNCVKQPUJKR DGVYGGP FKHHGTGPV UVCVGU QH VJG
URGGEJ IGPGTCVKQP RTQEGUU +P QVJGT YQTFU VJG UVTWEVWTG QH C JKFFGP /CTMQX OQFGN RTQXKFGU C TGCUQPCDNG OGCPU HQT EJCTCEVGTK\KPI VJG FKUVTKDWVKQP QH C URGGEJ UKIPCN 0QTOCNN[ VJG VQVCN PWODGT QH UVCVGU KU OWEJ UOCNNGT VJCP VJG VKOG FWTCVKQP QH VJG URGGEJ WVVGTCPEG 6JG UVCVG UGSWGPEG Õ FKURNC[U C EGTVCKP FGITGG QH UVCDKNKV[ COQPI CFLCEGPV ¼ U FWG VQ VJG CDQXG OGPVKQPGF őSWCUKUVCVKQPCTKV[Œ 6JG WUG QH *//U CU URGGEJ FKUVTKDWVKQPU KU UJQYP VQ DG RTCEVKECNN[ GHHGEVKXG +V UJQWNF DG PQVGF VJCV VJG EJQKEG QH UVCVG QDUGTXCVKQP FKUVTKDWVKQPU Ø Ü KU PQV URGEKſ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ſPF C DGUV UVCVG UGSWGPEG Õ YJKEJ KU QRVKOCN KP C EGTVCKP UGPUG IKXGP VJG URGGEJ HGCVWTG UGSWGPEG 5KPEG UVCVGU KP *// CTG TGNCVGF VQ YQTFU CPF YQTF ENCUUGU VJG YQTF UGSWGPEG KP URGGEJ WVVGTCPEG ECP DG KFGPVKſGF D[ VTCEKPI VJTQWIJ VJG YQTF NCDGNU KP UVCVG UGSWGPEG Õ 6JG GUVKOCVKQP RTQDNGO KU VQ GUVKOCVG
E\&5&3UHVV//&
*// RCTCOGVGTU HTQO C IKXGP UGV QH VTCKPKPI UCORNGU CEEQTFKPI VQ UQOG OGCP KPIHWN ETKVGTKQP 6JG EQPXGPVKQPCN CRRTQCEJ KU DCUGF QP VJG OCZKOWO NKMGNKJQQF
/. RTKPEKRNG CPF VJG OQFGN RCTCOGVGT UGV KU GUVKOCVGF UQ VJCV VJG NKMGNKJQQF QP VJG VTCKPKPI FCVC KU OCZKOK\GF 8CTKQWU JKIJN[ GHſEKGPV /. DCUGF CNIQTKVJOU CTG FGXGNQRGF KP URGGEJ TGEQIPKVKQP HQT *//U UWEJ CU $CWO9GNEJ CNIQTKVJO =? CPF UGIOGPVCN MOGCPU CNIQTKVJO =? /QTG FKUEWUUKQPU QH RCTCOGVGT GUVKOCVKQP RTQD NGO HQT *//U ECP DG HQWPF KP =? +V UJQWNF DG PQVGF VJCV VJG EQPXGPVKQPCN /. OGVJQF KP URGGEJ TGEQIPKVKQP FQGU PQV PGEGUUCTKN[ NGCF VQ C OKPKOWO GTTQT TCVG RGT HQTOCPEG HQT VJG TGEQIPK\GT 6JKU KU FWG OCKPN[ VQ VJG NKMGN[ OKUOCVEJ DGVYGGP VJG EJQUGP FKUVTKDWVKQP HQTO CPF VJG CEVWCN URGGEJ FCVC FKUVTKDWVKQP CPF VJG ſPKVG VTCKPKPI MPQYP FCVC UGV YJKEJ KU QHVGP KPCFGSWCVG
1.5 MCE Classifier Design Using Discriminant Functions #U KV KU PQVGF YKVJQWV VJG MPQYNGFIG QH VJG HQTO QH VJG ENCUU RQUVGTKQT RTQDCDKNK VKGU TGSWKTGF KP VJG ENCUUKECN $C[GU FGEKUKQP VJGQT[ ENCUUKſGT FGUKIP D[ FKUVTKDWVKQP GUVKOCVKQP QHVGP FQGU PQV NGCF VQ CP QRVKOCN RGTHQTOCPEG 6JKU OQVKXCVGU VJG GHHQTV QH UGCTEJKPI HQT QVJGT CNVGTPCVKXG ETKVGTKC KP ENCUUKſGT FGUKIP +P RCTVKEWNCT ETKVGTKC QH //+ OCZKOWO OWVWCN KPHQTOCVKQP CPF /&+ OKPKOWO FKUETKOKPCVKXG KPHQT OCVKQP CTG WUGF KP OCP[ CRRNKECVKQPU = ? #NVJQWIJ VJGUG OGVJQFU FGOQPUVTCVG UKIPKſECPV RGTHQTOCPEG CFXCPVCIGU QXGT VJG VTCFKVKQPCN /. CRRTQCEJ VJG[ CTG PQV DCUGF QP C FKTGEV OKPKOK\CVKQP QH C NQUU HWPEVKQP YJKEJ NKPMU VQ VJG ENCUUKſECVKQP GTTQT TCVG &Q6W GV CN =? UVWFKGF /%' UQNWVKQP HQT VJG VYQ ENCUU PQPRCTCOGVTKE ENCUUK ſECVKQP RTQDNGO WUKPI NKPGCT FKUETKOKPCPV HWPEVKQPU 6JG[ GORNQ[GF C YKPFQYGF UEJGOG VQ QXGTEQOG VJG RTQDNGO QH UKPIWNCT ITCFKGPV HWPEVKQPU CUUQEKCVGF YKVJ VJG GTTQT EQWPV KPFKECVQT HWPEVKQP QH VJG ENCUUKſGT # IGPGTCN CRRTQCEJ HQT OWNVKENCUU CPF PQPNKPGCT FKUETKOKPCPV HWPEVKQPU CTG RTQRQUGF D[ ,WCPI GV CN =? 6JKU IGP GTCN CRRTQCEJ KU ECNNGF őOKPKOWO ENCUUKſECVKQP GTTQT /%' OGVJQFŒ KP YJKEJ VJG ENCUUKſGT FGUKIP CPF RCTCOGVGT GUVKOCVKQP CTG VQ EQTTGEVN[ FKUETKOKPCVG VJG QDUGTXC VKQPU HQT DGUV TGEQIPKVKQPENCUUKſECVKQP TGUWNVU TCVJGT VJCP VQ ſV VJG FKUVTKDWVKQPU VQ VJG FCVC
1.5.1 MCE Classifier Design Strategy .GV WU EQPUKFGT C UGV QH ENCUU FKUETKOKPCPV HWPEVKQPU FG ſPGF D[ VJG RCTCOGVGT UGV 6JG ENCUUKſGT KU VJG QPG VJCV HQT CP QDLGEV
KHH
6JG IGPGTCN /%' ENCUUKſGT FGUKIP UVTCVGI[ KU DCUGF QP C URGEKCN V[RG QH NQUU HWPEVKQP 2CTCOGVGTU QH VJG ENCUUKſGT CTG GUVKOCVGF KP UWEJ C YC[ VJCV OKPKOK\KPI VJG GZRGEVGF
E\&5&3UHVV//&
NQUU TGNCVGU VQ C OKPKOK\CVKQP QH VJG TGEQIPKVKQP GTTQT TCVG QH VJG ENCUUKſGT 6JKU KU CEJKGXGF VJTQWIJ C VJTGG UVGR RTQEGUU 6JG OKUENCUUKſECVKQP OGCUWTG KP VJG /%' DCUGF CRRTQCEJ KU FGſPGF CU
YJGTG KU C RQUKVKXG PWODGT =? 6JKU OKUENCUUKſECVKQP OGCUWTG KU C EQPVKPWQWU HWPEVKQP QH VJG ENCUUKſGT RCTCOGVGTU CPF CVVGORVU VQ GOWNCVG VJG FGEKUKQP TWNG IKXGP VJG FKUETKOKPCPV HWPEVKQP (QT CP ENCUU WVVGTCPEG KORNKGU OKUENCUUKſECVKQP CPF OGCPU C EQTTGEV FGEKUKQP 9JGP CR RTQCEJGU VJG VGTO KP VJG DTCEMGV KU VJG PQTO QP VJG FKUETGVG KPVGIGT UGV YJKEJ EQPXGTIGU VQ VJG PQTO CPF DGEQOGU $[ XCT[KPI VJG XCNWG QH CPF QPG ECP VCMG CNN VJG EQORGVKPI ENCUUGU KPVQ EQPUKFGTCVKQP CEEQTFKPI VQ VJG KPFKXKFWCN UKIPKſECPEG YJGP UGCTEJKPI HQT VJG ENCUUKſGT RCTCOGVGT 6JG NQUU HWPEVKQP KU WUGF HQT TGEQIPKVKQP GTTQT TCVG OKPKOK\CVKQP 6JG OKUENCUUKſ ECVKQP OGCUWTG QH KU GODGFFGF KP C UOQQVJ \GTQQPG HWPEVKQP HQT YJKEJ CP[ OGODGT QH VJG UKIOQKF HWPEVKQP HCOKN[ KU CP QDXKQWU ECPFKFCVG # IGPGTCN HQTO QH VJG loss function ECP VJGP DG FGſPGF CU
YJGTG KU C UKIOQKF HWPEVKQP QPG GZCORNG QH YJKEJ KU
YKVJ PQTOCNN[ UGV VQ CPF UGV VQ ITGCVGT QT GSWCN VQ QPG %NGCTN[ YJGP KU OWEJ UOCNNGT VJCP \GTQ YJKEJ KORNKGU EQTTGEV ENCUUKſECVKQP XKTVWCNN[ PQ NQUU KU KPEWTTGF 9JGP KU RQUKVKXG KV NGCFU VQ C RGPCNV[ YJKEJ DGEQOGU GUUGPVKCNN[ C ENCUUKſECVKQPTGEQIPKVKQP GTTQT EQWPV 6JG ENCUUKſGT RCTCOGVGT GUVKOCVKQP KU DCUGF QP VJG OKPKOK\CVKQP QH VJG GZRGEVGF NQUU (QT CP[ WPMPQYP QDLGEV VJG ENCUUKſGT RGTHQTOCPEG KU OGCUWTGF D[
YJGTG KU VJG KPFKECVQT HWPEVKQP 6JG GZRGEVGF NQUU YJKEJ KU TGNCVGF VQ TGEQIPK VKQP GTTQT TCVG KU IKXGP D[
6JKU VJTGGUVGR FGſPKVKQP GOWNCVGU VJG ENCUUKſECVKQP QRGTCVKQP CU YGNN CU VJG TGEQI PKVKQP GTTQT TCVG DCUGF RGTHQTOCPEG GXCNWCVKQP KP C UOQQVJ HWPEVKQPCN HQTO UWKVCDNG
E\&5&3UHVV//&
HQT ENCUUKſGT RCTCOGVGT QRVKOK\CVKQP +V UJQWNF DG RQKPVGF QWV VJCV KH VJG EQTTGEV HQTO QH VJG RQUVGTKQT RTQDCDKNKV[ KU WUGF VJG $C[GU OKPKOWO TKUM KU VJGP GZRTGUUGF CU
CPF TGRTGUGPVU VJG GPVKTG
YJGTG UKIPCN URCEG 6JKU ECP DG CRRTQZKOCVGF D[ VJG NQUU HWPEVKQP KP /%' CRRTQCEJ CU HQNNQYU
#P KORQTVCPV RQKPV JGTG KU VJCV CRRTQZKOCVKQP CEEWTCE[ QH 'S ECP DG EQP VTQNNGF D[ XCT[KPI VJG EQPUVCPVU KP VJG UOQQVJ /%' NQUU HWPEVKQP $CUGF QP VJG ETKVGTKQP QH YG ECP EJQQUG VQ OKPKOK\G QPG QH VYQ SWCPVKVKGU HQT VJG ENCUUKſGT RCTCOGVGT UGCTEJ QPG KU VJG GZRGEVGF NQUU CPF VJG QVJGT VJG GORKTKECN NQUU
1.5.2 Optimization Methods 6JG RWTRQUG QH VJG VTCKPKPI RTQEGUU KP VJG /%' CRRTQCEJ KU VQ ſPF C UGV QH RCTCOGVGTU UQ VJCV C RTGUETKDGF NQUU KU OKPKOK\GF #U OGPVKQPGF RTGXKQWUN[ VJG VYQ MKPFU QH NQUU YG HQEWU QP CTG VJG GZRGEVGF NQUU CPF VJG GORKTKECN NQUU 1.5.2.1 Expected Loss (QT C ENCUUKſECVKQP RTQDNGO KPXQNXKPI FKHHGTGPV ENCUUGU VJG GZRGEVGF NQUU KU FG ſPGF CU
¾
8CTKQWU OKPKOK\CVKQP CNIQTKVJOU ECP DG WUGF VQ OKPKOK\G VJG GZRGEVGF NQUU 6JG IGPGTCNK\GF RTQDCDKNKUVKE FGUEGPV )2& CNIQTKVJO KU C RQYGTHWN CNIQTKVJO VJCV ECP DG WUGF VQ CEEQORNKUJ VJKU VCUM =? +P )2& DCUGF OKPKOK\CVKQP CNIQTKVJO VJG VCTIGV HWPEVKQP KU OKPKOK\GF CEEQTFKPI VQ CP KVGTCVKXG RTQEGFWTG
YJGTG KU C RQUKVKXG FGſPKVG OCVTKZ =? KU C UGSWGPEG QH RQUKVKXG PWODGTU CPF KU VJG ITCFKGPV HWPEVKQP QH VJG NQUU HWPEVKQP CV CPF KU
VJG VJ VTCKPKPI UCORNG WUGF KP VJG UGSWGPVKCN VTCKPKPI RTQEGUU 6JG EQPXGTIGPEG RTQRGTVKGU QH )2& CNIQTKVJO YCU UVWFKGF KP VJG NKVGTCVWTG GI = ? CPF UQOGVKOGU WPFGT VJG PCOG QH UVQEJCUVKE CRRTQZKOCVKQP 7PFGT XGT[ IGPGTCN EQPFKVKQPU VJG HQNNQYKPI EQPXGTIGPEG RTQRGTVKGU ECP DG GUVCDNKUJGF =?
E\&5&3UHVV//&
½
such that for all t, the inner product where is the Hessian matrix of second order partial derivatives; £ is the unique such that
Property 1 Suppose the following conditions are satisfied:
½
Then, given by
will converge to £ almost surely (i.e. with probability one). %QPFKVKQP ECP DG EQPUKFGTCDN[ YGCMGPGF 'XGP YKVJQWV EQPFKVKQP NQYKPI KU UVKNN VTWG
VJG HQN
YJGTG KU C UWDUGSWGPEG QH +P VJKU ECUG YKNN EQPXGTIG VQ C NQECN OKPKOWO RQKPV £ YJGTG #FCRVKPI VJG OQFGN RCTCOGVGTU WUKPI C UCORNG D[ UCORNG WRFCVKPI HQTOWNC CU KP 'S KU OQUV GHſEKGPV KP VGTOU QH VJG WUG QH CXCKNCDNG VTCKPKPI UCORNGU $WV VJG UKPING UCORNG DCUGF ITCFKGPV GUVKOCVKQP ECP DG PQKU[ NGCFKPI VQ ƀWEVWCVKQPU FWT KPI VJG RCTCOGVGT GUVKOCVKQP RTQEGUU $CVEJ OQFG CFCRVCVKQP UEJGOGU DCUGF QP C ITCFKGPV GUVKOCVG YJKEJ KU CP CXGTCIG QH GXGT[ UCORNGU ECP CNUQ DG WUGF 1VJGT XCTKCVKQPU VQ VJG QTKIKPCN )2& CNIQTKVJOU CTG CNUQ RQUUKDNG EJQKEGU VQ URGGF WR VJG EQPXGTIGPEG CPF TGFWEG VJG ƀWEVWCVKQP FWTKPI VJG ENCUUKſ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ſECVKQP HQT UQNXKPI OKPKOK\C VKQP RTQDNGOU YKVJ EQPUVTCKPVU #U YKNN DG UJQYP UJQTVN[ QPG ECP WVKNK\G RCTCOGVGT URCEG VTCPUHQTOCVKQPU VQ TGUQNXG VJKU KUUWG +P VJKU OGVJQF VJG QTKIKPCN RCTCOGVGTU CTG WRFCVGF VJTQWIJ VJG KPXGTUG VTCPUHQTO HTQO VJG VTCPUHQTOGF RCTCOGVGT URCEG VQ VJG QTKIKPCN RCTCOGVGT URCEG 6JKU KU FQPG KP UWEJ C YC[ VJCV EQPUVTCKPVU QP VJG QTKI KPCN RCTCOGVGTU CTG CNYC[U OCKPVCKPGF /QTG FGVCKNGF KNNWUVTCVKQPU QH VJKU CRRTQCEJ CTG IKXGP KP NCVGT UGEVKQPU
1.5.2.2 Empirical Loss
VJG GORKTKECN (QT C IKXGP VTCKPKPI FCVC UGV EQPUKUVKPI QH UCORNGU RTQDCDKNKV[ OGCUWTG FGſPGF QP VJG VTCKPKPI FCVC UGV KU C FKUETGVG RTQDCDKNKV[ OGC
E\&5&3UHVV//&
UWTG YJKEJ CUUKIPU GSWCN OCUU CV GCEJ UCORNG 6JG GORKTKECN NQUU QP VJG QVJGT JCPF KU VJWU GZRTGUUGF CU
¾
YJGTG FGPQVGU VJG KPFGZ QH VJG VTCKPKPI WVVGTCPEG KP VJG VTCKPKPI UGV QH UK\G CPF KU VJG GORKTKECN OGCUWTG FGſPGF QP VJG VTCKPKPI UGV +H VJG VTCKPKPI UCORNGU CTG QDVCKPGF D[ CP KPFGRGPFGPV UCORNKPI HTQO C URCEG YKVJ C ſZGF RTQDCDKNKV[ FKUVTK DWVKQP VJG GORKTKECN RTQDCDKNKV[ FKUVTKDWVKQP YKNN EQPXGTIG VQ KP FKUVTKDWVKQP +P QVJGT YQTFU HQT CP[ OGCUWTCDNG HWPEVKQP CU
6JG GORKTKECN NQUU FGſPGF QP VJG KPFGRGPFGPV VTCKPKPI UCORNGU YKNN EQPXGTIG VQ VJG GZRGEVGF NQUU CU VJG UCORNG UK\G KPETGCUGU 9KVJ UWHſEKGPV VTCKPKPI UCORNGU VJG GORKTKECN NQUU KU CP GUVKOCVG QH VJG GZRGEVGF NQUU 6JG IQQFPGUU QH VJKU GUVKOCVG KU FGVGTOKPGF D[ VJG VTCKPKPI UCORNG UK\G CPF VJG EQPXGTIGPEG TCVG QH VJG GORKTKECN RTQDCDKNKV[ OGCUWTG VQ VJG NKOKV FKUVTKDWVKQP 8CTKQWU WRRGT DQWPFU QP VJG EQPXGTIGPEG TCVG QH VJG GORKTKECN RTQDCDKNKV[ OGCUWTG ECP DG HQWPF KP =?
1.5.3 Other Optimization Methods +V UJQWNF DG RQKPVGF QWV VJCV CNVJQWIJ VJG )2& V[RG QH CFCRVCVKQP CNIQTKVJO KU GH HGEVKXG CPF OQUV RQRWNCT QVJGT QRVKOK\CVKQP OGVJQFU ECP CNUQ DG WUGF HQT GTTQT TCVG OKPKOK\CVKQP /%' DCUGF ENCUUKſGT FGUKIP KU XGT[ URGEKſE QP VJG HQTO CPF UVTWEVWTG QH VJG FKUETKOKPCPV HWPEVKQP CPF NQUU HWPEVKQP TGICTFKPI VJG ENCUUKſGT CPF TGNCVKXGN[ WPTGUVTKEVGF VQ YJCV RCTVKEWNCT QRVKOK\CVKQP OGVJQFU YJKEJ CTG WUGF VQ OKPKOK\G VJG NQUU /CP[ KPPQXCVKQPU CTG RQUUKDNG HQT DGVVGT QRVKOK\CVKQP TGUWNVU +P RCTVKE WNCT OGVJQFU QH NKPGCT RTQITCOOKPI =? ITCFKGPV RTQLGEVKQP =? CPF ITQYVJ VTCPUHQTOCVKQP = ? CTG CNUQ WUGF HQT OKPKOK\CVKQP QH VJG GZRGEVGF NQUU KP /%' ENCUUKſGT FGUKIP +P VJG ITQYVJ VTCPUHQTOCVKQP DCUGF CRRTQCEJ VJG IQCN KU VQ UGGM CVTCPUHQTOCVKQP UWEJ VJCV YJGTG KU C RTQDCDKNKV[ XGEVQT
KG CPF 6JKU CRRTQCEJ UQOGVKOGU TGHGTTGF VQ CU GZVGPFGF $CWO9GNEJ $9 CNIQTKVJO KP UQOG NKVGTCVWTG YCU QTKIKPCVGF HTQO $CWO'CIQPŏU KPGSWCNKV[ HQT DGKPI C RQN[PQOKCN YKVJ PQPPGICVKXG EQGHſ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
E\&5&3UHVV//&
GUVKOCVKQP NKMG HQTOWNC
¼ ¼ ¼
YJGTG KU C EQPUVCPV VQ DG FGVGTOKPGF CPF VJG UWO KP VJG FGPQOKPCVQT KU VCMGP QXGT CNN RCTCOGVGTU DGNQPIKPI VQ VJG UCOG FKUVTKDWVKQP +V KU UJQYP KP =? VJCV VJGTG KU C XCNWG UWEJ VJCV KH 'S KU C ITQYVJVTCPUHQTOCVKQP *QYGXGT 'S ECPPQV DG FKTGEVN[ CRRNKGF VQ CPF RCTCOGVGT GUVKOCVKQP QH EQPVKPWQWU RTQDCDKNKV[ FGPUKVKGU CPF PGY HQTOWNCVKQPU HQT EQPVKPWQWU FGPUKV[ *//U CTG PGGFGF 7UKPI C FKUETGVG CRRTQZKOCVKQP CTIWOGPV VJG ITQYVJVTCPUHQTOCVKQP OGVJQF KU GZVGPFGF VQ //+DCUGF RCTCOGVGT GUVKOCVKQP YKVJ EQPVKPWQWU FGPUKV[ *//U CPF PGY HQTOWNCVKQPU QH VJG ITQYVJVTCPUHQTOCVKQP HQT EQPVKPWQWU FGPUKV[ *//U KP //+ VTCKPKPI CTG FGTKXGF =? /QTG TGEGPVN[ CP GNGICPV RTQQH QH VJG VJGQTGVKECN RTQRGTVKGU QH VJG ITQYVJVTCPUHQTOCVKQP OGVJQF HQT //+ VTCKPKPI KU IKXGP KP =? YJKEJ GUVCDNKUJGU VJG ITQYVJVTCPUHQTOCVKQP OGVJQF KP C OQTG IGPGTCN UGVVKPI $WV OQTG YQTM TGOCKPU VQ DG FQPG KP QTFGT VQ CRRN[ VJG UKOKNCT OGVJQF HQT IGPGTCN /%' DCUGF RCTCOGVGT GUVKOCVKQP YKVJ EQPVKPWQWU RTQDCDKNKV[ FGPUKVKGU +V KU KPVGTGUVKPI VQ PQVG VJCV KH KU NCTIG VJGP VJG EQPXGTIGPEG QH VJKU CNIQTKVJO KU UNQY CPF KH KU VQQ NCTIG VJG CNIQTKVJO KU RTCEVKECNN[ PQV WUGHWN +P QTFGT VQ IGV HCUV EQPXGTIGPEG PGGFU VQ DG CU UOCNN CU RQUUKDNG /QFKſECVKQPU QH VJG QTKIKPCN CNIQTKVJO CPF WUKPI UGCTEJ JGWTKUVKEU CTG CVVGORVGF VQ URGGF WR VJG EQPXGTIGPEG CPF RGTHQTOCPEG KORTQXGOGPVU QXGT VJG QTKIKPCN CRRTQCEJ CTG CNUQ TGRQTVGF KP = ?
1.5.4 HMM as a Discriminant Function (QNNQYKPI YG JCXG UGXGTCN YC[U QH WUKPI CP *// CU VJG FKUETKOKPCPV HWPE VKQP # DCUKE EQORQPGPV KP KU VJG LQKPV QDUGTXCVKQPUVCVG RTQDCDKNKV[ ¼
½
YJKEJ KU PQY FGſPGF CU C EQORQPGPV HWPEVKQP HQT ENCUU CU YGNN 6JG FKUETKOKPCPV HWPEVKQP HQT ENCUU ECP VCMG UGXGTCN RQUUKDNG HQTOU DCUGF QP
YJGTG KU VJG VQVCN PWODGT QH RQUUKDNG UVCVG UGSWGPEGU CPF KU C RQUKVKXG PWODGT CPF
(WPEVKQPU QH VJG CDQXG
E\&5&3UHVV//&
0QVG VJCV KU GSWKXCNGPV VQ VJG NKMGNKJQQF HWPEVKQP KU GSWKXCNGPV VQ VJG OCZKOWO LQKPV QDUGTXCVKQPUVCVG UGSWGPEG RTQDCDKNKV[ CPF KU C IGPGTCNK\GF OKZVWTG OQFGN YJKEJ CRRTQCEJGU YJGP 9G WUG VJG NQICTKVJO QH
CU CP GZCORNG KP QWT FGTKXCVKQP DGECWUG KV KU VJG OQUV RQRWNCT EJQKEG HQT *// DCUGF TGEQIPKVKQP U[UVGOU CUUQEKCVGF YKVJ 8KVGTDK FGEQFKPI 6JG CNIQTKVJO DCUGF QP KU QHVGP ECNNGF segmental GPD =? 9G FGſPG HQT ܽ Ü Ü CPF Ü ¼ YKVJ DGKPI VJG FKOGPUKQP QH Ü
Õ Õ Ü
YJGTG Õ KU VJG QRVKOCN UVCVG UGSWGPEG VJCV CEJKGXGU Õ 9G CNUQ CUUWOG VJCV
Ü YJGTG
Ü
FGPQVGU C PQTOCN FKUVTKDWVKQP CTG VJG OKZVWTG YGKIJVU
VJG OGCP XGEVQT CPF VJG EQXCTKCPEG OCVTKZ YJKEJ HQT UKORNKEKV[ KU CUUWOGF VQ DG FKCIQPCN KG +V OC[ DG FGUKTCDNG VQ OCKPVCKP VJG QTKIKPCN EQPUVTCKPVU KP VJG *// CU RTQDCDKNKV[ OGCUWTG UWEJ CU VJG HWPEVKQP DGKPI PQPPGICVKXG HQT CNN CPF HQT CNN CPF GVE #NUQ YG CUUWOG 6JG HQNNQYKPI RCTCOGVGT
VTCPUHQTOCVKQPU CNNQY WU VQ OCKPVCKP VJGUG EQPUVTCKPVU FWTKPI RCTCOGVGT CFCRVCVKQP
YJGTG YJGTG
CPF
KP VJG VTCKPKPI UGV FKUETKOKPCVKXG CFLWUVOGPV QH VJG
E\&5&3UHVV//&
+V ECP DG UJQYP VJCV HQT OGCP XGEVQT HQNNQYU
YJGTG
Æ
Ü
CPF
Ü
Ü
YJGTG KU VJG EGPVGT UNQRG QH VJG GZRQPGPVKCN UKIOQKF HWPEVKQP HQT CU FGſPGF KP
CPF Æ FGPQVGU VJG -TQPGEMGT FGNVC HWPEVKQP (KPCNN[
5KOKNCTN[ HQT VJG XCTKCPEG
YJGTG
Ü
Æ
Ü
Ü
(KPCNN[
5KOKNCT FGTKXCVKQPU HQT VJG VTCPUKVKQP RTQDCDKNKVKGU CPF VJG OKZVWTG YGKIJVU ECP DG GCUKN[ CEEQORNKUJGF = ?
E\&5&3UHVV//&
#U OGPVKQPGF GCTNKGT VJG )2& CNIQTKVJO KU C ITCFKGPV DCUGF CPF WPEQPUVTCKPGF OKP KOK\CVKQP OGVJQF +P QTFGT VQ WUG HQT FKUETKOKPCPV HWPEVKQPU HTQO EGTVCKP HCOKNKGU UWEJ CU RTQDCDKNKV[ FGPUKV[ HWPEVKQPU HTQO *//U ECTG OWUV DG VCMGP UWEJ VJCV VJQUG RTQDCDKNKUVKE EQPUVTCKPVU CTG OCKPVCKPGF 6TCPUHQTOCVKQPU KP 'SU CTG WUGF HQT VJKU RWTRQUG #PQVJGT KORQTVCPV CPF RGTJCRU VJG OQUV FKHſEWNV KUUWG KP )2& DCUGF NQUU HWPEVKQP OKPKOK\CVKQP CRRTQCEJ KU JQY VQ FGUKIP VJG UVGR UK\G 1PG CRRCTGPV TGCUQP KU VJCV YG PGGF C IQQF UVGR UK\G VQ UVCTV YKVJ UKPEG VJG OQFGN CFCRVCVKQP YKNN DG RGTHQTOGF QPN[ C ſPKVG PWODGT QH VKOGU +H VJG UVGR UK\G KU VQQ NCTIG VJG ENCUUKſ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ſGT JCXG FKHHGTGPV UGPUKVKXKVKGU VQ VJG UVGR UK\G KP RCTCOGVGT CFCRVCVKQP 1PG UVGR UK\G ECP DG VQQ UOCNN HQT UQOG RCTCOGVGTU CPF VQQ NCTIG HQT QVJGTU +P RCTVKEWNCT VJG OCIPKVWFG QH XCTKCPEGU KP VJG OKZVWTG )CWUUKCP QDUGTXCVKQP FGPUKVKGU ECP XCT[ KP VJG TCPIG DGVYGGP ½¼¼ VQ ½¼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ſGT VQYCTFU UQOG URGEKſE RQRWNCVKQPU 9KVJ VJG CFXCPEG QH OKETQRTQEGUUQTU UWEJ OGVJQFU JCXG DGEQOG EQORWVCVKQPCNN[ HGCUKDNG 6JG CDQXGOGPVKQPGF UGIOGPVCN )2& CNIQTKVJO ſPFU OCP[ CRRNKECVKQPU KP URGGEJ TGEQIPKVKQP 6JG TGEQIPKVKQP RGTHQTOCPEG CFXCP VCIG QXGT VJG VTCFKVKQPCN FKUVTKDWVKQP DCUGF /. CRRTQCEJ KU TGRQTVGF HTQO XCTKQWU UKVGU CPF KP FKHHGTGPV CRRNKECVKQPU = ? 6JG UWEEGUU QH VJKU CNIQTKVJO KP URGGEJ TGEQIPKVKQP RTQXKFGU GZRGTKOGPVCN GXKFGPEG VJCV ENCUUKſGT FGUKIP DCUGF QP GTTQT TCVG OKPKOK\CVKQP KU HGCUKDNG GXGP HQT F[PCOKE RCVVGTPU CU FKHſEWNV CU URGGEJ +ORTQXGF TGEQIPKVKQP RGTHQTOCPEG QXGT VTCFKVKQPCN FKUVTKDWVKQP DCUGF /. CRRTQCEJ CTG CNUQ TGRQTVGF KP CTGCU QWVUKFG QH URGGEJ UWEJ CU 1%4 KOCIG TGEQIPKVKQP CPF JCPFYTKVKPI TGEQIPKVKQP =?
1.5.5 Relation between MCE and MMI +P CFFKVKQP VQ VJG /%' ETKVGTKQP QVJGT ETKVGTKC CTG CNUQ WUGF KP UQECNNGF FKUETKOKPC VKXG ENCUUKſGT FGUKIP (QT *// DCUGF U[UVGOU VJG ETKVGTKC QH OCZKOWO OWVWCN KPHQTOCVKQP //+ =? EQPFKVKQPCN OCZKOWO NKMGNKJQQF GUVKOCVG %/.' =?
E\&5&3UHVV//&
OKPKOWO FKUETKOKPCVKQP KPHQTOCVKQP /&+ =? CPF *ETKVGTKC =? CTG QVJGT CN VGTPCVKXGU YJKEJ JCXG HQWPF VJGKT WUG KP URGGEJ TGEQIPKVKQP #OQPI VJGO //+ KU OQUV RQRWNCT CPF CRRNKGF KP OCP[ CRRNKECVKQPU YKVJ UWEEGUU 6JG //+ CRRTQCEJ KU DCUGF QP VJG OWVWCN KPHQTOCVKQP DGVYGGP VJG CEQWUVKE QDUGTXCVKQP : CPF KVU EQTTGEV NGZKECN U[ODQN (QT VJG ENCUU ENCUUKſECVKQP RTQDNGO VJG NQICTKVJO QH VJG OWVWCN KPHQTOCVKQP JCU VJG HQNNQYKPI HQTO
YJGTG TWPU QXGT CNN RQUUKDNG ENCUU U[ODQNU CPF CTG VJG NQINKMGNKJQQF UEQTGU QH QP VJG EQTTGEV NGZKECN U[ODQN CPF VJG VJ NGZKECN U[ODQN TGURGEVKXGN[ (TQO 'S
YJKEJ TGNCVGU VQ VJG RQUVGTKQT RTQDCDKNKV[ +P //+ VTCKPKPI VJG ETKVGTKQP QH VJG ENCUUKſGT FGUKIP CPF RCTCOGVGT GUVKOCVKQP KU VQ OCZKOK\G VJG CXGTCIG OWVWCN KPHQTOCVKQP QP VJG VTCKPKPI UGV 'ZRGTKOGPVCN URGGEJ TGEQIPKVKQP TGUWNVU KPFKECVG VJCV ENCUUKſGT FGUKIP DCUGF QP VJG //+ ETKVGTKQP ECP NGCF VQ DGV VGT TGEQIPKVKQP RGTHQTOCPEG VJCP VJG EQPXGPVKQPCN CRRTQCEJ WUKPI VJG /. ETKVGTKQP = ? #NVJQWIJ VJKU ETKVGTKQP KU YGNN HQWPFGF KP KPHQTOCVKQP VJGQT[ RQUUGUUKPI IQQF VJGQTGVKECN RTQRGTVKGU CPF WPKSWG KP OCP[ YC[U KV KU PQV DCUGF QP C FKTGEV OKP KOK\CVKQP QH VJG ENCUUKſECVKQP GTTQT TCVG CPF KU SWKVG FKHHGTGPV HTQO VJG /%' DCUGF CRRTQCEJ 6JG TGNCVKQP DGVYGGP //+ CPF /%' KU C XGT[ KPVGTGUVKPI VQRKE CPF UVWF KGF KP = ? +V KU HQWPF VJCV WPFGT EGTVCKP EQPFKVKQPU FKTGEV EQORCTKUQPU ECP DG OCFG DGVYGGP VJGUG VYQ CRRTQCEJGU 6JG FKUEWUUKQP DGNQY KU DCUGF QP =? 9G FGTKXG VJG GZRNKEKV TGNCVKQPU DGVYGGP /%' CPF //+ CPF HTQO VJGTG RTQRGTVKGU QH DQVJ CRRTQCEJGU ECP DG KNNWUVTCVGF .GV WU CUUWOG C ECUG QH WUKPI C WPKHQTON[ FKUVTKDWVGF NCPIWCIG OQFGN 6JG OWVWCN KPHQTOCVKQP HTQO 'S KU IKXGP D[
CPF VJG //+ OQFGN RCTCOGVGT GUVKOCVKQP ETKVGTKQP KU
6JG EQTTGURQPFKPI /%' CRRTQCEJ WUKPI VJG UCOG CPF CU KP 'S
JCU VJG HQNNQYKPI HQTO
E\&5&3UHVV//&
¯ 6JG OKUENCUUKſECVKQP OGCUWTG
¯ 6JG /%' OQFGN RCTCOGVGT GUVKOCVKQP KU
YJGTG
%QPUKFGT VJG URGEKCN ECUG QH
YKVJ
VJG HQNNQYKPI CNIGDTCKE TGNCVKQPU ECP DG FGTKXGF
CPF
(TQO 'S VJG NQICTKVJO QH VJG OWVWCN KPHQTOCVKQP ECP DG GZRTGUUGF DCUGF QP VJG OKUENCUUKſECVKQP OGCUWTG QH 'S KP VJG /%' HQTOWNCVKQP
(TQO 'S KV ECP DG UGGP VJCV VJG NQUU HWPEVKQP KP //+ KU XGT[ FKHHGTGPV HTQO VJG QPG WUGF KP /%' CRRTQCEJ +P CFFKVKQP VQ C EQPUVCPV UJKHV KV KU VJG NQIC TKVJO QH VJG UKIOQKF HWPEVKQP QP VJG OKUENCUUKſECVKQP OGCUWTG PQV QP VJG UKIOQKF HWPEVKQP KVUGNH CU KP /%' CRRTQCEJ 6JG F[PCOKE TCPIG QH VJKU NQUU HWPEVKQP KU HTQO VQ ½ +V KU CRRCTGPV VJCV VJG //+ QDLGEVKXG HWPEVKQP KU PQV VT[KPI VQ CRRTQZKOCVG VJG TGEQIPKVKQP GTTQT TCVG HWPEVKQP CPF KVŏU QRVKOCNKV[ VQ ENCUUKſGT FGUKIP ECPPQV DG FKTGEVN[ GUVCDNKUJGF HTQO VJG GTTQT TCVG OKPKOK\CVKQP EQPUKFGTCVKQP (TQO VJG FKUETKOKPCPV HWPEVKQP RQKPV QH XKGY //+ KU VQ OKPKOK\G VJG CXGTCIG OKU ENCUUKſECVKQP OGCUWTG KH KP 'S KU CRRTQZKOCVGF D[ 6JKU KPVGTRTGVCVKQP KU SWKVG FKHHGTGPV HTQO VJG RQUVGTKQT RTQDCDKNKV[ DCUGF KPVGTRTGVCVKQP FGUETKDGF KP 'S YJKEJ KU DCUGF QP FKUVTKDWVKQPCN CUUWORVKQPU +V GZRNCKPU VJG GZRGTKOGPVCN TGEQIPKVKQP RGTHQTOCPEG KORTQXGOGPVU QDVCKPGF HTQO VJG //+ CRRTQCEJ GXGP KP VJG ECUG YJGTG VJG FKUVTKDWVKQPCN CUUWORVKQP KU MPQYP PQV XCNKF
E\&5&3UHVV//&
0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1 -10
-8
-6
-4
-2
0
2
4
6
8
10
FIGURE 1.1 A plot of the value of the derivative of the sigmoid function. $CUGF QP 'SU CPF VJG /%' QDLGEVKXG HWPEVKQP KU U[OOGVTKECN CTQWPF VJG OKUENCUUKſECVKQP OGCUWTG YJGTGCU VJG QDLGEVKXG HWPEVKQP KP //+ KU CU[O OGVTKECN (QT EQTTGEV TGEQIPKVKQP YJGTG CPF KU JKIJGT VJCP VJG CX GTCIG KPEQTTGEV EQORGVKVKXG ECPFKFCVGU DQVJ VJG /%' CPF //+ QDLGEVKXG HWPEVKQPU CTG DQWPFGF 6JG[ FKHHGT VJQWIJ KP VJGKT UGPUKVKXKV[ VQ VJG UKIP EJCPIGU KP 6JG QDLGEVKXG HWPEVKQP KP VJG /%' CRRTQCEJ KU FKTGEVN[ TGNCVGF VQ VJG UKIP EJCPIGU KP VJG OKUENCUUKſECVKQP OGCUWTG +P VJG //+ QDLGEVKXG HWPEVKQP VJG UKIP EJCPIG KP YKNN PQV NGCF VQ C EQTTGURQPFKPI EJCPIG KP VJG UKIP QH WPNGUU KU UOCNNGT VJCP 5KPEG KU VJG PWODGT QH ENCUUGU VJKU ECP JCRRGP KP //+ DCUGF CRRTQCEJ QPN[ KH 9JGP KV KPFK ECVGU C TGEQIPKVKQP GTTQT KU EQOOKVVGF D[ VJG TGEQIPK\GT QP VJG TCPFQO KPRWV 6JG QDLGEVKXG HWPEVKQP KP VJG /%' CRRTQCEJ KU DQWPFGF PQ OCVVGT VJG XCNWG QH #U C EQPVTCUV VJG QDLGEVKXG HWPEVKQP KP //+ KU PQV DQWPFGF HQT 6JKU DGJCXKQT OC[ JCXG UQOG CFXGTUG GHHGEVU KP //+ DCUGF RCTCOGVGT GUVKOCVKQP UKPEG KV KU DCUGF QP VJG OWVWCN KPHQTOCVKQP CXGTCIGF QXGT VJG GPVKTG VTCKPKPI UGV (WTVJGT KPUKIJVU ECP DG ICKPGF D[ GZCOKPKPI VJG ITCFKGPV QH VJG QDLGEVKXG HWPEVKQPU CUUQEKCVGF YKVJ VJGUG VYQ CRRTQCEJGU 6JG ITCFKGPV QH VJG QDLGEVKXG HWPEVKQP KP /%' CRRTQCEJ JCU VJG HQNNQYKPI HQTO
YKVJ
E\&5&3UHVV//&
¼
YJGTG KU VJG FGTKXCVKXG QH VJG UKIOQKF HWPEVKQP 1P VJG QVJGT JCPF VJG ITCFKGPV QH VJG QDLGEVKXG HWPEVKQP KP VJG //+ CRRTQCEJ DCUGF QP VJG OKUENCUUKſECVKQP OGCUWTG KU
+V ECP DG UGGP VJCV VJG ITCFKGPV HWPEVKQP KP VJG /%' CRRTQCEJ KU DCUGF QP VJG FKH HGTGPVKCVGF UKIOQKF HWPEVKQP YJKEJ KU EQPEGPVTCVGF QP VJG ENCUU FGEKUKQP DQWPFCT[ 6JG CDUQNWVG XCNWG QH VJG ITCFKGPV HWPEVKQP FGETGCUGU OQPQVQPKECNN[ KH VJG XCNWG QH OQXGU CYC[ HTQO VJG FGEKUKQP DQWPFCT[ 1P VJG QVJGT JCPF VJG ITCFKGPV QH VJG QDLGEVKXG HWPEVKQP KP //+ CRRTQCEJ KU VJG UKIOQKF HWPEVKQP KVUGNH C HWPEVKQP YJKEJ KU OQPQVQPKE KPETGCUKPI CPF RWVU GORJCUKU QP GZVTGOG HCNUG ENCUUK ſECVKQPU 'ZVTGOG HCNUG ENCUUKſECVKQPU CTG V[RKECNN[ QWVNKGTU 9KVJQWV RTQRGT EQPVTQN VJG RCTCOGVGT GUVKOCVKQP ECP DG UVTQPIN[ KPƀWGPEGF D[ VJG QWVNKGTU CPF VJG GUVKOCVKQP TGUWNVU OC[ DGEQOG DKCUGF 6JKU RTQDNGO ECP DG CEWVG KP URGGEJ TGEQIPKVKQP UKPEG UWEJ QWVNKGTU CTG QHVGP HTQO YTQPI NCDGNKPI CPF GZVTGOG OKUOCVEJ KP CEQWUVKE EQPFKVKQPU 2CTCOGVGT GUVKOCVKQP KP FKUETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ KU VQ ſ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ſGT VTCKPKPI YGTG RGTHQTOGF DCUGF QP KFGPVKECN GZRGTKOGPVCN UGVWRU CPF WUKPI VJG UCOG ITQYVJVTCPUHQTOCVKQP DCUGF QRVKOK\CVKQP OGVJQF HQT RCTCOGVGT GU VKOCVKQP +V YCU HQWPF VJCV DQVJ //+ CPF /%' ECP NGCF VQ URGGEJ TGEQIPKVKQP RGTHQTOCPEG KORTQXGOGPVU QXGT VJG /. DCUGF CRRTQCEJ CPF VJG CDUQNWVG GTTQT TCVG TGFWEVKQP KP VJG /%' CRRTQCEJ KU XGTUWU KP VJG //+ CRRTQCEJ +V UJQWNF DG RQKPVGF QWV VJCV CNVJQWIJ GZRNKEKV TGNCVKQP DGVYGGP VJG //+ CPF /%' CRRTQCEJGU ECP DG GUVCDNKUJGF WUKPI VJG OKUENCUUKſECVKQP OGCUWTG CPF WPFGT UQOG URGEKCN EQP FKVKQPU VJG VJGQTGVKECN UVWF[ CU YGNN CU OQTG GZVGPUKXG GZRGTKOGPVCN UVWFKGU DGVYGGP VJGUG VYQ ETKVGTKC CTG HCT HTQO EQORNGVG CPF OCP[ SWGUVKQPU TGOCKP VQ DG CPUYGTGF
1.5.6 Discussions and Comments 6JG /%' CRRTQCEJ FGUETKDGF KP VJKU UGEVKQP KU C FKUETKOKPCPV HWPEVKQP DCUGF CR RTQCEJ VQ RCVVGTP ENCUUKſECVKQP 6JG FGEKUKQP TWNG QH VJG ENCUUKſGT KU VTGCVGF CU C FKU ETKOKPCPV HWPEVKQP CPF VJG RCTCOGVGT GUVKOCVKQP KPXQNXGU OKPKOK\KPI VJG GZRGEVGF NQUU KPEWTTGF YJGP VJGUG FGEKUKQP TWNGU CTG CRRNKGF KP VJG ENCUUKſGT 6JG HQTO QH VJG
E\&5&3UHVV//&
NQUU HWPEVKQP KU ETKVKECN KP FKUETKOKPCPV HWPEVKQP DCUGF ENCUUKGT FGUKIP +P VJG /%' CRRTQCEJ VJG NQUU HWPEVKQP KU EQPUVTWEVGF KP UWEJ C YC[ VJCV VJG TGEQIPKVKQP GTTQT TCVG QH VJG ENCUUKſGT KU GODGFFGF KP C UOQQVJ HWPEVKQPCN HQTO CPF OKPKOK\KPI VJG GZRGEVGF NQUU QH VJG ENCUUKſGT JCU C FKTGEV TGNCVKQP VQ VJG ENCUUKſGT GTTQT TCVG TGFWE VKQP 6JKU FKTGEV TGNCVKQP VQ TGEQIPKVKQP GTTQT TCVG KP VJG /%' CRRTQCEJ JCU UGXGTCN CFXCPVCIGU KP ENCUUKſGT FGUKIP
+V KU OGCPKPIHWN KP VJG UGPUG QH OKPKOK\KPI VJG GORKTKECN TGEQIPKVKQP GTTQT TCVG QH VJG ENCUUKſGT CPF VJKU RTQRGTV[ KU PQV FGRGPFGPV QP VJG RCTCOGVTKE HQTO QH VJG FKUETKOKPCPV HWPEVKQP PQT KVU TGNCVKQP VQ VJG HQTO QH VJG VTWG ENCUU RQUVGTKQT FKUVTKDWVKQP +H VJG VTWG ENCUU RQUVGTKQT FKUVTKDWVKQPU CTG WUGF CU FKUETKOKPCPV HWPEVKQPU VJG CU[ORVQVKE DGJCXKQT QH VJG ENCUUKſ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őRNWIKPŒ OGVJQF KP UVCVKUVKECN NKVGTCVWTG &GPQVG VJG FKUVTKDWVKQPU HTQO VJG UQWTEG CU CPF VJG FKUVTKDWVKQPU HTQO VJG RCTCOGVTKE OQFGN CU 6JG UVCVKUVKEU QH VJG UQWTEG IGPGTCVKPI VJG VTCKPKPI FCVC CTG PQV PGEGUUCTKN[ VJQUG QH VJG OQFGNU CPF VJG QRVKOCNKV[ TGUWNVU QH VJG $C[GU ENCUUKſ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
YJGTG KU VJG RTKQT RTQDCDKNKV[ QH VJG OVJ ENCUU CPF KU VJG -WNNDCEM.GKDNGT FKUVCPEG DGVYGGP VJG GORKTKECN FKUVTKDWVKQP QH VJG UQWTEG
E\&5&3UHVV//&
CPF VJG FKUVTKDWVKQP QH VJG RCTCOGVTKE OQFGN GUVKOCVGF HTQO VJG VTCKPKPI FCVC HQT VJG IKXGP VTCKPKPI UCORNG EQPFKVKQPGF QP VJG ENCUU NCDGN =? 6JWU C IQQFPGUU ETKVGTKQP HQT VJG /. GUVKOCVG KU KPVTQFWEGF +H VJG 2&U QH VJG OQFGN KPENWFG VJG VTWG 2&U QH VJG UQWTEG CU[ORVQVKECNN[ KV YKNN NGCF VQ C $C[GU ENCUUKſGT *QYGXGT UWEJ RTQRGTVKGU OC[ GZKUV QPN[ WPFGT C OQFGN EQTTGEVPGUU CUUWORVKQP VJCV VJG VTWG 2&U QH VJG UQWTEG CTG EQXGTGF KP VJG 2&U QH VJG RCTCOGVTKE OQFGN +H VJG 2&U QH VJG OQFGN CTG TKEJ GPQWIJ VQ RTQXKFG C IQQF CRRTQZKOCVKQP VQ VJG VTWG 2&U QH VJG UQWTEG YKVJ UWHſEKGPV VTCKPKPI FCVC VJG TGEQIPKVKQP RGTHQTOCPEG OC[ KORTQXG CU KV TGUWNVU KP C DGVVGT CRRTQZKOCVKQP VQ VJG VTWG 2&U QH VJG UQWTEG VJTQWIJ VJG GORKTKECN FKUVTKDWVKQP DCUGF QP VJG VTCKPKPI FCVC $WV KH VJG 2&U QH VJG OQFGN CTG SWKVG FKHHGTGPV EQORCTKPI VQ 2&U QH VJG UQWTEG VJG CEJKGXCDNG TGEQIPK VKQP RGTHQTOCPEG OC[ DG XGT[ NKOKVGF WUKPI VJG FKUVTKDWVKQP GUVKOCVKQP CRRTQCEJ CPF FKUETKOKPCPV HWPEVKQP DCUGF ENCUUKſ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ſECPV KORTQXGOGPVU KP TGEQIPKVKQP RGTHQTOCPEG QXGT VJG /. DCUGF CR RTQCEJ 6JG UKIPKſECPEG QH VJG /%' CRRTQCEJ KP URGGEJ TGEQIPKVKQP KU VYQHQNF (KTUV C ENCUUKſGT FGUKIP DCUGF QP FKTGEV OKPKOK\CVKQP QH VJG TGEQIPKVKQP GTTQT TCVG KU C OGCPKPIHWN CNVGTPCVKXG VQ FKUVTKDWVKQP GUVKOCVKQP DCUGF CRRTQCEJ 5GEQPF VJG 2&U WUGF KP RCTCOGVTKE OQFGNKPI QH URGGEJ CTG XGT[ NKOKVGF EQORCTGF VQ VJG VTWG 2&U KP VJG UQWTEG CPF VJG FGEKUKQP TWNG DCUGF QP FKUETKOKPCPV HWPEVKQP CRRTQCEJ KU C TGCUQPCDNG CNVGTPCVKXG VQ VJG őRNWIKPŒ /#2 TWNG YJKEJ KU DCUGF QP VJG OQFGN EQTTGEVPGUU CUUWORVKQP
1.6 Embedded String Model Based MCE Training +P VJG CDQXG OGPVKQPGF FGXGNQROGPV QH VJG /%' VTCKPKPI HQTOCNKUO VJG WVVGTCPEG ENCUUGU (QT TGEQIPKVKQP QH EQP QDUGTXCVKQP KU CUUWOGF VQ DG HTQO QPG QH VJG VKPWQWU URGGEJ QT HQT URGGEJ TGEQIPKVKQP WUKPI UWDYQTF OQFGN WPKVU KU C EQPECVG PCVGF UVTKPI QH QDUGTXCVKQPU DGNQPIKPI VQ FKHHGTGPV ENCUUGU (QT GZCORNG C UGPVGPEG KU C UGSWGPEG QH YQTFU GCEJ QH YJKEJ KU VQ DG OQFGNGF D[ C FKUVTKDWVKQP 6JG FGEQF KPI RTQEGUU KP EQPVKPWQWU URGGEJ TGEQIPKVKQP KU VQ EQORCTG KORNKEKVN[ CNN RQUUKDNG
YQTF QT UWDYQTF UVTKPI OQFGNU CPF VJG YQTF UVTKPI YJQUG UVTKPI OQFGN JCU VJG JKIJGUV NKMGNKJQQF UEQTG KU EJQUGP CU VJG FGEQFGF UVTKPI 6JG NKMGNKJQQF UEQTG QH
E\&5&3UHVV//&
FIGURE 1.2 A structure diagram of a context dependent head-body-tail digit model in speech recognition. VJG YQTF UVTKPI KU V[RKECNN[ C EQODKPCVKQP QH UEQTGU HTQO XCTKQWU OQFGNU KPENWFKPI VJG UEQTG HTQO VJG CEQWUVKE OQFGN NCPIWCIG OQFGN FWTCVKQP OQFGN GVE 6JG OCKP TGCUQP QH CFQRVKPI VJKU V[RG QH UVTKPI OQFGN KU UKORN[ VJCV VJG DCUKE URGGEJ TGEQI PKVKQP OQFGN WPKVU YJKEJ CTG WUGF VQ HQTO UVTKPI OQFGNU ECP DG GUVKOCVGF HTQO C ſ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ſGT FGUKIP KP EQPVKPWQWU URGGEJ TGEQIPKVKQP +P VJKU UGEVKQP YG ſTUV FGUETKDG C IGPGTCN GODGFFGF UVTKPI OQFGN DCUGF /%' RCTCFKIO HQT EQPVKPWQWU URGGEJ TGEQIPKVKQP CPF HTQO VJGTG VJG /%' VTCKPKPI HQT GCEJ EQORQPGPV KP VJG WVVGTCPEG DCUGF UVTKPI OQFGN ECP DG CEJKGXGF WPFGT VJKU WPKſGF HTCOGYQTM
1.6.1 String Model Based MCE Approach &KUETKOKPCPV HWPEVKQPU DCUGF QP UVTKPI NGXGN OQFGNKPI CTG PGEGUUCT[ KP EQPVKPWQWU URGGEJ TGEQIPKVKQP DGECWUG VJG ENCUUKſGT FGEKUKQP TWNGU CTG DCUGF QP VJG YJQNG WV VGTCPEG NGXGN INQDCN OCVEJKPI 6JG UVTKPI OQFGN YJKEJ FGUETKDGU VJG IKXGP YQTF
E\&5&3UHVV//&
UVTKPI VJCV DGUV OCVEJGU VJG KPRWV URGGEJ WVVGTCPEG JCU VQ DG FGVGTOKPGF D[ 8KVGTDK CNKIPOGPV RTQEGUU DGVYGGP CNN RQUUKDNG UVTKPI OQFGNU CPF VJG KPRWV URGGEJ WVVGTCPEG ½ (QT GCUG QH TGRTGUGPVCVKQP YG FTQR VJG PQPCEQWUVKE RCTVU KP VJG UVTKPI OQFGN ſTUV CPF EQPUKFGT VJGO UGRCTCVGN[ NCVGT 6JG UVTKPI OQFGN HQT C IKXGP YQTF UVTKPI KP CP *// DCUGF URGGEJ TGEQIPKVKQP U[UVGO WUKPI EQPVKPWQWU QDUGTXCVKQP FGPUKVKGU KU IKXGP D[
YJGTG KU C RQUUKDNG UVTKPI OQFGN HQT YQTF UVTKPI 5 KU VJG QRVKOCN UVCVG UGSWGPEG KP VJG UVTKPI OQFGN QH KU VJG OQFGN UGV QH CNN TGEQIPKVKQP OQFGN WPKVU CPF KU VJG NQINKMGNKJQQF UEQTG CNQPI VJG QRVKOCN UVCVG UGSWGPEG +P VJG GODGFFGF UVTKPI OQFGN DCUGF /%' VTCKPKPI FGUETKDGF KP =? VJG FKUETKOKPCPV HWPEVKQP CV VJG UVTKPI NGXGN KU DCUGF QP VJG UVTKPI OQFGN HQT VJG EQTTGEV YQTF UVTKPI CPF VJG UVTKPI OQFGNU QH VJG OQUV EQPHWUCDNG YQTF UVTKPIU QDVCKPGF WUKPI C HCUV VTGGVTGNNKU DGUV UGCTEJ =? .GV ½ DG CP CTDKVTCT[ YQTF UVTKPI )KXGP VJG OQFGN UGV VJG QRVKOCN UVCVG UGSWGPEG KU C HWPEVKQP QH VJG QDUGTXCVKQP CPF VJG YQTF UVTKPI KU QHVGP FGVGTOKPGF D[ C 8KVGTDK FGEQFKPI RTQEGUU 6JG VQR DGUV UVTKPI J[RQVJGUGU ½ ECP DG FGſPGF KPFWEVKXGN[ CU HQNNQYU
½ ½
½
CTG
6JG FKUETKOKPCPV HWPEVKQPU HQT
YJGTG KU VJG VJ DGUV UVTKPI KU VJG *// UGV WUGF KP VJG DGUV FGEQFKPI KU VJG QRVKOCN RCVJ UVCVG UGSWGPEG QH VJG VJ UVTKPI IKXGP VJG OQFGN UGV CPF KU VJG TGNCVGF NQINKMGNKJQQF UEQTG QP VJG QRVKOCN RCVJ QH VJG
VJ UVTKPI (QT VJG EQTTGEV UVTKPI VJG FKUETKOKPCPV HWPEVKQP KU IKXGP D[
YJGTG KU VJG EQTTGEV UVTKPI KU VJG QRVKOCN CNKIPOGPV RCVJ CPF KU VJG EQTTGURQPFKPI NQINKMGNKJQQF UEQTG 6JGUG FKUETKOKPCPV HWPEVKQPU CTG GODGFFGF KP VJG /%' DCUGF NQUU HWPEVKQP VJTQWIJ VJG HQNNQYKPI UVGRU
6JG OKUENCUUKſECVKQP OGCUWTG KP GODGFFGF UVTKPI OQFGN DCUGF /%' VTCKPKPI KU FGſPGF CU
E\&5&3UHVV//&
½
5RGGEJ (GCVWTG
0$GUV5VTKPI *[RQVJGUGU&GEQFKPI
5EQTG1RVKOCN2CVJ QH0$GUV5VTKPI/QFGN
7RFCVGF *//U
5GIOGPVCN )2&6TCKPGT
*//U
FIGURE 1.3 A diagram of the embedded string model based MCE training process. 6JG NQUU HWPEVKQP KP OKPKOWO UVTKPI GTTQT TCVG VTCKPKPI KU FGſPGF CU
YJGTG KU C RQUKVKXG EQPUVCPV YJKEJ EQPVTQNU VJG UNQRG QH VJG UKIOQKF HWPE VKQP 6JG GZRGEVGF NQUU YJKEJ KU CUUQEKCVGF YKVJ VJG UVTKPI GTTQT TCVG KU IKXGP D[
+V UJQWNF DG PQVGF VJCV VJG UKVWCVKQP KP EQPVKPWQWU URGGEJ TGEQIPKVKQP KU SWKVG FKHHGT GPV HTQO VJG ſPKVG ENCUU ENCUUKſECVKQP RTQDNGO YJGTG C ſZGF UGV QH FKUETKOKPCPV HWPEVKQPU ECP DG RTGURGEKſGF 6JG FKUETKOKPCPV HWPEVKQPU KP VJG GODGFFGF UVTKPI OQFGN DCUGF /%' CRRTQCEJ CTG F[PCOKE FGRGPFKPI QP VJG RCTVKEWNCT NGZKECN YQTF UVTKPI TCPFQO KPRWV CPF C NKUV QH VJG OQUV EQORGVKVKXG UVTKPI OQFGNU 6JG OQUV EQORGVKVKXG UVTKPI OQFGNU CNUQ FGRGPF QP VJG UVTKPI NGXGN OQFGN OCVEJKPI QH VJG WVVGTCPEG CICKPUV VJG EWTTGPV OQFGN UGV +P VJG /. DCUGF FKUVTKDWVKQP GUVKOCVKQP CRRTQCEJ VJG OQFGN RCTCOGVGTU CTG GUVKOCVGF QPN[ HTQO VJG VTCKPKPI FCVC YKVJ VJG EQTTGEV UVTKPI OQFGN 6JG FKUETKOKPCVKXG KPHQTOCVKQP GZKUVKPI KP VJG EQORGV KPI UVTKPI OQFGNU KU IGPGTCNN[ PQV WUGF 6JG WUG QH VJG UGSWGPVKCN VTCKPKPI RTQEGFWTG DCUGF QP )2& CNIQTKVJO HQT RCTCOGVGT CFCRVCVKQP CNUQ OCMGU VJKU VTCKPKPI RTQEG FWTG őUGIOGPVCNŒ KP VJG UGPUG VJCV VJG UVCVG UGIOGPVCVKQP QH VJG URGGEJ WVVGTCPEG KU WUGF VQ WRFCVG VJG EWTTGPV OQFGN CPF VJG WRFCVGF OQFGN KU WUGF VQ KPVTQFWEG PGY UGIOGPVCVKQP HQT VJG PGZV VTCKPKPI UCORNG
E\&5&3UHVV//&
1PG QH VJG KUUWGU KP CEQWUVKE OQFGNKPI KU JQY VQ OQFGN VJG YQTF UVTKPIU VJCV CTG PQV KP VJG VTCKPKPI UGV +P EQPVKPWQWU URGGEJ TGEQIPKVKQP VJG EQXGTCIG QH VJG VTCKPKPI OCVGTKCN QP VJG RQUUKDNG YQTF UVTKPIU KU CNYC[U NKOKVGF IKXGP VJG HCEV VJCV C JWIG PWODGT QH YQTF UVTKPIU ECP QEEWT KP VJG NCPIWCIG 6JGUG WPUGGP UVTKPIU CTG KP IGP GTCN XGT[ JCTF VQ OQFGN CPF /. GUVKOCVKQP KU DCUGF QP VJG UGGP VTCKPKPI FCVC CPF ECPPQV EQXGT VJG ECUGU YJKEJ CTG WPUGGP 6JG WUG QH EQORGVKPI UVTKPI OQFGNU KP /%' VTCKPKPI RTQXKFGU C DGVVGT EQXGTCIG QH YQTF UVTKPIU UKPEG OCP[ QH VJGO OC[ PQV CEVWCNN[ QEEWT KP VJG VTCKPKPI FCVC 6JQUG EQPHWUCDNG YQTF UVTKPIU CTG UGNGEVGF DCUGF QP VJGKT EQPHWUKDKNKV[ YKVJ VJG EQTTGEV NGZKECN UVTKPI IKXGP VJG EWTTGPV OQFGN UGV 6JG[ CTG WUGF VQ HQTO VJG UVTKPI OQFGN DCUGF FKUETKOKPCPV OGCUWTG CPF OQFGNGF KP VJG UOQQVJ /%' DCUGF NQUU HWPEVKQP YJKEJ TGNCVGU VQ UVTKPI GTTQT TCVG # FKCITCO QH VJG GODGFFGF UVTKPI OQFGN DCUGF /%' VTCKPKPI KU IKXGP KP (KI 6JG GODGFFGF UVTKPI OQFGN DCUGF /%' CRRTQCEJ KU YGNN UWKVGF HQT CEQWUVKE OQFGN KPI WUKPI FGVCKNGF EQPVGZV FGRGPFGPV OQFGNU YJGTG UGRCTCVG CEQWUVKE OQFGN WPKV KU WUGF VQ OQFGN RJQPGOG YKVJ FKHHGTGPV NGHV CPF TKIJV EQPVGZV +V ECP FGUETKDG XCTKQWU NQPI URCP NGHV CPF TKIJV EQPVGZV FGRGPFGPEKGU UWEJ CU VTKRJQPG SWKPRJQPG GVE 1PG GZCORNG QH C ETQUUYQTF EQPVGZV FGRGPFGPV OQFGN WUGF KP EQPPGEVGF FKIKV TGEQIPK VKQP KU FGRKEVGF KP (KI YJKEJ JCU C HWNN GZRCPUKQP HQT CNN RQUUKDNG NGHV CPF TKIJV EQPVGZVU CV VJG YQTF DQWPFCTKGU 6JG KPVTQFWEVKQP QH VJG GODGFFGF UVTKPI OQFGN KP /%' VTCKPKPI JCU VYQ CFXCPVCIGU
¯ +V GZVGPFU VJG /%' DCUGF FKUETKOKPCPV HWPEVKQP CRRTQCEJ VQ EQPVKPWQWU URGGEJ TGEQIPKVKQP YJGTG OQFGNKPI GCEJ KPFKXKFWCN YQTF UVTKPI ENCUU KU PQV HGCUKDNG ¯ +V RTQXKFGU CP GZCEV GOWNCVKQP QH VJG ENCUUKſGT KP EQPVKPWQWU URGGEJ TGEQIPK VKQP CPF GODGFU VJG WVVGTCPEG NGXGN URGGEJ OCPKHGUVCVKQP KP VJG DCUKE TGEQIPK VKQP OQFGN WPKVU +P GODGFFGF UVTKPI OQFGN DCUGF /%' CRRTQCEJ VJG NQPI VGTO FGRGPFGPEKGU CTG GODGFFGF KP VJG DCUKE URGGEJ TGEQIPKVKQP OQFGN WPKVU GXGP KH VJGKT QTKIKPCN EQPVGZV FGRGPFGPE[ FGſPKVKQPU CTG PQV +V KU QDUGTXGF KP VJG GZRGTKOGPVU VJCV OCP[ OQPQ RJQPG DCUGF EQPVGZV KPFGRGPFGPV OQFGN WPKVU QDVCKPGF HTQO VJG /%' CRRTQCEJ GZ JKDKV URGGEJ TGEQIPKVKQP RGTHQTOCPEG QH EQPVGZV FGRGPFGPV OQFGN WPKVU =? 6JG GODGFFGF UVTKPI OQFGN DCUGF /%' CRRTQCEJ HQWPF CRRNKECVKQPU KP XCTKQWU TGEQIPK VKQP VCUMU CPF UKIPKſECPV GTTQT TCVG TGFWEVKQP YGTG QDUGTXGF = ? #NVJQWIJ VJG UVTKPI OQFGN DCUGF CRRTQCEJ KU VJG PCVWTCN EJQKEG HQT UVTKPI GTTQT TCVG OKPKOK\CVKQP KV KU RQUUKDNG VQ KPENWFG YQTF NGXGN GTTQT GHHGEVU KP /%' VTCKPKPI 1PG OQFKſECVKQP RTQRQUGF KP =? WUGU YQTF GTTQT EQWPVU CU VJG YGKIJVU DGVYGGP VJG EQTTGEV NGZKECN UVTKPI OQFGN CPF VJG OQUV EQPHWUCDNG UVTKPI OQFGN KP VJG OKUENCUUKſ ECVKQP OGCUWTG .GV VJG NGZKECN UVTKPI OQFGN DG CPF VJG OQUV EQPHWUCDNG UVTKPI OQFGN ½ VJG OKUENCUUKſECVKQP OGCUWTG YKVJ VJG YQTF GTTQT EQWPV YGKIJVKPI JCU VJG HQNNQYKPI HQTO
LD ½ ½
YJGTG LD ½ KU VJG UQ ECNNGF .GXGPUJVGKPFKUVCPEG DGVYGGP VJG EQTTGEV TGHGT GPEG YQTF UVTKPI CPF VJG TKXCN YQTF UVTKPI ½ KG VJG PWODGT QH GTTQTU EQP VCKPGF KP ½ 6JG TGUV QH VJKU /%' HQTOWNCVKQP HQNNQYU VJG GODGFFGF UVTKPI GTTQT
E\&5&3UHVV//&
DCUGF /%' CRRTQCEJ *QYGXGT KV UJQWNF DG PQVGF VJCV OWNVKRN[KPI C RQUKVKXG EQP UVCPV QP VJG OKUENCUUKſECVKQP OGCUWTG FQGU PQV EJCPIG KVU UKIP PQT VJG UVTKPI GTTQT DCUGF /%' HQTOWNCVKQP 9JGP VJG NQUU HWPEVKQP YKNN UVKNN EQPXGTIG VQ VJG UVTKPI GTTQT EQWPV HWPEVKQP CNVJQWIJ C YQTF GTTQT DCUGF YGKIJVKPI KU CRRNKGF
1.6.2 Combined String Model Based MCE Approach #U OGPVKQPGF CV VJG DGIKPPKPI QH VJKU UGEVKQP VJG ſPCN FGEKUKQP KP URGGEJ TGEQIPK VKQP KU DCUGF QP VJG EQODKPCVKQP QH UEQTGU HTQO XCTKQWU MPQYNGFIG UQWTEGU TGRTG UGPVGF D[ FKHHGTGPV OQFGNU #UUWOKPI KPFGRGPFGPEG QH GCEJ OQFGN VJG ſPCN UEQTG KP VJG NQICTKVJO FQOCKP DGEQOGU C UWO QH NQINKMGNKJQQF UEQTGU HTQO GCEJ KPFKXKF WCN OQFGN +P RCTVKEWNCT KP CFFKVKQP VQ VJG CEQWUVKE OQFGN KH C NCPIWCIG OQFGN KU WUGF CPF KVU UEQTG KU YGKIJVGF D[ C YGKIJVKPI HCEVQT VJG ſPCN NKMGNKJQQF UEQTG QH C ECPFKFCVG UVTKPI KU
+H VJG OQFGN EQTTGEVPGUU CUUWORVKQP KU XCNKF VJG NQINKMGNKJQQF UEQTG UJQWNF UVTKEVN[ HQNNQY 'S CPF VJG UEQTG YGKIJVKPI HCEVQT *QYGXGT KP URGGEJ TGEQI PKVKQP GZRGTKOGPVU CPF CRRNKECVKQPU KV KU HQWPF VJCV C XCNWG QH YKVJ FGOQPUVTCVGU OWEJ DGVVGT TGEQIPKVKQP RGTHQTOCPEG =? CP KPFKECVKQP VJCV VJG VTWG FKUVTKDWVKQP QH VJG UKIPCN UQWTEG FGRCTVU HTQO VJG CUUWORVKQP OCFG D[ VJG OQFGN 6JG NCPIWCIG OQFGN HCEVQT KP URGGEJ TGEQIPKVKQP KU QHVGP VWPGF CPF CFLWUVGF DCUGF QP VJG TGEQIPKVKQP TGUWNVU QP VJG VTCKPKPI CPF FGXGNQROGPV FCVC 6JG CEVWCN XCNWG QH VJG NCPIWCIG OQFGN HCEVQT WUGF KP TGEQIPKVKQP KU SWKVG FKHHGTGPV HTQO VJG QPG FGTKXGF HTQO VJG OQFGN EQTTGEVPGUU CUUWORVKQP 6JG VWPKPI RTQEGFWTG KVUGNH DGKPI GORKTK ECN KU C FGRCTVWTG HTQO VJG FKUVTKDWVKQP DCUGF RCVVGTP TGEQIPKVKQP CRRTQCEJ CPF VJG UEQTG EQODKPCVKQP ECP DG EQPUKFGTGF CU C RTQDNGO QH UGNGEVKPI FKUETKOKPCPV HWPE VKQPU KP RCVVGTP ENCUUKſECVKQP 6JG KPVTQFWEVKQP QH GODGFFGF UVTKPI OQFGN OCMGU KV RQUUKDNG VQ GZVGPF VJG FKUETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ VQ VJG NGXGN QH JCPFNKPI OWNVKOQFGN EQODKPCVKQPU CPF VQ VJG RCTCFKIO QH EQODKPGF UVTKPI OQFGN VTCKPKPI 6JG EQODKPGF UVTKPI OQFGN ECP DG FQPG KP VJG EQODKPCVKQP QH VJG HQNNQYKPI VYQ FKTGEVKQPU 1PG KU JQTK\QPVCN UEQTGU HTQO OWNVKRNG OQFGNU CPF FKHHGTGPV MPQYNGFIG UQWTEGU CTG EQODKPGF VQ HQTO VJG ſPCN UEQTG YJGTG GCEJ KPFKXKFWCN OQFGN OC[ DG GUVKOCVGF UGRCTCVGN[ DCUGF QP FKHHGTGPV GUVKOCVKQP OGVJQFU KPENWFKPI WUKPI FKHHGTGPV VTCKPKPI FCVC CPF EQPUVTCKPVU &KUETKOKPCVKXG OQFGN EQODKPCVKQP =? HCNNU KP VJKU ECV GIQT[ #PQVJGT KORQTVCPV FKTGEVKQP HQT OQFGN EQODKPCVKQP KU VQ GUVKOCVG VJG KPFKXKF WCN OQFGN RCTCOGVGTU KP VJG EQODKPGF UVTKPI OQFGN CU CP KPVGITCVGF EQORQPGPV QH VJG ſPCN EQODKPGF UVTKPI OQFGN &KUETKOKPCVKXG HGCVWTG GZVTCEVKQP =? FKUETKOKPCVKXG NCPIWCIG OQFGN GUVKOCVKQP = ? CPF GODGFFGF UVTKPI OQFGN DCUGF GUVKOCVKQP WUKPI OWNVKRNG MPQYNGFIG UQWTEGU =? CTG UWEJ CRRTQCEJGU KP YJKEJ VJG FKUETKOK PCPV HWPEVKQP KU EQPUVTWEVGF CV VJG EQODKPGF UVTKPI OQFGN NGXGN CPF VJG GUVKOCVKQP QH RCTCOGVGTU CV GCEJ KPFKXKFWCN OQFGN KU CEJKGXGF D[ VTCEKPI FQYP VJG OQFGN EQODKPC VKQP VTGG VQ GCEJ QH KVU NGCH PQFGU HQNNQYKPI C EJCKP TWNG NKMG TGNCVKQPUJKR #NVJQWIJ KV ECP DG EQORWVCVKQPCNN[ FGOCPFKPI VQ GUVKOCVG CNN OQFGN RCTCOGVGTU KP UWEJ C INQDCN
E\&5&3UHVV//&
OCPPGT VJG EQODKPGF UVTKPI OQFGN DCUGF CRRTQCEJ PGXGTVJGNGUU RTQXKFGU CP GZCEV EJCTCEVGTK\CVKQP QH VJG FGEKUKQP RTQEGUU HQT GXGP VJG OQUV UQRJKUVKECVGF TGEQIPKVKQP CRRNKECVKQPU CPF KV KU CRRNKGF UWEEGUUHWNN[ KP OCP[ URGGEJ TGEQIPKVKQP U[UVGOU +P QTFGT VQ TGFWEG VJG EQORWVCVKQPCN EQORNGZKV[ OQFGN VTCKPKPI ECP DG FQPG KP C UGNGE VKXG YC[ YJGTG UQOG RQTVKQP QH VJG EQODKPGF UVTKPI OQFGN KU CUUWOGF ſZGF YJKNG GUVKOCVKPI RCTCOGVGTU KP QVJGT UGNGEVGF EQORQPGPVU QH VJG EQODKPGF UVTKPI OQFGN 6JG VTCKPKPI RTQEGUU KU QHVGP KVGTCVGF UGXGTCN VKOGU QP VJG VTCKPKPI FCVC YJGTG FKHHGT GPV OQFGN EQORQPGPVU CTG UGNGEVGF CV GCEJ KVGTCVKQP =? 6JKU KPVGITCVGF CRRTQCEJ YKNN DG HWTVJGT GZGORNKſGF KP VJG HQNNQYKPI UWDUGEVKQPU 1.6.2.1 Discriminative Model Combination 6JG CDKNKV[ VQ EQODKPG OWNVKRNG OQFGNU HTQO XCTKQWU MPQYNGFIG UQWTEGU KP URGGEJ TGEQIPKVKQP KU KORQTVCPV 6JKU KU DGECWUG URGGEJ KU C EQORNKECVGF UQWTEG CPF ECP DG CHHGEVGF D[ OCP[ HCEVQTU UWEJ CU EQPVGZV RTQUQFKEU XQECN VTCEV NGPIVJ CODK GPV GPXKTQPOGPV URGCMKPI UV[NG OQFG QH VJG URGCMGT CEEGPV GVE /WNVKRNG UKIPCN UQWTEGU HTQO OWNVKRNG UKIPCN DCPFU CTG CNUQ WUGF KP URGGEJ TGEQIPKVKQP =? /CP[ QH VJGUG OQFGNU QT MPQYNGFIG UQWTEGU OC[ PQV DG DCUGF QP RTQDCDKNKVKGU CPF C FKU ETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ KU C UWKVCDNG EJQKEG HQT OQFGN EQODKPCVKQP .GV ½ DG VJG KPFKXKFWCN OQFGN EQORQPGPVU KP VJG OQFGN EQODKPCVKQP 9G WUG VJG PQVCVKQP ½ VQ FGPQVG VJG EQODKPGF UVTKPI OQFGN IKXGP TCPFQO KPRWV YJGTG KU VJG HWPEVKQP UGNGEVGF HQT OQFGN EQODKPCVKQP +H ) KU NKPGCT
½
YJGTG KU VJG UEQTG HTQO VJG VJ OQFGN CPF KU VJG OQFGN EQODKPCVKQP YGKIJVU &KUETKOKPCVKXG OQFGN EQODKPCVKQP DCUGF QP VJG /%' CRRTQCEJ KU VQ GODGF VJG EQODKPGF UVTKPI OQFGN DCUGF FKUETKOKPCPV HWPEVKQP KP VJG NQUU HWPEVKQP CPF GUVKOCVG VJG OQFGN EQODKPCVKQP HCEVQT CU RCTCOGVGTU KP VJG EQODKPGF UVTKPI OQFGN +P RCTVKEWNCT VJG OKUENCUUKſECVKQP OGCUWTG KP VJG EQODKPGF UVTKPI OQFGN DCUGF /%' CRRTQCEJ KU
½
CPF VJG NQUU HWPEVKQP KU FGſPGF CU
YJGTG KU C RQUKVKXG EQPUVCPV YJKEJ EQPVTQNU VJG UNQRG QH VJG UKIOQKF HWPEVKQP 6Q FGVGTOKPG OQFGN EQODKPCVKQP EQGHſEKGPVU OCP[ QRVKOK\CVKQP OGVJQFU ECP DG CRRNKGF VQ GUVKOCVG YJKEJ OKPKOK\G VJG GZRGEVGF NQUU 6JG RQRWNCT )2& CNIQTKVJO JCU C XGT[ UKORNG HQTO KP VJKU ECUG =? %QPUVTCKPVU QP VJG XCNWG QH OQFGN EQODKPCVKQP EQGHſEKGPVU ECP CNUQ DG CRRNKGF FWTKPI RCTCOGVGT QRVKOK\C VKQP FGRGPFKPI QP VJG PCVWTG QH VJG MPQYNGFIG UQWTEGU WUGF KP VJG EQODKPGF UVTKPI
E\&5&3UHVV//&
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őVWPKPI RCTCOGVGTUŒ KP URGGEJ TGEQIPKVKQP GURGEKCNN[ YJGP GKVJGT C OQFGN EQTTGEVPGUU CUUWORVKQP KU PQV XCNKF QT C WPKſGF HTCOGYQTM KU PGGFGF VQ EQODKPG MPQYNGFIG UQWTEGU VJCV CTG FKHHGTGPV KP QTKIKP QT PCVWTG +V UJQWNF DG RQKPVGF QWV VJCV FKUETKOKPCVKXG OQFGN EQODKPCVKQP KU XGT[ FKHHGTGPV HTQO OGVJ QFU WUGF VQ EQODKPG TGUWNVU HTQO OWNVKRNG TGEQIPK\GTU UWEJ CU 418'4 =? +P 418'4 KV KU DCUGF QP C XQVKPI UEJGOG CPF KV WVKNK\GU VJG FKXGTUKV[ QH VJG TGEQIPK VKQP GTTQTU HTQO KPFGRGPFGPV TGEQIPKVKQP U[UVGOU VQ KORTQXG VJG TGEQIPKVKQP RGTHQT OCPEG &KUETKOKPCVKXG OQFGN EQODKPCVKQP EQODKPGU FKHHGTGPV MPQYNGFIG UQWTEGU KPVQ QPG FKUETKOKPCPV HWPEVKQP 6JG EQORQPGPV KP VJG EQODKPGF UVTKPI OQFGN OC[ PQV DG CP KPFGRGPFGPV TGEQIPK\GT CPF KV ECP DG CP[ MPQYNGFIG UQWTEG TGNCVGF VQ VJG TCPFQO KPRWV : *QYGXGT KH GCEJ EQORQPGPV KP VJG EQODKPGF UVTKPI OQFGN KU C TGEQIPK\GT DQVJ OGVJQFU CRRN[ CPF KV KU CP KPVGTGUVKPI TGUGCTEJ VQRKE VQ UGG JQY VQ KPVGITCVG VJGO VQIGVJGT VQYCTFU C OQTG FKUETKOKPCVKXG EQODKPCVKQP DCUGF QP QWVRWVU HTQO OWNVKRNG TGEQIPKVKQP U[UVGOU 1.6.2.2 Discriminative Language Model Estimation .CPIWCIG OQFGNKPI KU C ETKVKECN EQORQPGPV KP URGGEJ TGEQIPKVKQP CPF HTQO VJG UQWTEG CPF OQFGN RQKPV QH XKGY KV RTQXKFGU VJG NCPIWCIG NGXGN OQFGNKPI QH VJG UQWTEG /QTGQXGT C NQV QH YQTFU KP URGGEJ CTG CEQWUVKECNN[ UKOKNCT CPF UQOG QH VJGO JQOQRJQPGU CTG GXGP KFGPVKECN UWEJ CU “too” CPF “two” +H QPN[ DCUGF QP CEQWUVKE KPHQTOCVKQP KFGPVKſECVKQP QH VJGUG YQTFU CPF RJTCUGU KP EQPVKPWQWU URGGEJ ECP DG XGT[ FKHſEWNV CPF QVJGT MPQYNGFIG UQWTEGU KP RCTVKEWNCT C NCPIWCIG OQFGN CTG PGGFGF # UVCVKUVKECN DCUGF ITCO NCPIWCIG OQFGN KU C RQRWNCT EJQKEG KP URGGEJ TGEQIPKVKQP CPF KV JCU VJG HQNNQYKPI HQTO ´ ½ ·½ µ YJKEJ KU VJG GUVKOCVGF RTQDCDKNKV[ QH QDUGTXKPI YQTF IKXGP VJG RCUV ½ YQTF JKUVQT[ $GECWUG VJG PWODGT QH RQUUKDNG ITCO RTQDCDKNKVKGU ITQYU GZRQPGPVKCNN[ YKVJ VJG QTFGT NQYGT QTFGT NCPIWCIG OQFGNU UWEJ CU WPKITCO DKITCO VTKITCO CPF HQWT ITCO CTG WUGF KP XCTKQWU URGGEJ TGEQIPKVKQP VCUMU 6JG ITCO UVCVKUVKECN NCPIWCIG OQFGN KU V[RKECNN[ GUVKOCVGF HTQO C NCTIG VGZV EQTRWU KPFGRGPFGPV QH QVJGT MPQYNGFIG UQWTEGU KP URGGEJ TGEQIPKVKQP 'XGP YKVJ C XGT[ NCTIG VGZV EQNNGEVKQP QPN[ RQTVKQPU QH WPKSWG VTKITCO CPF HQWTITCO GPVTKGU ECP DG GUVKOCVGF FWG VQ URCTUGPGUU QH VJG VTCKPKPI FCVC /CP[ NCPIWCIG OQFGN GPVTKGU TCTGN[ QEEWT KP VJG EQTRWU CPF DCEMQHH UEJGOGU GI -CV\ DCEMQHHU =? CTG WUGF VQ UWDUVKVWVG VJG WPUGGP NCPIWCIG OQFGN GPVTKGU YKVJ VJGKT NQYGT QTFGT DCEMQHH EQWPVGTRCTVU 6JG URCTUG FCVC RTQDNGO KU C UGTKQWU KUUWG KP NCPIWCIG OQFGN GUVKOCVKQP CPF KP QTFGT VQ IGPGTCVG OQTG GPVTKGU VJG UCORNG EQWPV EWVQHH VJTGUJQNF KP NCPIWCIG OQFGN GUVKOCVKQP KU WUWCNN[ UGV XGT[ NQY OCMKPI VJG GUVKOCVG HCT HTQO DGKPI TGNKCDNG +V KU CNUQ QDXKQWU HTQO VJG NCPIWCIG RQKPV QH XKGY VJCV CNVJQWIJ VJG UVCVKUVKECN ITCO NCPIWCIG OQFGN KU SWKVG UWEEGUUHWN
E\&5&3UHVV//&
KP URGGEJ TGEQIPKVKQP KV KU PQV VJG őVTWGŒ UQWTEG OQFGN HQT URGGEJ +V KU QDXKQWU VJCV JWOCP NCPIWCIG JCU C OWEJ OQTG EQORNKECVGF UVTWEVWTG VJCV QHVGP ECPPQV DG EQXGTGF D[ VJG UKORNG UVTWEVWTG QH C ſ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ſECPV QEEWTTGPEGU QT NCPIWCIG OQFGN RCTCOGVGTU YJKEJ QEEWT QHVGP UWEJ CU NCPIWCIG OQFGN DCEMQHH YGKIJVU GVE = ? &KUETKOKPCVKXG NCPIWCIG OQFGN GUVKOCVKQP ECP CNUQ DG WUGF KP CRRNKECVKQPU YJGTG NCPIWCIG OQFGNU CTG WUGF CU ENCUUKſGTU UWEJ CU ECNN TQWVKPI =? CPF FKCNQIWG UVCVG KFGPVKſECVKQP =? 'ZRGTKOGPVCN TGUWNVU KPFKECVG VJCV FKU ETKOKPCVKXG NCPIWCIG OQFGN GUVKOCVKQP ECP KORTQXG NCPIWCIG OQFGN DCUGF ENCUUKſGTU EQORCTKPI VQ /. DCUGF NCPIWCIG OQFGN GUVKOCVKQP = ?
1.6.3 Discriminative Feature Extraction /QUV URGGEJ TGEQIPKVKQP U[UVGOU WUG UQOG V[RG QH URGEVTCN CPCN[UKU QP VJG YKPFQYGF TCY URGGEJ YCXGHQTO 5RGGEJ KU TGRTGUGPVGF CU C UGSWGPEG QH UJQTVVKOG RQYGT URGE VTC QT TGNCVGF TGEQIPKVKQP HGCVWTG XGEVQTU 6JG VYQ V[RGU QH URGEVTCN CPCN[UKU OGVJQFU OQUV HTGSWGPVN[ GORNQ[GF CTG ſNVGT DCPM CPCN[UKU CPF NKPGCT RTGFKECVKQP (KNVGT DCPM CRRTQCEJGU V[RKECNN[ WUG C DCPM QH DCPFRCUU ſNVGTU 6JG HTGSWGPE[ URCEKPI QH VJG ſNVGTU CTG GKVJGT WPKHQTO URCEGF QT ETKVKECNDCPFURCEGF HQNNQYKPI $CTM UECNG QT /GN UECNG 6JGUG ſNVGTU CTG IGPGTCNN[ JKIJN[ QXGTNCRRGF CPF EQXGT VJG TGNGXCPV HTGSWGPE[ TCPIG QH VJG KPRWV UKIPCN 6KOG CPF HTGSWGPE[ TGUQNWVKQP KU CP KORQTVCPV HCEVQT KP VJG ſNVGT DCPM FGUKIP 5RGEVTWO KPVGPUKV[ KU QHVGP UECNGF NQICTKVJOKECNN[ CPF VJG KFGC QH URGEVTWO YGKIJVKPI KU CNUQ WUGF VQ EQPVTQN VJG HGCVWTG UGPUKVKXKV[ *QYGXGT OQUV
E\&5&3UHVV//&
HGCVWTG GZVTCEVKQPU CTG DCUGF QP VJG CPCN[UKU QH JWOCP JGCTKPI ECRCDKNKVKGU CPF VJWU CTG PQV PGEGUUCTKN[ CRRNKECDNG VQ UVCVKUVKECNN[ DCUGF OCEJKPG TGEQIPKVKQP 6JG IQCN QH FKUETKOKPCVKXG HGCVWTG GZVTCEVKQP KU VQ CEEQORNKUJ VJG URGGEJ TGEQIPKVKQP HGCVWTG GZVTCEVKQP HTQO VJG UVCPFRQKPV QH OKPKOK\KPI VJG TGEQIPKVKQP GTTQT TCVG HQT ENCUUKſECVKQP D[ OCEJKPGU +P RNCEG QH VJG $CTM UECNG HTQO JGCTKPI C PGY HTGSWGPE[ UECNKPI =? ECP DG FGTKXGF DCUGF QP VJG ENCUUKſ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ſEKGPVU 6JGTGHQTG URGGEJ TGEQIPK\GTU UGNGEVKXGN[ WUG VJKU PCTTQY TGIKQP QH EGRUVTCN EQORQPGPVU D[ CRRN[KPI EGRUVTCN YGKIJVKPI QT NKVVGTKPI DCUGF QP C YKPFQYKPI HWPEVKQP QT C NKHVGT 6JG FGUKIP QH EGRUVTCN NKHVGT KU VQ EQPVTQN VJG PQPKPHQTOCVKQP DGCTKPI EGRUVTCN XCTKCDKNKVKGU KP QTFGT VQ RGTHQTO TGNKCDNG FKUETKOKPCVKQP QH UQWPFU 1PG RQRWNCT V[RG QH VJG NKHVGT HQT EGRUVTWO HGCVWTGU KU DCUGF QP C TCKUGF UKPG HWPEVKQP QH VJG HQTO
HQT HQT
YJGTG KU WUWCNN[ EJQUGP CU CPF KU V[RKECNN[ HQT URGGEJ QH M*\ DCPFYKFVJ 6JG YGKIJVGF HGCVWTG UGSWGPEG HTQO VJG EGRUVTCN NKHVGT EQTTG URQPFU VQ C UOQQVJGF NQI RQYGT URGEVTWO 6JG LWUVKſECVKQP QH VJKU V[RG QH NKHVGT HQT EGRUVTCN HGCVWTG XGEVQTU KU IKXGP KP =? +P FKUETKOKPCVKXG HGCVWTG GZVTCEVKQP NKHVGT FGUKIP ECP DG FQPG CEEQTFKPI VQ VJG EQODKPGF UVTKPI OQFGN HQTOWNCVKQP CPF KPUVGCF QH TGN[KPI QP JWOCP JGCTKPI ECRCDKNKVKGU HQT EJQQUKPI VJG TKIJV NKHVGT VJG NKHVGT RC TCOGVGTU ECP DG GUVKOCVGF WUKPI VJG /%' ETKVGTKQP VQ OKPKOK\G VJG TGEQIPKVKQP GTTQT TCVG 5RGGEJ GZRGTKOGPVU WUKPI FKUETKOKPCVKXG HGCVWTG GZVTCEVKQP DCUGF NKHVGT YGTG RGTHQTOGF KP UGXGTCN VCUMU = ?
1.7 Verification and Identification 5RGCMGT XGTKſECVKQP CPF KFGPVKſECVKQP DCUGF QP XQKEG KU CP KORQTVCPV CTGC KP URGGEJ TGUGCTEJ CPF JCU DGGP UVWFKGF HQT UGXGTCN FGECFGU 6JG IGPGTCN RTQDNGO QH RCVVGTP XGTKſECVKQP ECP DG HQTOWNCVGF CU HQNNQYU IKXGP C TCPFQO KPRWV UKIPCN YG YCPV VQ XGTKH[ KH VJG UKIPCN KU HTQO C UKIPCN UQWTEG ¼ +P OCMKPI C FGEKUKQP TGICTFKPI VJG QTKIKP QH VJG UKIPCN UQWTEG VYQ V[RGU QH GTTQTU ECP QEEWT 1PG EQWNF OKUVCMGPN[
E\&5&3UHVV//&
FGEKFG VJCV KU PQV HTQO VJG UKIPCN UQWTEG ¼ YJKNG VJG VTWG UQWTEG QH VJG UKIPCN KU ¼ 6JKU V[RG QH GTTQT KP XGTKſECVKQP KU TGHGTTGF VQ CU C V[RG + GTTQT VJG GTTQT QH HCNUG TGLGEVKQP QT OKUUGF FGVGEVKQP 6JG UGEQPF V[RG QH GTTQT KU VJCV KU CEEGRVGF CU EQOKPI HTQO VJG UKIPCN UQWTEG ¼ YJKNG VJG VTWG UQWTEG QH UKIPCN KU PQV ¼ 6JKU V[RG QH GTTQT KU QHVGP TGHGTTGF VQ CU C V[RG ++ GTTQT QT VJG GTTQT QH HCNUG CEEGRVCPEG 6JG RGTHQTOCPEG QH C XGTKſECVKQP U[UVGO KU V[RKECNN[ GXCNWCVGF DCUGF QP VJG EQODK PCVKQP QH V[RG + CPF V[RG ++ GTTQTU 6JG RTQDNGO QH XGTKſECVKQP ECP DG EQPXGPKGPVN[ HQTOWNCVGF KPVQ C UVCVKUVKECN J[RQVJGUKU VGUVKPI RTQDNGO IKXGP VJG VGUV UKIPCN YG YCPV VQ VGUV VJG PWNN J[RQVJGUKU ¼ CICKPUV VJG CNVGTPCVKXG J[RQVJGUKU ½ YJGTG KU HTQO VJG UQWTEG ¼ CPF ½ CUUWOGU VJCV KU IGPGTCVGF D[ ¼ CUUWOGU VJCV CPQVJGT UQWTEG ½ +P OCP[ CRRNKECVKQPU VJG CNVGTPCVKXG J[RQVJGUKU ½ CUUWOGU QPN[ KU PQV IGPGTCVGF HTQO VJG MPQYP UQWTEG ¼ CPF KP UWEJ UKVWCVKQPU ½ KU C VJCV EQORQUKVG J[RQVJGUKU CU QRRQUGF VQ DGKPI C UKORNG J[RQVJGUKU +P IGPGTCN C VGUV RTQEGFWTG FKXKFGU VJG UKIPCN URCEG KPVQ VYQ TGIKQPU CPF CPF YG TGLGEV ¼ KH CPF CEEGRV ¼ KH KU QHVGP TGHGTTGF VQ CU VJG ETKVKECN TGIKQP QH VJG VGUV 6JG RTQDCDKNKVKGU QH VJGUG VYQ V[RGU QH GTTQTU ECP DG GZRTGUUGF CU
½ ¼ CPF
¾ ½ ½
6JG RQYGT QH VJG VGUV YJKEJ KU CP KORQTVCPV SWCPVKV[ VQ EJCTCEVGTK\G VJG VGUV KU IKXGP D[ ½
+P UVCVKUVKECN J[RQVJGUKU VGUVKPI QPG KU QHVGP KPVGTGUVGF KP ſPFKPI VJG ETKVKECN TGIKQP
UWEJ VJCV VJG RQYGT QH VJG VGUV KU OCZKOK\GF QT KP QVJGT YQTFU VJG V[RG ++ GTTQT
KU OKPKOK\GF CV C IKXGP NGXGN QH V[RG + GTTQT # VGUV YJKEJ KU QRVKOCN KP VJKU UGPUG KU QHVGP TGHGTTGF VQ CU VJG OQUV RQYGTHWN VGUV 6JGTG CTG RNGPV[ QH UVWFKGU CXCKNCDNG KP VJG UVCVKUVKECN NKVGTCVWTG TGICTFKPI VJG FGUKIP QH VJG QRVKOCN VGUVU KH ¼ CPF ½ CTG MPQYP CPF HCNN KPVQ UQOG URGEKſE FKUVTKDWVKQPU UWEJ CU VJG GZRQPGPVKCN HCOKN[ =? +P RTCEVKEG VJG VGUV RTQEGFWTG KU QHVGP DCUGF QP C VGUV UVCVKUVKEU UWEJ VJCV ¼ KU TGLGEVGF KH #EEQTFKPI VQ VJG 0G[OCP2GCTUQP NGOOC ECP DG DCUGF QP RTQDCDKNKV[ TCVKQ VGUV ½ ¼ QT NKMGNKJQQF TCVKQ VGUV ½ ¼ CPF KU UGNGEVGF UWEJ VJCV VJG NGXGN QH VJG V[RG + GTTQT ¼ *QYGXGT KP OQUV RTCEVKECN XGTKſECVKQP RTQDNGOU YG JCXG PQ GZCEV MPQYNGFIG TGICTFKPI VJG FKUVTKDWVKQPU QH PWNN CPF CNVGTPCVKXG J[RQVJGUGU 6JKU RTQDNGO KU GXGP OQTG CEWVG HQT VJG URGGEJ UKIPCN YJKEJ KU PQPUVCVKQPCT[ CPF VJG GZCEV PCVWTG QH URGGEJ IGPGTCVKQP RTQEGUU KU UVKNN NCTIGN[ WPMPQYP /QTGQXGT RCTCOGVGTU QH VJG URGGEJ OQFGN CTG GUVKOCVGF HTQO XGT[ URCTUG FCVC RQKPVU EQNNGEVGF HTQO MPQYP UQWTEGU 9KVJ VJG EQTTGEVPGUU QH VJG OQFGN KP SWGUVKQP CPF VJG GUVKOCVKQP GTTQTU FWG VQ URCTUG VTCKPKPI UCORNGU HQT VJG RCTCOGVGTU QH VJG OQFGN VJG QRVKOCNKV[ QH VJG VGUV KP VJG ENCUUKECN UGPUG ECPPQV DG TGCNK\GF CPF FKUETKOKPCPV HWPEVKQP DCUGF OGVJQFU ECP DG WUGF VQ KORTQXG VJG
E\&5&3UHVV//&
5RGGEJ (GCVWTG
5RGCMGT6GUV5EQTG 'XCNWCVKQP
+FGPVKV[ %NCKO
#EEGRV4GLGEV
*[RQVJGUKU 6GUVKPI
5RGCMGT 6JGUJQNFU
5RGCMGT /QFGN
FIGURE 1.4 Block diagram of a speaker verification system XGTKſECVKQP RGTHQTOCPEG = ? +P VJG HQNNQYKPI UWDUGEVKQPU OQTG FGVCKNGF FKUEWUUKQPU CTG IKXGP VQ VJGUG CRRTQCEJGU
1.7.1 Speaker Verification and Identification #WVJGPVKECVKQP D[ XQKEG JCU XCTKQWU CRRNKECVKQPU KP JWOCPOCEJKPG EQOOWPKEC VKQP &GRGPFKPI QP CRRNKECVKQP TGSWKTGOGPVU KV ECP DG ENCUUKſGF KPVQ VYQ ECVGIQTKGU PCOGN[ URGCMGT XGTKſECVKQP CPF URGCMGT KFGPVKſECVKQP 5RGCMGT XGTKſECVKQP KPXQNXGU XGTKH[KPI VJG KFGPVKV[ QH C ENCKOGF URGCMGT HTQO C MPQYP URGCMGT RQRWNCVKQP CPF URGCMGT KFGPVKſECVKQP KPXQNXGU KFGPVKH[KPI CP WPMPQYP URGCMGT HTQO C MPQYP RQRW NCVKQP # V[RKECN URGCMGT XGTKſECVKQP U[UVGO KU UJQYP KP (KI )KXGP C UGSWGPEG CPF VJG ENCKOGF URGCMGT KFGPVKV[ VJG VGUV UVCVKUVKEU QH URGGEJ HGCVWTG XGEVQTU UEQTG ECP DG EQORWVGF HTQO VJG EQTTGURQPFKPI URGCMGT OQFGN 6JG VGUV UEQTG KU VJGP EQORCTGF YKVJ C VJTGUJQNF CUUQEKCVGF YKVJ VJG ENCKOGF URGCMGT VQ FGEKFG KH VJG ENCKOGF KFGPVKV[ UJQWNF DG CEEGRVGF QT TGLGEVGF 5RGCMGT OQFGNKPI KU VJG OQUV ETKVKECN RCTV QH C URGCMGT XGTKſECVKQP U[UVGO #NVJQWIJ VJG URGCMGTŏU URGGEJ RTQFWEVKQP RTQEGUU ECP DG OQFGNGF FKTGEVN[ DCUGF QP VJG RJ[U KQNQIKECN UVTWEVWTG QH VJG URGCMGTŏU CTVKEWNCVQT[ CRRCTCVWU UWEJ CU VJG UJCRG QH VJG XQECN VTCEV GVE KV KU SWKVG FKHſEWNV KP RTCEVKEG VQ WPKSWGN[ GZVTCEV UWEJ UVTWEVWTCN RC TCOGVGTU HTQO URGGEJ UCORNGU +PUVGCF KPFKTGEV URGCMGT OQFGNKPI KU QHVGP WUGF KP YJKEJ C UGV QH URGGEJ OQFGNU KU ETGCVGF HQT GCEJ URGCMGT DCUGF QP C EQNNGEVKQP QH URGCMGT URGEKſE URGGEJ VTCKPKPI FCVC 6JGUG URGGEJ OQFGNU EJCTCEVGTK\G VJG CEQWUVKE OCPKHGUVCVKQP QH URGGEJ HQT C IKXGP URGCMGT YJKEJ ECP DG FQPG DCUGF QP XCTKQWU ETKVGTKC CPF FGRGPF QP VJG V[RG QH VJG XGTKſECVKQP UVTCVGIKGU +P QTFGT VQ OQFGN VJG VGORQTCN UVTWEVWTG KP URGGEJ *//U CTG VJG OQUV RQRWNCT EJQKEG HQT URGCMGT OQF
E\&5&3UHVV//&
GNKPI CNVJQWIJ QVJGT OQFGNKPI VGEJPKSWGU UWEJ CU 83 EQFGDQQM PGWTCN PGVYQTMU GVE ECP CNUQ DG WUGF *//U CTG WUGF KP URGCMGT XGTKſECVKQP VQ OQFGN VJG IGP GTCN VTGPF QH VJG URGGEJ HTQO C MPQYP URGCMGT CPF FGRGPFKPI QP CRRNKECVKQPU CPF VJG CXCKNCDKNKV[ QH CPPQVCVGF VTCKPKPI FCVC KV ECP DG DCUGF QP ſZGF RJTCUGU DTQCF RJQPGVKE ENCUUGU YJQNG YQTFU QT GXGP UWDYQTFU +P URGCMGT XGTKſECVKQP VJG FGEKUKQP KU DCUGF QP VJG UEQTG QH VJG VGUV UVCVKUVKEU (QT URGCMGT KFGPVKſECVKQP RTQDNGO VJG URGCMGT KU KFGPVKſGF CU VJG VTWG URGCMGT KH
(QT URGCMGT XGTKſECVKQP YG WUWCNN[ CEEGRV VJG ENCKOGF URGCMGT KFGPVKV[ KH VJG VGUV UVCVKUVKEU HQT VJ URGCMGT
6JG VGUV UVCVKUVKEU ECP DG FKTGEVN[ DCUGF QP VJG NKMGNKJQQF UEQTG QH VJG URGGEJ KP VJ URGCMGTŏU OQFGN *QYGXGT VJG PQTOCNK\GF UEQTG HWPEVKQP YJKEJ KU C HQTO QH VJG IGPGTCNK\GF NKMGNKJQQF TCVKQ VGUV IKXGU C OWEJ DGVVGT URGCMGT XGTKſECVKQP RGTHQTOCPEG =? 6JG NKMGNKJQQF OQFGNU VJG CEQWUVKE URCEG YJKEJ FQGU PQV DGNQPI VQ URGCMGT CPF KU QHVGP QDVCKPGF HTQO VJG URGCMGT EQJQTV OQFGNKPI (QT GCEJ URGCMGT C UGV QH URGCMGTU YJQ CTG OQUV ENQUG VQ URGCMGT KU ECNNGF C EQJQTV UGV CPF ECP DG KFGPVKſGF HTQO VJG VTCKPKPI FCVC 6JG NKMGNKJQQF KU OQFGNGF CU C HWPEVKQP QH VJG NKMGNKJQQF HTQO EQORGVKPI KP VJG EQJQTV UGV )QQF URGCMGT XGTKſECVKQP RGTHQTOCPEG KU QDUGTXGF YJGP KU CRRTQZKOCVGF D[
¾
6JG FKUETKOKPCVKXG HWPEVKQP DCUGF /%' CRRTQCEJ ECP DG CRRNKGF WUKPI C UWKVCDNG OKUENCUUKſECVKQP OGCUWTG CEEQTFKPI VQ VJG VGUV UVCVKUVKEU =? (QT URGCMGT XGTKſEC VKQP VJG XGTKſECVKQP GTTQT ECP CNUQ DG EJCTCEVGTK\GF D[ C OKUXGTKſECVKQP OGCUWTG DCUGF QP VJG VGUV UVCVKUVKEU KP C YC[ UKOKNCT VQ VJG OKUENCUUKſECVKQP OGCUWTG KP TGEQIPKVKQP CPF VJG /%' DCUGF FKUETKOKPCPV HWPEVKQP CRRTQCEJ ECP DG FKTGEVN[ CFCRVGF VQ OKPKOK\G VJG XGTKſECVKQP GTTQT = ? +P HCEV VJG EQJQTV OQFGNKPI HQT OWNCVKQP JCU C ENQUG TGNCVKQP VQ VJG OKUENCUUKſECVKQP OGCUWTG KP /%' CRRTQCEJ 'Z RGTKOGPVCN TGUWNVU KPFKECVGF VJCV VJG FKUETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ VQ URGCMGT KFGPVKſECVKQP CPF XGTKſECVKQP ECP NGCF VQ C UKIPKſECPV TGFWEVKQP QH VJG QXGTCNN XGTKſ ECVKQP GTTQT 6JG URGCMGT VGUV UVCVKUVKEU CTG OWEJ DGVVGT UGRCTCVGF VJCP VJG /. DCUGF FKUVTKDWVKQP GUVKOCVKQP CRRTQCEJ =? 6JG U[UVGO UGPUKVKXKV[ VQ VJTGUJQNF UGNGEVKQP KU OWEJ TGFWEGF CPF TQDWUVPGUU QH VJG U[UVGO KU KORTQXGF = ? +P CFFKVKQP VQ CFFTGUUKPI VJG KFGPVKſECVKQP CPF XGTKſECVKQP RTQDNGO CU C URGEKCN ENCU UKſECVKQP RTQDNGO OGVJQFU QH KPVTQFWEKPI VJG UVTWEVWTG QH UVCVKUVKECN VGUV KP FKUETKOK PCPV HWPEVKQP DCUGF CRRTQCEJ CTG CNUQ CVVGORVGF /KPKOWO XGTKſECVKQP GTTQT /8' VTCKPKPI KU UWEJ CP CRRTQCEJ = ? 6JG OCKP FKHHGTGPEG VQ /%' CRRTQCEJ KU VJG WUG QH VYQ UGRCTCVG NQUU HWPEVKQPU VQ OQFGN VYQ V[RGU QH GTTQTU KP J[RQVJG UKU VGUVKPI 6JG FGVCKNU QH /8' CRRTQCEJ CTG FGUETKDGF DGNQY CPF KV GZGORNKſGU
E\&5&3UHVV//&
VJG FKUETKOKPCPV HWPEVKQP CRRTQCEJ VQ URGCMGT XGTKſECVKQP CPF KFGPVKſECVKQP 6JG OKUXGTKſECVKQP OGCUWTG CU QRRQUGF VQ VJG OKUENCUUKſECVKQP OGCUWTG HQT VJG VJ URGCMGT KU FGſPGF CU
KH HTQO VJG ENCKOGF VJ URGCMGT KH KU PQV HTQO VJG ENCKOGF VJ URGCMGT
YJGTG KU VJG NQINKMGNKJQQF UEQTG HTQO VJG ENCKOGF VJ URGCMGT OQFGN CPF KU VJG UEQTG HTQO VJG EQJQTV URGCMGT ITQWR 6JG OKUXGTKſECVKQP OGCUWTG KU GODGFFGF KP C UOQQVJ UKIOQKF DCUGF NQUU HWPEVKQP CU KP VJG /%' HQTOWNCVKQP 6YQ UGRCTCVG NQUU HWPEVKQPU CTG WUGF VQ FGUETKDG VJG V[RG + CPF V[RG ++ GTTQTU 6JG CXGTCIG NQUU HQT GCEJ V[RG QH GTTQTU CRRTQZKOCVGU VJG GORKTKECN V[RG + CPF V[RG ++ GTTQT TCVG QP VJG VTCKPKPI UCORNGU ½
CPF
½
½
¾
%NCKOGF 5RGCMGT
%NCKOGF 5RGCMGT
YJGTG KU VJG KPFKECVQT HWPEVKQP 6JG QXGTCNN GZRGEVGF NQUU QH VJG /8' KU IKXGP D[
YJGTG CPF CTG FGUKIP RCTCOGVGTU YJKEJ EQPVTQN VJG KPƀWGPEG QH V[RG + CPF V[RG ++ GTTQTU KP VJG QXGTCNN NQUU HWPEVKQP 6JG OQFGN RCTCOGVGT GUVKOCVKQP KP /8' VTCKP KPI KU VQ OKPKOK\G VJG GZRGEVGF NQUU QH 'S YJKEJ TGNCVGU VQ VJG OKPKOK\CVKQP QH GORKTKECN GTTQT TCVG QH V[RG + CPF V[RG ++ GTTQTU 6JG IQQFPGUU QH VJKU ETKVGTKQP KU LWUVKſGF HTQO VJG FKUETKOKPCPV HWPEVKQP CRRTQCEJ CPF KV KU OGCPKPIHWN GXGP YJGP VJG OQFGN EQTTGEVPGUU CUUWORVKQP ECPPQV DG GUVCDNKUJGF 8CTKQWU URGCMGT XGTKſECVKQP GZ RGTKOGPVU CTG EQPFWEVGF CPF VJG FKUETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ FGOQPUVTCVGU UKIPKſECPV RGTHQTOCPEG CFXCPVCIGU QXGT VJG FKUVTKDWVKQP GUVKOCVKQP DCUGF CRRTQCEJ = ?
1.7.2 Utterance Verification 7VVGTCPEG XGTKſECVKQP KU VQ XGTKH[ VJG EQPVGPV QH VJG URGGEJ WVVGTCPEG CICKPUV C ENCKOGF
QT J[RQVJGUK\GF VGZV UVTKPI +V ECP DG WUGF VQ XGTKH[ VJG URGGEJ TGEQIPKVKQP TGUWNV CPF FGEKFG YJGVJGT VJG FGEQFGF YQTF UVTKPI KU TGNKCDNG CPF UJQWNF DG CEEGRVGF YKVJ EQPſFGPEG 1PG PGY CRRTQCEJ KP WVVGTCPEG XGTKſECVKQP KU VQ XGTKH[ VJG KPHQTOCVKQP EQPVGPV QH VJG URGGEJ WVVGTCPEG CICKPUV UQOG MPQYP FCVC UVQTGF KP VJG WUGT RGTUQPCN RTQſNG UWEJ CU DKTVJ FCVG QT CICKPUV C FCVCDCUG VQ YJKEJ VJG KPHQTOCVKQP EQPVGPV QH VJG WVVGTCPEG RTQXKFGF D[ VJG WUGT YKVJ ENCKOGF KFGPVKV[ UJQWNF OCVEJ 6JKU CRRTQCEJ QH XGTKH[KPI VJG EQPVGPV QH VJG WVVGTCPEG CICKPUV C MPQYP FCVCDCUG KU ECNNGF XGTDCN KP HQTOCVKQP XGTKſECVKQP 8+8 =? 8GTDCN KPHQTOCVKQP XGTKſECVKQP ECP DG CEJKGXGF
E\&5&3UHVV//&
YKVJQWV VJG PGGF QH EQNNGEVKPI URGCMGT URGEKſE VTCKPKPI FCVC 7VVGTCPEG XGTKſECVKQP CU C UVCVKUVKECN J[RQVJGUKU VGUVKPI RTQDNGO JCU C ENQUG TGNCVKQP VQ VJG RTQDNGO QH URGCMGT XGTKſECVKQP 6JG FKUETKOKPCPV HWPEVKQP CRRTQCEJ DCUGF QP /%' QT /8' ECP DG CRRNKGF VQ WVVGTCPEG XGTKſECVKQP YKVJ VJG UCOG HCUJKQP CU KV KU CRRNKGF VQ URGCMGT XGT KſECVKQP *QYGXGT VJG RWTRQUG QH WVVGTCPEG XGTKſECVKQP KU VQ XGTKH[ VJG KPHQTOCVKQP EQPVGPV QH VJG WVVGTCPEG PQV VJG KFGPVKV[ QH VJG URGCMGT #U C EQPUGSWGPEG KPUVGCF QH WUKPI VJG URGCMGT OQFGN VJG EQPſFGPEG UEQTGU TGICTFKPI VJG YQTF EQPVGPV KP VJG WVVGTCPEG CTG WUGF KP WVVGTCPEG XGTKſECVKQP 6JGTG CTG OCP[ YC[U VQ HQTO YQTF CPF UVTKPI NGXGN EQPſFGPEG OGCUWTG HQT WVVGTCPEG XGTKſECVKQP +P VJG CRRTQCEJ FGUETKDGF KP =? VJG YQTF NGXGN EQPſFGPEG UEQTG KU DCUGF QP VJG YQTF NGXGN NKMGNKJQQF TCVKQ FGſPGF CU HQNNQYU
¼ ½
YJGTG ¼ CPF ½ CTG PWNN CPF CNVGTPCVKXG J[RQVJGUGU TGURGEVKXGN[ HQT XGT KH[KPI TCPFQO KPRWV JCU YQTF EQPVGPV 6JG NKMGNKJQQF QH ¼ KU HTQO VJG OQFGN HQT 6JG NKMGNKJQQF HQT VJG CNVGTPCVKXG J[RQVJGUKU ½ KU HTQO C FKHHGTGPV OQFGN YJKEJ KU OQFGNGF D[ VYQ UGRCTCVG *//U KU CP *// VTCKPGF QP CNN VJG VTCKPKPI FCVC KP VJG EQ JQTV UGV QH +V KU WUGF VQ OQFGN VJG EQORQUKVG CEQWUVKE URCEG EQPUKUVU QH QVJGT YQTFU GZEGRV CPF KV KU UQOGVKOGU ECNNGF CPVKOQFGN VQ YQTF KU CP QVJGT *// YJKEJ KU C ſNNGT OQFGN VQ OQFGN PQPMG[YQTF GXGPV 6JG NKMGNKJQQF QH ½ DCUGF QP VJGUG VYQ V[RGU QH *//U KU IKXGP CU
½ ½
YJGTG KU C RQUKVKXG EQPUVCPV +P FKUETKOKPCPV HWPEVKQP DCUGF WVVGTCPEG XGTKſECVKQP VJG OKUXGTKſECVKQP OGCUWTG KU
6JG OKUXGTKſECVKQP OGCUWTG KU GODGFFGF KP C UKIOQKF V[RG HWPEVKQP QH VJG HQTO
YJGTG KU C RQUKVKXG EQPUVCPV EQPVTQNNKPI VJG UNQRG QH VJG UKIOQKF HWPEVKQP CPF VCMGU QP VJG XCNWG QH CPF CU HQNNQYU
KH KH KH
YJGTG TGHGTU VQ VJG ECUGU VJCV KU EQTTGEVN[ TGEQIPK\GF /4 TGHGTU VQ VJG ECUGU VJCV KU OKUTGEQIPK\GF CPF TGHGTU VQ VJG ECUGU VJCV VJG KPRWV URGGEJ EQP VCKPU PQ MG[YQTF $CUGF QP 'S VJG FKUETKOKPCPV HWPEVKQP DCUGF /%' CPF /8' CRRTQCEJ ECP DG CRRNKGF CU KP VJG URGCMGT XGTKſECVKQP 6JG CDQXG CRRTQCEJ
E\&5&3UHVV//&
ECP DG GZVGPFGF VQ UVTKPI NGXGN XGTKſECVKQP CPF DCUGF QP UWDYQTF WPKVU GPCDNKPI VJG XGTKſECVKQP RTQEGUU XQECDWNCT[ KPFGRGPFGPV =? 5VTKPI NGXGN XGTKſECVKQP OCMGU VJG ſPCN TGLGEVKQPCEEGRVCPEG FGEKUKQP QP VJG MG[YQTF J[RQVJGUKU OCFG D[ VJG TGE QIPK\GT #UUWOKPI KPFGRGPFGPEG VJG UVTKPI NGXGN NKMGNKJQQF TCVKQ ECP DG YTKVVGP CU C RTQFWEV QH UWDYQTFNGXGN NKMGNKJQQF TCVKQ
½
$[ EQNNGEVKPI VGTOU QH VJG PWOGTCVQT CPF FGPQOKPCVQT VJG OKUXGTKſECVKQP OGCUWTG ECP DG IKXGP CU HQNNQYU
YJGTG ECP DG DCUGF QP VJG YQTFU QT UWDYQTFU CEEQTFKPI VQ VJG OQFGNKPI UVTCVGI[ WUGF KP CRRNKECVKQP 6JG FKUETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ ECP VJGP DG CRRNKGF VQ XGTKH[KPI YQTF UVTKPIU 6JG EQPUVTWEVKQP QH UVTKPI NGXGN EQPſFGPEG OGCUWTG KU CP CEVKXG TGUGCTEJ CTGC CPF XCTKQWU UVCVKUVKECN ETKVGTKC ECP CRRN[ 1PG ETKVGTKQP QHVGP WUGF KP UVTKPI NGXGN EQPſFGPEG OGCUWTG KU VJG OKPKOCZ RTKPEKRNG YJKEJ KU VQ OKPKOK\G VJG OCZKOWO TKUM QH CEEGRVKPI GCEJ KPFKXKFWCN ENCUU QT YQTF KP VJG UVTKPI 9QTF ENCUU FGRGPFGPV YGKIJVKPI ECP CNUQ DG WUGF YJKEJ IKXGU OQTG GORJCUKU VQ XGTKH[ UCNKGPV YQTFU CPF KPHQTOCVKQP DCTKPI ENCUUGU &KUEWUUKQPU QH HQTOKPI XCTKQWU V[RGU QH UVTKPI NGXGN EQPſFGPEG OGCUWTGU CTG IKXGP KP = ? CPF VJG TGHGTGPEGU EKVGF VJGTG 6JG KPVTQFWEVKQP QH XGTKſECVKQP RTQEGUU KP URGGEJ TGEQIPKVKQP QRGPU C RCTCFKIO QH CRRN[KPI VJG FKUETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ KP URGGEJ RTQEGUUKPI +PUVGCF QH WUKPI OQFGN NKMGNKJQQF UEQTGU CU KU V[RKECN QPG CRRTQCEJ RTQRQUGF KP =? KU VQ WUG IGPGTCNK\GF EQPſFGPEG UEQTG HQT FGEQFKPI 6JG IGPGTCNK\GF EQPſFGPEG UEQTGU KU HQTOGF D[ KPVGITCVKPI XCTKQWU MPQYNGFIG UQWTEGU 'ZCORNGU QH UWEJ MPQYNGFIG UQWTEGU CTG HTCOG CEQWUVKE NKMGNKJQQF HTCOG CEQWUVKE NKMGNKJQQF TCVKQ RJQPG CPF YQTF FWTCVKQP RGPCNVKGU YQTF NCPIWCIG RTQDCDKNKVKGU YQTF KPUGTVKQP RGPCNVKGU HTCOG GPGTI[ RGPCNVKGU RTQUQFKE EQPſFGPEG UEQTG GVE # EQPſFGPEG UEQTG RTGRTQEGUUQT KU WUGF VQ EQPXGTV VJG EQPſFGPEG UEQTG HTQO GCEJ EQORQPGPV KPVQ C UWKVCDNG HQTO HQT EQODKPCVKQP +P RCTVKEWNCT VJG PQPNKMGNKJQQF TCVKQ DCUGF MPQYNGFIG UQWTEG KU EQP XGTVGF KPVQ VJG NQICTKVJO FQOCKP CPF KPVGITCVGF WUKPI C NKPGCT EQODKPCVKQP 6JG NKMGNKJQQF TCVKQ DCUGF MPQYNGFIG UQWTEG KU ſTUV GODGFFGF KP C UKIOQKF HWPEVKQP VQ EQPVTQN KVU F[PCOKE TCPIG CPF VJGP KV KU NKPGCT EQODKPGF KP VJG NQICTKVJO FQOCKP YKVJ QVJGT EQORQPGPVU $QVJ FKUETKOKPCVKXG OQFGN EQODKPCVKQP CPF FKUETKOKPCVKXG WVVGT CPEG XGTKſECVKQP CTG CRRNKGF VQ KPVGITCVG EQPſFGPEG UEQTGU HTQO FKHHGTGPV MPQYNGFIG UQWTEGU CPF VQ GUVKOCVG OQFGN RCTCOGVGTU CPF EQODKPCVKQP YGKIJVU KP VJG IGPGTCN K\GF EQPſFGPEG UEQTG DGECWUG VJG IGPGTCNK\GF EQPſFGPEG UEQTG OC[ PQV DG C RTQD CDKNKV[ DCUGF NKMGNKJQQF QT C RTQDCDKNKV[ FKUVTKDWVKQP /QTGQXGT VJG EQPſFGPEG UEQTG EQORQPGPVU KP VJG IGPGTCNK\GF EQPſFGPEG UEQTG ECP DG DCUGF QP FKHHGTGPV NGXGN QH KPHQTOCVKQP UWEJ CU HTCOG NGXGN UVCVG NGXGN RJQPG NGXGN YQTF NGXGN GVE &WTKPI
E\&5&3UHVV//&
VJG FGEQFKPI RTQEGUU VJQUG EQPſFGPEG UEQTG EQORQPGPVU CTG CRRNKGF CV FKHHGTGPV VKOGU CPF CV FKHHGTGPV NC[GTU QH VJG FGEQFKPI PGVYQTM CEEQTFKPI VQ VJGKT URGEKſEC VKQPU 6JGTGHQTG VJG XGTKſECVKQP QH VJG FGEQFGF RCTVKCN YQTF UVTKPI ECP DG FQPG CV FKHHGTGPV RJQPG YQTF CPF RJTCUG ITQWR LWPEVKQPU CEEQTFKPI VQ VJG XGTKſECVKQP DCUGF NKMGNKJQQF TCVKQ UEQTG EQORQPGPVU KP VJG IGPGTCNK\GF EQPſFGPEG UEQTG 6JG FGEQF KPI RTQEGUU QH VJKU CRRTQCEJ ECP DG RGTHQTOGF KP C VYQ RCUU HCUJKQP DCUGF QP C YQTF ITCRJ QT VJG DGUV NKUV QDVCKPGF HTQO VJG ſTUV RCUU UGCTEJ QT KP C QPG RCUU UGCTEJ YJKEJ CRRNKGU VJG IGPGTCNK\GF EQPſFGPEG UEQTG KP UGCTEJ FKTGEVN[ = ? #PQVJGT CRRTQCEJ QH WVKNK\KPI XGTKſECVKQP KP URGGEJ TGEQIPKVKQP CPF WPFGTUVCPFKPI KU DCUGF QP C FGVGEVKQP CPF XGTKſECVKQP UVTCVGI[ =? +P VJKU CRRTQCEJ C MG[ RJTCUG DCUGF FGVGEVKQP WUKPI C IGPGTCN CEQWUVKE RJQPGVKE OQFGN KU RGTHQTOGF ſTUV 6JG FGVGEVGF MG[ RJTCUGU CTG RTQEGUUGF KP C XGTKſECVKQP OQFWNG VQ XGTKH[ VJG FGVGEVGF MG[ RJTCUGU CPF GNKOKPCVG HCNUG CNCTOU 'CEJ MG[ RJTCUG KU VCIIGF YKVJ C UGOCPVKE VCI CPF VJG XGTKſGF MG[ RJTCUGU CTG EQPPGEVGF KPVQ UGPVGPEG J[RQVJGUKU WUKPI VCUM URGEKſE UGOCPVKE MPQYNGFIG # UVCEM FGEQFGT KU VJGP WUGF VQ UGCTEJ HQT VJG QRVKOCN J[RQVJGUKU VJCV UCVKUH[ VJG UGOCPVKE EQPUVTCKPVU 6JG QRVKOCN J[RQVJGUKU HTQO VJG UVCEM FGEQFGT KU HWTVJGT XGTKſGF CV VJG UGPVGPEG NGXGN DCUGF QP DQVJ CEQWUVKE CPF UG OCPVKE KPHQTOCVKQP HQT VJG ſPCN QWVRWV 6JG FKUETKOKPCPV HWPEVKQP DCUGF CRRTQCEJ KU CRRNKGF KP DQVJ MG[ RJTCUG FGVGEVKQP CPF XGTKſECVKQP VQ KORTQXG VJG U[UVGO RGT HQTOCPEG &GVCKNU QH VJKU CRRTQCEJ CTG IKXGP KP = ? CPF VJG TGHGTGPEGU EKVGF VJGTG
1.8 Summary +P VJKU EJCRVGT YG GZCOKPGF VJG ENCUUKECN $C[GU FGEKUKQP VJGQT[ CRRTQCEJ VQ VJG RTQDNGO QH RCVVGTP ENCUUKſ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ſECVKQP CRRTQCEJ VQ ENCUUKſGT FGUKIP (QT C IKXGP UGV QH FKUETKOKPCPV HWPEVKQPU VJG ENCUUKſGT FGUKIP KU VQ ſPF C UGV QH RCTCOGVGTU YJKEJ OKPKOK\G VJG GORKTKECN TGEQIPKVKQP GTTQT TCVG 6JKU KU CEJKGXGF VJTQWIJ C URGEKCN NQUU HWPEVKQP YJGTG OKPKOK\KPI VJG GZRGEVGF NQUU TGNCVGU VQ VJG TGFWEVKQP QH VJG TGEQIPKVKQP GTTQT TCVG 6JG FKUETKOKPCPV HWPEVKQP DCUGF /%' CRRTQCEJ CRRNKGU VQ ECUGU YJGTG VJG VTCFKVKQPCN FKUVTKDWVKQP GUVKOCVKQP
E\&5&3UHVV//&
DCUGF CRRTQCEJ FQGU PQV CRRN[ GURGEKCNN[ YJGP VJG HCOKN[ QH VJG FKUETKOKPCPV HWPE VKQPU GPEQWPVGTGF KP VJG ENCUUKſGT CTG PQV DCUGF QP RTQDCDKNKV[ FKUVTKDWVKQPU 6JG IQQFPGUU QH VJKU CRRTQCEJ KU LWUVKſGF YKVJQWV VJG OQFGN EQTTGEVPGUU CUUWORVKQP CPF KV CRRNKGU VQ ECUGU YJGTG VJG OQFGN EQTTGEVPGUU CUUWORVKQP KU MPQYP VQ DG KPXCNKF 9G HQTOWNCVGF VJG DCUKE VJGQTGVKECN HTCOGYQTM QH VJKU CRRTQCEJ CPF FKUEWUUGF KVU TGNCVKQP VQ QVJGT ETKVGTKC WUGF KP VJG ENCUUKſGT FGUKIP 6JG FGXGNQROGPV QH /%' CR RTQCEJ JCU NGF VQ C PGY RCTCFKIO KP RCVVGTP TGEQIPKVKQP CPF KV NGCFU VQ TGEQIPKVKQP RGTHQTOCPEG CFXCPVCIGU QXGT VJG EQPXGPVKQPCN CRRTQCEJ KP OCP[ CRRNKECVKQPU 9G UVWFKGF XCTKQWU GZVGPUKQPU QH VJG /%' CRRTQCEJ CPF RTQXKFGF VJGQTGVKECN LWUVKſEC VKQPU CU YGNN CU KORNGOGPVCVKQP FGVCKNU YJGP KV YCU CRRNKGF VQ FKHHGTGPV ENCUUKſECVKQP RTQDNGOU KP URGGEJ CPF NCPIWCIG RTQEGUUKPI 6JKU EJCRVGT KU DCUGF QP PGY FGXGN QROGPVU KP FKUETKOKPCPV HWPEVKQP DCUGF /%' CRRTQCEJ FWTKPI VJG RCUV VGP [GCTU #NVJQWIJ CVVGORVU YGTG OCFG VQ RTQXKFG C UPCRUJQV QH TGUGCTEJ KP VJKU CTGC VJG OC VGTKCN EQXGTGF KP VJKU EJCRVGT KU D[ PQ OGCPU GZJCWUVKXG 4GUGCTEJ QP FKUETKOKPCVKXG OGVJQFU KP RCVVGTP ENCUUKſECVKQP KU C HCUV OQXKPI ſGNF YKVJ PGY RTQDNGOU CPF CRRNK ECVKQPU HTQO XCTKQWU FKTGEVKQPU CPF YG CTG LWUV CV VJG DGIKPPKPI QH TGCNK\KPI VJG PGY RQVGPVKCN QH VJKU CRRTQCEJ KP RCVVGTP TGEQIPKVKQP
Acknowledgement 6JG CWVJQT YQWNF NKMG VQ CEMPQYNGFIG VJG EQPVTKDWVKQPU QH JKU RCUV CPF RTGUGPV EQN NCDQTCVQTU /QUV OCVGTKCNU RTGUGPVGF KP VJKU EJCRVGT CTG DCUGF QP LQKPV RWDNKECVKQPU CPF FKUEWUUKQPU
References =? 5+ #OCTK ő# VJGQT[ QH CFCRVKXG RCVVGTP ENCUUKſGTUŒ IEEE Trans. on Electronic Computers 8QN 0Q RR Ō =? 5+ #OCTK ő.GCTPKPI RCVVGTPU CPF RCVVGTP UGSWGPEGU D[ UGNHQTICPK\KPI PGVU QH VJTGUJQNF GNGOGPVUŒ IEEE Transactions on Computers 8QN % 0Q RR Ō 0QXGODGT =? # )WPCYCTFCPC ő/CZKOWO OWVWCN KPHQTOCVKQP GUVKOCVKQP QH CEQWUVKE *// GOKUUKQP FGPUKVKGUŒ CLSP Research Note No. 40 %GPVGT HQT .CPIWCIG CPF 5RGGEJ 2TQEGUUKPI ,QJPU *QRMKPU 7PKXGTUKV[ ,WPG =? .4 $CJN 2( $TQYP 28 FG5QW\C CPF 4 . /GTEGT ő/CZKOWO OWVWCN KPHQTOCVKQP GUVKOCVKQP QH *// RCTCOGVGTU HQT URGGEJ TGEQIPKVKQPŒ Proceedings of ICASSP-86 RR Ō
E\&5&3UHVV//&
=? .4 $CJN 2( $TQYP 28 FG5QW\C CPF 4. /GTEGT ő'UVKOCVKPI JKFFGP /CTMQX OQFGN RCTCOGVGTU UQ CU VQ OCZKOK\G URGGEJ TGEQIPKVKQP CEEWTCE[Œ IEEE Trans. Speech and Audio Processing 8QN 0Q RR Ō =? .4 $CJN ( ,GNKPGM CPF 4. /GTEGT ő# OCZKOWO NKMGNKJQQF CRRTQCEJ VQ EQPVKPWQWU URGGEJ TGEQIPKVKQPŒ IEEE Transactions on Pattern and Machine Intelligence 8QN 2#/+ RR Ō =? .' $CWO 6 2GVTKG ) 5QWNGU CPF 0 9GKUU ő# OCZKOK\CVKQP VGEJPKSWG QE EWTTKPI KP VJG UVCVKUVKECN CPCN[UKU QH RTQDCDKNKUVKE HWPEVKQPU QH /CTMQX EJCKPUŒ Ann. Math. Stat. 8QN RR Ō =? .' $CWO ő#P KPGSWCNKV[ CPF CUUQEKCVGF OCZKOK\CVKQP VGEJPKSWGU KP UVCVKU VKECN GUVKOCVKQP HQT RTQDCDKNKUVKE HWPEVKQPU QH /CTMQX RTQEGUUŒ Inequalities 8QN RR Ō =? .' $CWO CPF ,# 'CIQP ő#P KPGSWCNKV[ YKVJ CRRNKECVKQPU VQ UVCVKUVKECN RTGF KECVKQP HQT HWPEVKQPU QH /CTMQX RTQEGUU CPF VQ OQFGN QH GEQNQI[Œ Bull. Amer. Math Soc., 8QN RR Ō =? .' $CWO CPF ) 5GNN ő)TQYVJ VTCPUHQTOCVKQPU HQT HWPEVKQPU QP OCPKHQNFUŒ Pacific J. Math. 8QN 0Q RR Ō =? $KEMGN CPF &QMUWO Mathematical Statistics 2TGPVKEG*CNN =? # $GPXGPKUVG / /GVKXKGT CPF 2 2TKQWGV Adaptive Algorithms and Stochastic Approximations 5RTKPIGT8GTNCI =? 2 $G[GTNGKP ő&KUETKOKPCVKXG OQFGN EQODKPCVKQPŒ Proc. 1997 Workshop on Automatic Speech Recognition and Understanding Proceedings RR Ō =? # $KGO 5 -CVCIKTK CPF $* ,WCPI ő2CVVGTP TGEQIPKVKQP WUKPI FKUETKOKPC VKXG HGCVWTG GZVTCEVKQPŒ IEEE Trans. Signal Processing 8QN RR Ō =? ,4 $NWO ő/WNVKFKOGPUKQPCN UVQEJCUVKE CRRTQZKOCVKQP OGVJQFUŒ #PP /CVJ 5VCV 8QN RR Ō =? 2% %JCPI CPF $* ,WCPI ő&KUETKOKPCVKXG VGORNCVG VTCKPKPI HQT F[PCOKE RTQITCOOKPI URGGEJ TGEQIPKVKQPŒ Proc. ICASSP92 8QN RR Ō =? 2% %JCPI CPF $* ,WCPI ő&KUETKOKPCVKXG VTCKPKPI HQT F[PCOKE RTQITCO OKPI DCUGF URGGEJ TGEQIPK\GTUŒ IEEE Trans. Speech and Audio Processing 5#2 =? 9 %JQW %* .GG CPF $ * ,WCPI ő5GIOGPVCN )2& VTCKPKPI QH CP JKFFGP /CTMQX OQFGN DCUGF URGGEJ TGEQIPK\GTŒ IEEE Proc. ICASSP-92 RR Ō #RTKN
E\&5&3UHVV//&
=? 9 %JQW %* .GG CPF $* ,WCPI ő/KPKOWO GTTQT TCVG VTCKPKPI DCUGF QP 0DGUV UVTKPI OQFGNUŒ IEEE Proc. ICASSP-93 8QN ++ RRŌ =? 9 %JQW %* .GG CPF $* ,WCPI ő/KPKOWO GTTQT TCVG VTCKPKPI QH KPVGT YQTF EQPVGZV FGRGPFGPV CEQWUVKE OQFGN WPKVU KP URGGEJ TGEQIPKVKQPŒ Proc. ICSLP-94 RR Ō ;QMQJCOC =? 9 %JQW $* ,WCPI CPF %* .GG ő/KPKOWO GTTQT TCVG VTCKPKPI QH EQO DKPGF UVTKPI OQFGNUŒ 75 2CVGPV =? 9 %JQW %* .GG CPF $* ,WCPI ő5RGGEJ TGEQIPKVKQP DCUGF QP EQODKPGF UVTKPI OQFGNUŒ Proc. DARPA ANN Tech. Program CSR Mtg. RR Ō =? 9 %JQW ő&KUETKOKPCPVHWPEVKQPDCUGF OKPKOWO TGEQIPKVKQP GTTQT TCVG RCVVGTPTGEQIPKVKQP CRRTQCEJ VQ URGGEJ TGEQIPKVKQPŒ Proceedings of The IEEE 8QN 0Q RR #WIWUV =? 9 %JQW %* .GG $* ,WCPI CPF ( - 5QQPI ő# OKPKOWO GTTQT TCVG RCVVGTP TGEQIPKVKQP CRRTQCEJ VQ URGGEJ TGEQIPKVKQPŒ International Journal of Pattern Recognition and Artificial Intelligence 8QN 0Q RR Ō =? 9 %JQW CPF $* ,WCPI “Adaptive discriminative learning in pattern recognition,” 6GEJPKECN 4GRQTV QH #66 $GNN .CDQTCVQTKGU =? 6 %QXGT CPF , 6JQOCU Elements of Information Theory ,QJP 9KNG[ 5QPU =? 5/ %JW CPF ; <JCQ ő4QDWUV URGGEJ TGEQIPKVKQP WUKPI FKUETKOKPCVKXG UVTGCO YGKIJVKPI CPF RCTCOGVGT KPVGTRQNCVKQPŒ IEEE Proc. ICSLP’98 RR Ō =? *CK &Q6W CPF /KEJCGN +PUVCNNG ő.GCTPKPI CNIQTKVJOU HQT PQPRCTCOGVTKE UQNW VKQP VQ OKPKOWO GTTQT ENCUUKſECVKQP RTQDNGOŒ IEEE Transactions on Computers 8QN % 0Q Ō =? #2 &GORUVGT 0/ .CKTF CPF &$ 4WDKP ő/CZKOWO NKMGNKJQQF HTQO KP EQORNGVG FCVC XKC VJG '/ CNIQTKVJOŒ J. Roy. Soc. 5GT $ 8QN RR =? ,. &QQD Stochastic Process ,QJP 9KNG[ 5QPU =? 4 1 &WFC CPF 2 ' *CTV Pattern Classification and Scene Analysis ,QJP 9KNG[ 5QPU =? ; 'RJTCKO CPF . 4CDKPGT ő1P VJG TGNCVKQP DGVYGGP OQFGNKPI CRRTQCEJGU HQT URGGEJ TGEQIPKVKQPŒ IEEE Transactions on Information Theory 8QN 0Q RR Ō /CTEJ =? ; 'RJTCKO # &GODQ CPF . 4CDKPGT ő# OKPKOWO FKUETKOKPCVKQP KPHQTOC VKQP CRRTQCEJ HQT JKFFGP /CTMQX OQFGNKPIŒ IEEE Transactions on Information Theory XQN 0Q RR Ō /CTEJ
E\&5&3UHVV//&
=? , (KUEWU ő# RQUVRTQEGUUKPI U[UVGO VQ [KGNF TGFWEGF YQTF GTTQT TCVGŒ Porc. 1997 IEEE Workshop on Automatic Speech Recognition and Understanding RR Ō =? /$ )CPFJK CPF , ,CEQD ő0CVWTCN PWODGT TGEQIPKVKQP WUKPI /%' VTCKPGF KPVGTYQTF EQPVGZV FGRGPFGPV CEQWUVKE OQFGNUŒ IEEE Proc. ICASSP’98 RR Ō =? % / FGN #NCOQ GV CN ő&KUETKOKPCVKXG VTCKPKPI QH )// HQT URGCMGT KFGPVKſ ECVKQPŒ IEEE Proc. ICASSP’96 RR Ō =? 25 )QRCNCMTKUJPCP GV CN ő#P KPGSWCNKV[ HQT TCVKQPCN HWPEVKQPU YKVJ CRRNK ECVKQPU VQ UQOG UVCVKUVKECN GUVKOCVKQP RTQDNGOUŒ IEEE Trans. on Information Theory 8QN PQ RR =? 25 )QRCNCMTKUJPCP GV CN ő&GEQFGT UGNGEVKQP DCUGF QP ETQUUGPVTQRKGUŒ IEEE Proc. ICASSP’88 RR Ō =? 5 *GTOCP CPF 4 5WMMCT ő,QKPV /%' GUVKOCVKQP QH 83 CPF *// RCTCOGVGTU HQT )CWUUKCP OKZVWTG UGNGEVKQPŒ IEEE Proc. ICASSP’98 RR Ō =? 3 *WQ CPF % %JCP ő6JG ITCFKGPV RTQLGEVKQP OGVJQF HQT VJG VTCKPKPI QH JKF FGP /CTMQX OQFGNUŒ Speech Communication 8QN RR Ō =? : *WCPI / $GNKP ( #NNGXC CPF / *YCPI ő7PKſGF UVQEJCUVKE GPIKPG
75' HQT URGGEJ TGEQIPKVKQPŒ IEEE Proc. ICASSP-93 RR Ō =? ( ,GNKPGM ő6JG FGXGNQROGPV QH CP GZRGTKOGPVCN FKUETGVG FKEVCVKQP TGEQIPK\GTŒ Proc. IEEE 8QN 0Q RR Ō 0QXGODGT =? ( ,GNKPGM ő%QPVKPWQWU URGGEJ TGEQIPKVKQP D[ UVCVKUVKECN OGVJQFUŒ Proc. of the IEEE 8QN 0Q RR =? ( ,GNKPGM 4 . /GTEGT CPF 5 4QWMQU ő2TKPEKRNGU QH NGZKECN .CPIWCIG OQF GNKPI HQT URGGEJ TGEQIPKVKQPŒ Advances in Speech Signal Processing 5 (WTWK CPF / / 5QPFJK GFU RR Ō /CTEGN &GMMGT 0GY ;QTM =? ( ,GNKPGM Statistical Methods for Speech Recognition 6JG /+6 2TGUU %CO DTKFIG =? $* ,WCPI .4 4CDKPGT CPF ,) 9KNRQP ő1P VJG WUG QH DCPFRCUU NKVVGTKPI KP URGGEJ TGEQIPKVKQPŒ IEEE Trans. Acoust. Speech Signal Processing #552 0Q Ō =? $* ,WCPI CPF . 4 4CDKPGT ő*KFFGP /CTMQX OQFGNU HQT URGGEJ TGEQIPK VKQPŒ Technometrics 8QN 0Q RR Ō #WIWUV =? $* ,WCPI 5 5 .GXKPUQP CPF / / 5QPFJK ő/CZKOWO NKMGNKJQQF GUVKOC VKQP HQT OWNVKXCTKCVG OKZVWTG QDUGTXCVKQPU QH /CTMQX EJCKPUŒ IEEE Trans. on Information Theory 8QN +6 0Q RR
E\&5&3UHVV//&
=? $* ,WCPI CPF . 4CDKPGT ő 6JG UGIOGPVCN -OGCPU CNIQTKVJO HQT GUVKOCV KPI RCTCOGVGTU QH JKFFGP /CTMQX OQFGNUŒ IEEE Trans. Acoust., Speech & Sig. Proc. RR 5GRVGODGT =? $* ,WCPI CPF 5 -CVCIKTK ő &KUETKOKPCVKXG NGCTPKPI HQT OKPKOWO GTTQT VTCKPKPIŒ IEEE Trans. Acoust., Speech & Sig. Proc. &G EGODGT =? $* ,WCPI 9 %JQW CPF %* .GG ő/KPKOWO ENCUUKſECVKQP GTTQT TCVG OGVJQFU HQT URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing /C[ =? & -CXGPUM[ ő)GPGTCNK\CVKQP QH $CWO CNIQTKVJO VQ HWPEVKQPU QP PQPNKPGCT /CPKHQNFUŒ Proc. ICASSP’95 8QN RR &GVTQKV =? / -CV\ ő'UVKOCVKQP QH RTQDCDKNKVKGU HTQO URCTUG FCVC HQT VJG NCPIWCIG OQFGN EQORQPGPV QH C URGGEJ TGEQIPK\GTŒ IEEE Trans. Acoustic., Speech, Signal Processing 8QN 0Q RR Ō =? 6 -CYCJCTC %* .GG CPF $* ,WCPI ő-G[RJTCUG FGVGEVKQP CPF XGTKſ ECVKQP HQT ƀGZKDNG URGGEJ WPFGTUVCPFKPIŒ IEEE Transactions on Audio and Speech Processing =? 6 -CYCJCTC %* .GG CPF $* ,WCPI ő%QODKPKPI MG[RJTCUG FGVGE VKQP CPF UWDYQTFDCUGF XGTKſECVKQP HQT ƀGZKDNG URGGEJ WPFGTUVCPFKPIŒ Proc. ICASSP’97 RR =? 6 -QOQTK CPF 5 -CVCIKTK ő#RRNKECVKQP QH C IGPGTCNK\GF RTQDCDKNKUVKE FG UEGPV OGVJQF QH F[PCOKE VKOG YCTRKPI DCUGF URGGEJ TGEQIPKVKQPŒ IEEE Proc. ICASSP-92 RR Ō =? 5 -CVCIKTK %* .GG $* ,WCPI CPF 6 -QOQTK ő0GY FKUETKOKPCVKXG VTCKP KPI CNIQTKVJOU DCUGF QP C IGPGTCNK\GF RTQDCDKNKUVKE FGUEGPV OGVJQFŒ Proc. IEEE-SP Workshop on Neural Networks for Signal Processing 2TKPEGVQP =? 5 -CVCIKTK $* ,WCPI CPF # $KGO ő&KUETKOKPCVKXG HGCVWTG GZVTCEVKQPŒ KP Artificial Neural Networks for Speech and Vision 4 /COOQPG 'F .QPFQP 7- %JCROCP CPF *CNN =? 5 -CVCIKTK $* ,WCPI CPF %* .GG ő2CVVGTP TGEQIPKVKQP WUKPI C HCOKN[ QH FGUKIP CNIQTKVJOU DCUGF WRQP VJG IGPGTCNK\GF RTQDCDKNKV[ FGUEGPV OGVJQFŒ IEEE Proceedings RR =? /9 -QQ %* .GG CPF $* ,WCPI ő5RGGEJ TGEQIPKVKQP CPF WVVGTCPEG XGTKſECVKQP DCUGF QP C IGPGTCNK\GF EQPſFGPEG UEQTGŒ IEEE Transactions on Speech and Audio Processing =? /9 -QQ %* .GG CPF $* ,WCPI ő# PGY FGEQFGT DCUGF QP C IGPGTCN K\GF EQPſFGPEG UEQTGŒ Proc. ICASSP’98 /C[
E\&5&3UHVV//&
=? ( -QTOCP\UMK[ CPF $* ,WCPI ő&KUETKOKPCVKXG #FCRVCVKQP HQT 5RGCMGT 8GT KſECVKQPŒ IEEE Proc. ICSLP’96 RR Ō =? -( .GG The Development of the SPHINX System -NWYGT =? # .LQNLG ; 'RJTCKO CPF . 4CDKPGT ő'UVKOCVKPI JKFFGP /CTMQX OQFGN RC TCOGVGTU D[ OKPKOK\KPI GORKTKECN GTTQT TCVGŒ Proc. ICASSP’90 RR =? '. .GJOCPP Testing Statistical Hypotheses 9KNG[ 0GY ;QTM =? ' .NGKFC CPF 4 4QUG ő'HſEKGPV FGEQFKPI CPF VTCKPKPI RTQEGFWTGU HQT WVVGT CPEG XGTKſECVKQP KP EQPVKPWQWU URGGEJ TGEQIPKVKQPŒ Proc. ICASSP’96 /C[ =? %* .GG $* ,WCPI 9 %JQW CPF ,, /QNKPC2GTG\ ő# UVWF[ QP VCUM KPFGRGPFGPV UWDYQTF UGNGEVKQP CPF OQFGNKPI HQT URGGEJ TGEQIPKVKQPŒ Proc. ICSLP96 RR 2JKNCFGNRJKC =? %* .GG ' )KCEJKP .4 4CDKPGT 4 2KGTCEEKPK CPF #' 4QUGPDGTI ő+O RTQXGF CEQWUVKE OQFGNKPI HQT URGCMGT KPFGRGPFGPV NCTIG XQECDWNCT[ EQPVKPW QWU URGGEJ TGEQIPKVKQPŒ Computer Speech and Language 8QN 0Q RR =? %* .GG ( - 5QQPI CPF 2CNKYCN 'FU ő#WVQOCVKE 5RGGEJ CPF 5RGCMGT 4GEQIPKVKQPŒ 0QTYGNN /# -NWYGT =? 3 .K $* ,WCPI 3 <JQW CPF %* .GG ő8GTDCN KPHQTOCVKQP XGTKſECVKQPŒ Proc. EuroSpeech’97 =? %* .GG ő# VWVQTKCN QP URGCMGT CPF URGGEJ XGTKſECVKQPŒ Proc. NORSIG’98 RR ,WPG =? 3 .K CPF $* ,WCPI ő5RGCMGT XGTKſECVKQP WUKPI XGTDCN KPHQTOCVKQP XGTKſ ECVKQP HQT CWVQOCVKE GPTQNNOGPVŒ Proc. ICASSP’98 /C[ =? %5 .KW *% 9CPI CPF %* .GG ő5RGCMGT XGTKſECVKQP WUKPI PQTOCNK\GF NQINKMGNKJQQF UEQTGŒ IEEE Trans. Audio & Speech Proc. ,CP =? %5 .KW %* .GG 9 %JQW $* ,WCPI CPF # 4QUGPDGTI ő# UVWF[ QP OKPKOWO GTTQT FKUETKOKPCVKXG VTCKPKPI HQT URGCMGT TGEQIPKVKQPŒ J. Acoust. Soc. Am. 8QN 0Q RR Ō =? /) 4CJKO CPF %* .GG ő5KOWNVCPGQWU #00 HGCVWTG CPF *// TGEQIPK\GT FGUKIP WUKPI UVTKPIDCUGF OKPKOWO ENCUUKſECVKQP VTCKPKPI QH *//UŒ Proc. ICSLP’96 RR Ō =? / ) 4CJKO %* .GG $* ,WCPI CPF 9 %JQW ő&KUETKOKPCVKXG WVVGTCPEG XGTKſECVKQP WUKPI OKPKOWO UVTKPI XGTKſECVKQP GTTQT /58' VTCKPKPIŒ Proc. ICASSP’96 RR Ō
E\&5&3UHVV//&
=? / ) 4CJKO CPF %* .GG ő5VTKPI DCUGF OKPKOWO XGTKſECVKQP GTTQT 5$ /8' VTCKPKPI HQT ƀGZKDNG URGGEJ TGEQIPKVKQPŒ Computer, Speech and Language =? / ) 4CJKO %* .GG CPF $* ,WCPI ő&KUETKOKPCVKXG WVVGTCPEG XGTKſEC VKQP HQT EQPPGEVGF FKIKV TGEQIPKVKQPŒ IEEE Transaction on Speech and Audio Processing =? 2 /E/CJQP 0 *CTVG 5 8CUGIJK CPF 2 /E%QWTV ő&KUETKOKPCVKXG 5RGEVTCN 6GORQTCN /WNVK4GUQNWVKQP (GCVWTGU HQT 5RGGEJ 4GEQIPKVKQPŒ IEEE Proc. ICASSP’99 RR Ō =? ' /E&GOQV CPF 5 -CVCIKTK ő2TQVQV[RGDCUGF /%')2& VTCKPKPI HQT XCTKQWU URGGEJ WPKVUŒ Comput. Speech Language 8QN RR Ō =? ' /E&GOQV CPF 5 -CVCIKTK ő5VTKPINGXGN /%' HQT EQPVKPWQWU RJQPGOG TGEQIPKVKQPŒ Proc. EuroSpeech’97 8QN RR Ō =? , / /GPFGN CPF - 5 (W Adaptive, Leaning and Pattern Recognition #EC FGOKE 2TGUU +PE =? # 0CFCU & 0CJCOQQ CPF / # 2KEJGP[ ő1P C OQFGNTQDWUV VTCKPKPI OGVJQF HQT URGGEJ TGEQIPKVKQPŒ IEEE Trans., on Acoustics, Speech and Signal Processing 8QN 0Q RR Ō =? ; 0QTOCPFKP GV CN ő*KIJ RGTHQTOCPEG EQPPGEVGF FKIKV TGEQIPKVKQP WUKPI OCZKOWO OWVWCN KPHQTOCVKQP GUVKOCVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN PQ RR =? - # 2CRKPGPK ő&KUETKOKPCVKXG VTCKPKPI XKC NKPGCT RTQITCOOKPIŒ Proc. ICASSP’99 =? - - 2CNKYCN / $CEEJKCPK CPF ; 5CIKUCMC ő/KPKOWO ENCUUKſECVKQP GTTQT VTCKPKPI CNIQTKVJO HQT HGCVWTG GZVTCEVKQP CPF RCVVGTP ENCUUKſGT KP URGGEJ TGEQI PKVKQPŒ Proc. EuroSpeech’95 RR Ō =? & 2QNNCTF Convergence of Stochastic Process 5RTKPIGT 5GTKGU KP 5VCVKUVKEU =? & 2QXG[ CPF 2 9QQFNCPF ő+ORTQXGF FKUETKOKPCVKXG VTCKPKPI VGEJPKSWGU HQT NCTIG XQECDWNCT[ EQPVKPWQWU URGGEJ TGEQIPKVKQPŒ 2TQE +''' +%#552 5CNV .CMG %KV[ /C[ =? .4 4CDKPGT ő# VWVQTKCN QP JKFFGP /CTMQX OQFGNU CPF UGNGEVGF CRRNKECVKQPU KP URGGEJ TGEQIPKVKQPŒ Proc. IEEE 8QN 0Q Ō (GDTWCT[ =? .4 4CDKPGT CPF $* ,WCPI Fundamentals of Speech Recognition 2TGPVKEG *CNN 'PINGYQQF %NKHHU 0, =? % 4CVJKPCXGNW CPF . &GPI ő7UG QH IGPGTCNK\GF F[PCOKE HGCVWTG RCTCOGVGTU HQT URGGEJ TGEQIPKVKQP OCZKOWO NKMGNKJQQF CPF OKPKOWO ENCUUKſECVKQP GTTQT CRRTQCEJGUŒ IEEE Proc. ICASSP’95 RR Ō
E\&5&3UHVV//&
=? % 4CVJKPCXGNW CPF . &GPI ő6JG VTGPF *// YKVJ FKUETKOKPCVKXG VTCKPKPI HQT RJQPGVKE ENCUUKſECVKQPŒ IEEE Proc. ICSLP’96 =? % 4CVJKPCXGNW CPF . &GPI ő*// DCUGF URGGEJ TGEQIPKVKQP WUKPI UVCVG FGRGPFGPV FKUETKOKPCVKXGN[ FGTKXGF VTCPUHQTOU QP OGNYCTRGF &(6 HGCVWTGUŒ IEEE Proc. ICASSP’96 RR Ō =? 9 4GKEJN CPF ) 4WUMG ő&KUETKOKPCVKXG VTCKPKPI HQT EQPVKPWQWU URGGEJ TGEQI PKVKQPŒ Proc. 1995 EuroSpeech’95 8QN RR /CFTKF 5GRV =? 9 4GKEJN ő.CPIWCIG OQFGN CFCRVCVKQP WUKPI OKPKOWO FKUETKOKPCVKQP KPHQT OCVKQPŒ Proc. EuroSpeech’99 RR $WFCRGUV =? 9 4GKEJN CPF 5 1TVOCPPU ő+PVGITCVGF PCVWTCN NCPIWCIG ECNN TQWVKPIŒ ESCA Workshop on Interactive Dialogue in Multi-Modal Systems -NQUVGT +TUGG =? 4# 4GPFGT CPF *( 9CNMGT ő/KZVWTG FGPUKVKGU OCZKOWO .KMGNKJQQF CPF VJG '/ CNIQTKVJOŒ SIAM Review 8QN 0Q RR Ō =? . 4KIC\KQ ,% ,WPSWC CPF / )CNNGT ő/WNVKNGXGN FKUETKOKPCVKXG VTCKPKPI HQT URGNNGF YQTF TGEQIPKVKQPŒ IEEE Proc. ICASSP’98 RR Ō =? #' 4QUGPDGTI GV CN ő6JG WUG QH EQJQTV PQTOCNK\GF UEQTGU HQT URGCMGT XGTK ſECVKQPŒ Proc. ICSLP’92 RR Ō =? * 4QDDKPU CPF 5 /QPTQ ő # 5VQEJCUVKE #RRTQZKOCVKQP /GVJQFŒ #PP /CVJ 5VCV 8QN RR Ō =? 4% 4QUG $* ,WCPI CPF %* .GG ő# VTCKPKPI RTQEGFWTG HQT XGTKH[KPI UVTKPI J[RQVJGUGU KP EQPVKPWQWU URGGEJ TGEQIPKVKQPŒ IEEE Proc. ICASSP’95 RR Ō =? / 5WIK[COC CPF - -WTKPCOK ő/KPKOWO ENCUUKſECVKQP GTTQT QRVKOK\CVKQP HQT C URGCMGT OCRRKPI PGWTCN PGVYQTMUŒ Neural Network for Signal Processing II RR =? 4 5EJNWVGT CPF 9 /CEJGTG[ ő%QORCTKUQP QH FKUETKOKPCVKXG VTCKPKPI ETKVGTKCŒ IEEE Proc. ICASSP’98 RR Ō =? (- 5QQPI CPF '( *WCPI ő# VTGGVTGNNKU DCUGF HCUV UGCTEJ HQT ſPFKPI VJG 0 DGUV UGPVGPEG J[RQVJGUGU KP EQPVKPWQWU URGGEJ TGEQIPKVKQPŒ Proc. ICASSP 91 =? 4 5EJNWVGT 9 /CEJGTG[ 5 -CPVJCM * 0G[ CPF . 9GNNKPI ő%QORCTK UQP QP QRVKOK\CVKQP OGVJQFU HQT FKUETKOKPCVKXG VTCKPKPI ETKVGTKCŒ IEEE Proc. EuroSpeech’97 RR Ō 5GRV =? 4 5EJNWVGT $ /WGNNGT ( 9GUUGN CPF * 0G[ ő+PVGTFGRGPFGPEG QH NCPIWCIG OQFGNU CPF FKUETKOKPCVKXG VTCKPKPIŒ Proc. ASRU’99 RR Ō &GE =? #4 5GVNWT 4# 5WMMCT CPF , ,CEQD ő%QTTGEVKQP TGEQIPKVKQP GTTQTU XKC FKU ETKOKPCVKXG WVVGTCPEG XGTKſECVKQPŒ IEEE Proc. ICSLP’96
E\&5&3UHVV//&
=? 4 5WMMCT CPF %* .GG ő8QECDWNCT[ KPFGRGPFGPV FKUETKOKPCVKXG WVVGTCPEG XGTKſECVKQP HQT PQPMG[YQTF TGLGEVKQP KP UWDYQTF DCUGF URGGEJ TGEQIPKVKQPŒ IEEE Transactions on Speech and Audio Processing 8QN RR Ō 0QX =? 4 5WMMCT ő5WDYQTF DCUGF OKPKOWO XGTKſECVKQP GTTQT 5$/8' VTCKPKPI HQT VCUM KPFGRGPFGPV WVVGTCPEG XGTKſECVKQPŒ IEEE Proc. ICASSP’98 RR Ō =? 4 5WMMCT / 4CJKO CPF %* .GG ő7VVGTCPEG XGTKſECVKQP QH MG[YQTF UVTKPIU WUKPI YQTF DCUGF OKPKOWO XGTKſECVKQP GTTQT 9$/8' VTCKPKPIŒ Proc. ICASSP’96 RR Ō /C[ =? 4 5WMMCT # 4 5GVNWT %* .GG CPF , ,CEQD ő8GTKH[KPI CPF EQTTGEVKPI UVTKPI J[RQVJGUGU WUKPI FKUETKOKPCVKXG WVVGTCPEG XGTKſECVKQPŒ Speech Communication =? ;C < 6U[RMKP ő5GNH.GCTPKPI Ō 9JCV KU +V!Œ IEEE Transactions on Automatic Control 8QN #% 0Q RR Ō &GEGODGT =? 8 9CTPMG 5 *CTDGEM ' 0QVJ * 0KGOCPP CPF / .GXKV ő&KUETKOKPCVKXG GUVKOCVKQP QH KPVGTRQNCVKQP RCTCOGVGTU HQT NCPIWCIG OQFGN ENCUUKſGTŒ IEEE Proc. ICASSP’99 RR Ō /CT =? 8 8CNVEJGX ,, 1FGNN 2% 9QQFNCPF CPF 5 , ;QWPI ő.CVVKEG DCUGF FKU ETKOKPCVKXG VTCKPKPI HQT NCTIG XQECDWNCT[ URGGEJ TGEQIPKVKQPŒ IEEE Proc. ICASSP’96 RR Ō /C[ =? 9W , CPF 3 *WQ ő5WRGTXKUGF CFCRVCVKQP QH /%'VTCKPGF %&*//U WUKPI OKPKOWO ENCUUKſECVKQP GTTQT NKPGCT TGITGUUKQPŒ Proc. ICASSP-2002 1TNCPFQ (NQTKFC /C[ =? % ;GP 55 -WQ CPF %* .GG ő/KPKOWO GTTQT TCVG VTCKPKPI HQT 2*// DCUGF VGZV TGEQIPKVKQPŒ IEEE Transactions on Image Processing 8QN 0Q RR Ō
E\&5&3UHVV//&
2 Minimum Bayes-Risk Methods in Automatic Speech Recognition Vaibhava Goel £ and William ByrneÝ £ IBM; Ý Johns Hopkins University
CONTENTS
/KPKOWO $C[GU4KUM %NCUUKſECVKQP (TCOGYQTM 2TCEVKECN /$4 2TQEGFWTGU HQT #54 5GIOGPVCN /$4 2TQEGFWTGU 'ZRGTKOGPVCN 4GUWNVU 5WOOCT[ #EMPQYNGFIGOGPVU 4GHGTGPEGU
#WVQOCVKE URGGEJ TGEQIPKVKQP #54 U[UVGOU CTG DGIKPPKPI VQ CRRGCT KP C YKFG XC TKGV[ QH KPHQTOCVKQP U[UVGOU +P CWVQOQDKNGU QT KP OKPKCVWTG EGNNWNCT RJQPGU #54 CNNQYU WUGTU VQ EQPVTQN GNGEVTQPKE FGXKEGU YKVJQWV WUKPI KPVTWUKXG MG[DQCTFU QT MG[ RCFU +P QVJGT CRRNKECVKQPU UWEJ CU KP UGCTEJKPI WPUVTWEVWTGF CWFKQXKUWCN CTEJKXGU #54 RTQOKUGU CEEGUU VQ KPHQTOCVKQP VJCV YQWNF QVJGTYKUG DG KPCEEGUUKDNG FWG VQ VJG FKHſEWNV[ QH UGCTEJKPI VJTQWIJ VJQWUCPFU QH JQWTU QH TGEQTFKPIU 9JGP #54 KU KPEQTRQTCVGF KPVQ CP KPHQTOCVKQP U[UVGO KV DGEQOGU LWUV QPG CURGEV QH C EQORNGZ CPF KPVGTTGNCVGF EQNNGEVKQP QH CWVQOCVKE RTQEGFWTGU 1XGTCNN U[UVGO RGTHQTOCPEG YKNN DG OGCUWTGF PQV D[ #54 YQTF GTTQT TCVG DWV VJTQWIJ VCUM URGEKſ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ſE ETKVGTKC *QYGXGT VJG OCZKOWO NKMGNKJQQF VGEJPKSWGU VJCV WPFGTNKG VJG VTCKPKPI CPF FGEKUKQP RTQEGUUGU QH OQUV EWTTGPV #54 U[U VGOU CTG PQV UGPUKVKXG VQ CRRNKECVKQP URGEKſE IQCNU # RTQOKUKPI CRRTQCEJ VQYCTFU VJG EQPUVTWEVKQP QH URGGEJ TGEQIPK\GTU VJCV CTG VWPGF HQT URGEKſE VCUMU KU MPQYP CU /KPKOWO $C[GUTKUM /$4 CWVQOCVKE URGGEJ TGEQIPKVKQP 6JG /$4 HTCOGYQTM CUUWOGU VJCV C SWCPVKVCVKXG OGCUWTG QH TGEQIPKVKQP RGTHQT OCPEG KU MPQYP CPF VJCV TGEQIPKVKQP UJQWNF DG C FGEKUKQP RTQEGUU VJCV CVVGORVU
E\&5&3UHVV//&
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ſE /$4 TGEQIPK\GTU YKNN DG EQORCTGF VQ OQTG EQOOQPN[ WUGF OCZKOWO NKMGNKJQQF TGEQIPKVKQP U[UVGOU VQ UJQY VJCV /$4 TGEQIPK\GTU ECP DG EQPUVTWEVGF VQ [KGNF KORTQXGF RGTHQTOCPEG WPFGT C XCTKGV[ QH VCUM URGEKſE GTTQT OGCUWTGU 9G YKNN VJGP FKUEWUU IGPGTCNK\CVKQPU QH /$4 YKVJ CP GORJCUKU QP VJG 418'4 U[UVGO EQODKPCVKQP RTQEGFWTG #U CP CRRNKECVKQP QH 418'4 CPF 418'4 XCTKCPVU YG YKNN RTGUGPV TGUWNVU KP U[UVGO EQODKPCVKQP HQT OWNVKNKPIWCN #54
2.1 Minimum Bayes-Risk Classification Framework +P #54 CP CEQWUVKE QDUGTXCVKQP UGSWGPEG ½ ¾ KU VQ DG OCRRGF VQ C YQTF UVTKPI ½ ¾ YJGTG VJG YQTFU DGNQPI VQ C XQECDWNCT[ #UUWOG VJCV C NCPIWCIG KU MPQYP HQT NCTIG XQECDWNCT[ VCUMU KV KU WUWCNN[ VJG UGV QH CNN YQTF UVTKPIU QXGT 6JKU NCPIWCIG URGEKſGU VJG YQTF UVTKPIU VJCV EQWNF RTQ FWEG CP[ CEQWUVKE FCVC UGGP D[ VJCV #54 U[UVGO (WTVJGTOQTG CUUWOG VJCV VJG #54 ENCUUKſGT OCMGU KVU J[RQVJGUKU UGNGEVKQP HTQO C UGV QH YQTF UVTKPIU 6JKU UGV ECNNGF VJG hypothesis space QH VJCV ENCUUKſGT YQWNF WUWCNN[ DG C UWDUGV QH VJG NCP IWCIG +P IGPGTCN VJG J[RQVJGUKU URCEG EQWNF GXGP DG C HWPEVKQP QH VJG QDUGTXCVKQP UC[ 6JG #54 ENCUUKſGT ECP VJGP DG FGUETKDGF CU Æ .GV ¼ DG C TGCN XCNWGF NQUU HWPEVKQP VJCV FGUETKDGU VJG EQUV KPEWTTGF YJGP CP WVVGTCPEG DGNQPIKPI VQ NCPIWCIG KU OKUVTCPUETKDGF CU ¼ ¼ EQWNF DG VJG YQTF GTTQT TCVG 9'4 OGCUWTGF D[ YGKIJVGF .GXGPUJVGKP FKUVCPEG = ? HQT C URGGEJ VTCPUETKRVKQP VCUM QT UQOG OGCUWTG QH UGOCPVKE FKUVCPEG DGVYGGP UGPVGPEGU HQT C URGGEJ WPFGTUVCPFKPI VCUM 5WRRQUG VJG VTWG FKUVTKDWVKQP QH URGGEJ CPF NCPIWCIG KU MPQYP VJKU CU UWOGU VJCV VJG VTWG FKUVTKDWVKQP VJCV FGUETKDGU FCVC GPEQWPVGTGF KP RTCEVKEG KU CXCKN CDNG +V YQWNF VJGP DG RQUUKDNG VQ OGCUWTG ENCUUKſGT RGTHQTOCPEG CEEQTFKPI VQ $C[GU TKUM CU
´µ Æ
6JKU KU VJG GZRGEVGF NQUU YJGP Æ KU WUGF CU VJG ENCUUKſECVKQP TWNG HQT FCVC IGPGT CVGF WPFGT )KXGP C NQUU HWPEVKQP CPF C FKUVTKDWVKQP VJG ENCUUKſECVKQP TWNG
VJCV OKPKOK\GU VJG $C[GUTKUM QH 'SWCVKQP KU IKXGP D[ =?
Æ
¼ ¾Ï ¾Ï
E\&5&3UHVV//&
¼
9JKNG VJG UWO KP 'SWCVKQP KU ECTTKGF QWV QXGT VJG GPVKTG NCPIWCIG QH VJG TGEQI PK\GT QPN[ VJQUG YQTF UVTKPIU YKVJ PQP\GTQ EQPFKVKQPCN RTQDCDKNKV[ EQP VTKDWVG VQ VJG UWO .GV FGPQVG VJG UWDUGV QH UWEJ VJCV
'SWCVKQP ECP PQY DG TGYTKVVGP CU
Æ
¼
9G UJCNN TGHGT VQ VJKU ENCUUKſGT CU VJG OKPKOWO $C[GUTKUM /$4 ENCUUKſGT +V OCMGU KVU J[RQVJGUKU UGNGEVKQP D[ ſTUV EQORWVKPI CP expected loss
HQT GCEJ YQTF UVTKPI KP VJG J[RQVJGUKU URCEG 6JG J[RQVJGUKU YKVJ VJG NGCUV GZRGEVGF NQUU KU VJGP UGNGEVGF CU KVU FGEKUKQP 5KPEG VJG QDUGTXCVKQPU KP UGTXG CU VJG GXKFGPEG WUGF D[ VJG /$4 ENCUUKſGT YG TGHGT VQ CU VJG evidence space HQT VJG CEQWUVKE QDUGTXCVKQPU 5KOKNCTN[ VJG FKUVTKDWVKQP VJCV FGſPGU VJG GXKFGPEG URCEG KU TGHGTTGF VQ CU VJG evidence distribution 9G PQY UJQY VJCV URGEKſE NQUU HWPEVKQPU ECP DG FGſPGF UQ VJCV VYQ EQOOQPN[ WUGF ENCUUKſECVKQP OGVJQFU PCOGN[ NKMGNKJQQF TCVKQ J[RQVJGUKU VGUVKPI CPF OCZKOWO C RQUVGTKQTK ENCUUKſECVKQP ECP DG FGTKXGF YKVJKP VJG /$4 HTCOGYQTM
2.1.1 Likelihood Ratio Based Hypothesis Testing +P J[RQVJGUKU VGUVKPI VJG QDUGTXCVKQP KU ENCUUKſGF CU DGNQPIKPI VQ QPG QH VYQ ENCUUGU C ŎPWNNŏ ENCUU VJCV TGRTGUGPVU C FGUKTGF UVCVGOGPV CDQWV CPF CP ŎCNVGTPCVKXGŏ ENCUU VJCV TGRTGUGPVU PGICVKQP QH VJG ŎPWNNŏ (QT KPUVCPEG KP C URGCMGT XGTKſECVKQP VCUM VJG PWNN ENCUU TGRTGUGPVU VJG FGUKTGF URGCMGT CPF VJG CNVGTPCVKXG TGRTGUGPVU KORQUVQTU 5KOKNCTN[ KP CP WVVGTCPEG XGTKſECVKQP VCUM VJG PWNN ENCUU KU VJG FGUKTGF WVVGTCPEG CPF VJG CNVGTPCVKXG KU C UGV QH UKOKNCT UQWPFKPI WVVGTCPEGU .GV FGPQVG VJG PWNN ENCUU CPF FGPQVG VJG CNVGTPCVKXG 6JG NKMGNKJQQF TCVKQ VGUVU
.46 HQT J[RQVJGUKU VGUVKPI ENCUUKſGU CEEQTFKPI VQ VJG HQNNQYKPI FGEKUKQP TWNG
Æ.46 KH QVJGTYKUG
6JG VJTGUJQNF KU UGV KP CP CRRNKECVKQP URGEKſE OCPPGT KV FGVGTOKPGU VJG DCNCPEG DGVYGGP HCNUG TGLGEVKQP CPF HCNUG CEEGRVCPEG 6JCV VJG .46 KU C URGEKCN ECUG QH /$4 ENCUUKſECVKQP ECP DG UGGP D[ EQPUKFGTKPI CP GXKFGPEG URCEG J[RQVJGUKU URCEG CPF NQUU
E\&5&3UHVV//&
HWPEVKQP .46
KH KH KH KH
7PFGT VJKU NQUU HWPEVKQP VJG GZRGEVGF NQUU 'SWCVKQP QH KU
CPF VJCV QH KU 6JGTGHQTG KU FGEKFGF QP KH QT 6JKU KU VJG FGEK UKQP TWNG QH 'SWCVKQP YKVJ
2.1.2 Maximum A-Posteriori Probability Classification 6JG /#2 ENCUUKſGT OCMGU KVU FGEKUKQP HTQO VJG GXKFGPEG URCEG KVUGNH D[ UGNGEVKPI VJG YQTF UVTKPI YKVJ VJG JKIJGUV EQPFKVKQPCN RTQDCDKNKV[ 6JCV KU Æ/#2
¾Ï
6JG /#2 ENCUUKſGT ECP DG FGTKXGF CU CP /$4 ENCUUKſGT D[ EQPUKFGTKPI C J[RQVJGUKU URCEG VJCV KU KFGPVKECN VQ VJG GXKFGPEG URCEG CPF C NQUU HWPEVKQP VJCV CUUKIPU GSWCN EQUV VQ CNN UC[ VQ CNN OKUENCUUKſECVKQPU 6JCV KU WPFGT VJG NQUU HWPEVKQP
¼
KH ¼
QVJGTYKUG
VJG ENCUUKſGT QH 'SWCVKQP DGEQOGU Æ
YJGTG
¼
¾Ï ¼
¼
¾Ï ¼½ ¼
6JKU KU VJG /#2 ENCUUKſGT QH 'SWCVKQP
2.1.3 Previous Studies of Application Sensitive ASR 4KUM OKPKOK\CVKQP CPF CRRNKECVKQP URGEKſE OKPKOWO EQUV ENCUUKſECVKQP JCXG DGGP YGNN UVWFKGF CPF RTCEVKEGF KP ſPCPEG FGHGPUG GEQPQOKEU CPF XCTKQWU QVJGT EQO OGTEKCN CPF PQPEQOOGTEKCN UGEVQTU *QYGXGT WUG QH VJGUG OGVJQFU KP CWVQOCVKE URGGEJ TGEQIPKVKQP JCU PQV DGGP GZVGPUKXG 'CTN[ KPXGUVKICVKQPU KPVQ VJG OKPKOWO $C[GUTKUM VTCKPKPI ETKVGTKC HQT URGGEJ TGEQIPK\GTU YGTG RGTHQTOGF D[ 0CFCU = ? 5KPEG VJGP QVJGT TGUGCTEJGTU = ? JCXG WUGF $C[GUTKUM DCUGF ETKVGTKC KP #54 U[U VGO VTCKPKPI 1WT HQEWU KP VJKU EJCRVGT JQYGXGT KU KP OKPKOWOTKUM ENCUUKſECVKQP TCVJGT VJCP GUVKOCVKQP
E\&5&3UHVV//&
5VQNEMG GVCN =? RTQRQUGF CP CRRTQZKOCVKQP VQ C OKPKOWO $C[GU TKUM ENCUUKſGT HQT IGPGTCVKQP QH OKPKOWO YQTF GTTQT TCVG J[RQVJGUKU HTQO TGEQIPKVKQP 0DGUV NKUVU 1VJGT TGUGCTEJGTU = ? JCXG RTQRQUGF RQUVGTKQT RTQDCDKNKV[ CPF EQPſFGPEG DCUGF J[RQVJGUKU UGNGEVKQP UVTCVGIKGU HQT YQTF GTTQT TCVG TGFWEVKQP VJCV JCXG DGGP UJQYP VQ DG CRRTQZKOCVKQPU VQ VJG /$4 ENCUUKſGTU = ? 6JGUG CRRTQZKOCVKQPU JCXG TGUWNVGF KP UKIPKſECPV KORTQXGOGPVU KP U[UVGO RGTHQTOCPEG CPF UWIIGUV VJCV HWTVJGT YQTM QP OKPKOWOTKUM ENCUUKſGTU HQT #54 OC[ DG DGPGſEKCN 9JKNG /$4 TGEQIPK\GTU CVVGORV VQ RTQXKFG C VCUM URGEKſE J[RQVJGUKU UGNGEVKQP OGEJ CPKUO RCTCNNGN GHHQTVU JCXG DGGP IQKPI QP FGXGNQRKPI VCUM URGEKſE TGEQIPKVKQP VGEJ PKSWGU D[ ETGCVKPI DGVVGT VCUM URGEKſE OQFGNU 0QVCDNG COQPI VJGUG CTG MG[YQTF URQVVKPI = ? RJTCUG FGVGEVKQP = ? YGKIJVGF YQTF GTTQT TCVG OKPKOK\C VKQP =? CPF KFGPVKſECVKQP QH PCOGF GPVKVKGU KP URGGEJ = ?
2.2 Practical MBR Procedures for ASR 6JG CNIQTKVJOKE KORNGOGPVCVKQP QH /$4 TGEQIPK\GTU KU FKHſEWNV HQT VJTGG TGCUQPU (KTUV FWG VQ VJG NCTIG XQECDWNCT[ UK\G KP OCP[ NCTIG XQECDWNCT[ EQPVKPWQWU URGGEJ TGEQIPKVKQP .8%54 VCUMU VJG GXKFGPEG CPF J[RQVJGUKU URCEGU KP 'SWCVKQP VGPF VQ DG SWKVG NCTIG GXGP HQT UJQTV CEQWUVKE QDUGTXCVKQP UGSWGPEGU (QT KPUVCPEG KH VJGTG CTG ſXG YQTFU KP VJG WVVGTCPEG CPF ¾¼ ¼¼¼ YQTFU KP VJG XQECDWNCT[ VJGTG CTG ¾¼ ¼¼¼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ſEKGPV F[PCOKE RTQITCOOKPI VGEJPKSWGU VQ KORNGOGPV VJG /#2 TGEQIPK\GT UWEJ OGVJQFU CTG PQV [GV CXCKNCDNG HQT CP /$4 TGEQIPK\GT WPFGT CP CTDKVTCT[ NQUU HWPEVKQP +P VJKU UGEVKQP YG RTGUGPV VYQ KORNGOGPVCVKQPU QH VJG /$4 TGEQIPK\GT ſ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
E\&5&3UHVV//&
2.2.1 Summation over Hidden State Sequences 9JGTGCU KP VJG FKUEWUUKQP VJWU HCT KV JCU DGGP CUUWOGF VJCV VJG VTWG GXKFGPEG FKUVTK DWVKQP KU CXCKNCDNG VJKU KU PQV VJG ECUG KP RTCEVKEG 6JKU FKUVTKDWVKQP KU QDVCKPGF D[ CRRN[KPI VJG $C[GU TWNG
*GTG KU CRRTQZKOCVGF WUKPI C language model KV KU WUWCNN[ C /CTMQX EJCKP DCUGF 0ITCO OQFGN KU WUWCNN[ CRRTQZKOCVGF WUKPI C JKFFGP /CTMQX OQFGN ECNNGF VJG acoustic model .GV DG VJG UGV QH CNN VJG UVCVGU KP VJG CEQWUVKE *// .GV FGPQVG VJG UGV QH CNN RQUUKDNG UVCVG UGSWGPEGU VJCV EQWNF IGPGTCVG 6JG RTQDCDKNKV[ KU EQORWVGF CU
6JG UWOOCVKQP QH 'SWCVKQP KU QXGT CNN RQUUKDNG JKFFGP UVCVG UGSWGPEGU 'XGP KH UGSWGPEGU HQT YJKEJ KU \GTQ CTG FKUECTFGF VJKU ECP UVKNN DG XGT[ GZRGPUKXG UKPEG VJG PWODGT QH FKUVKPEV JKFFGP UVCVG UGSWGPEGU ITQYU GZRQPGPVKCNN[ YKVJ VJG PWODGT QH HTCOGU KP # EQORWVCVKQPCNN[ HGCUKDNG CNVGTPCVKXG KU VQ OQFKH[ VJG 'SWCVKQP CU HQNNQYU
Æ
¼
¼
YJGTG KU C URCTUG UCORNKPI QH VJG OQUV NKMGN[ UVCVG UGSWGPEGU KP 6JKU TGCTTCPIGOGPV EJCPIGU DQVJ VJG GXKFGPEG CPF VJG J[RQVJGUKU URCEGU HTQO CPF VQ CPF TGURGEVKXGN[ +V CPVKEKRCVGU QWT UGCTEJ QXGT GXKFGPEG CPF J[RQVJGUKU URCEGU VJCV EQPVCKP YQTF UVTKPIU CNQPI YKVJ VJGKT *// UVCVG CNKIPOGPV KPHQTOCVKQP +P CFFKVKQP KV IKXGU WU VJG ƀGZKDKNKV[ QH YQTMKPI YKVJ NQUU HWPEVKQPU VJCV FGRGPF QP VJG UVCVG CNKIPOGPV QH YQTF UVTKPIU #NUQ KP VJG CDQXG YG WUGF $C[GU TWNG CPF KIPQTGF VJG VGTO YJKEJ KU EQPUVCPV HQT C IKXGP (QT EQPXGPKGPEG YG WUG TCVJGT VJCP TCVJGT VJCP CPF TCVJGT VJCP KP 'SWCVKQP YKVJ VJG WPFGTUVCPFKPI VJCV YQTF UGSWGPEGU KP J[RQVJGUKU CPF GXKFGPEG URCEGU EQPVCKP UVCVG CNKIPOGPV KPHQTOCVKQP YKVJ VJGO 9KVJ VJGUG EJCPIGU 'SWCVKQP DGEQOGU
Æ
¼
*GTG KU C UKPING WVVGTCPEG CPF CTG NGZKECNCEQWUVKE LQKPV RTQDCDKNKVKGU FGTKXGF YKVJ UVCVG CNKIPOGPV KPHQTOCVKQP HTQO CP N-best list QT C lattice
E\&5&3UHVV//&
2.2.2 MBR Recognition with N-best Lists #P 0DGUV NKUV KU C UQTVGF GPWOGTCVKQP QH YQTF UVTKPIU CPF VJGKT CUUQEKCVGF UVCVG CNKIP OGPV UQTVGF KP FGETGCUKPI QTFGT QH (QT GZCORNG CP 0DGUV NKUV IGPGTCVGF KP TGURQPUG VQ CP WVVGTCPEG EQTTGURQPFKPI VQ ő+ .+8' +0 # 474#. #4'#Œ CTG RTGUGPVGF KP 6CDNG 6JG OQUV FKTGEV CRRTQZKOCVKQP QH 'SWCVKQP KU D[ 0DGUV NKUV TGUEQTKPI RTQEG FWTGU CU ſTUV RTQRQUGF HQT 9'4 OKPKOK\CVKQP D[ 5VQNEMG GV CN =? CPF NCVGT GZ VGPFGF VQ IGPGTCN NQUU HWPEVKQPU D[ )QGN GV CN =? +P VJKU CRRTQCEJ VJG GXKFGPEG CPF J[RQVJGUKU URCEGU CTG TGUVTKEVGF VQ VJG 0DGUV NKUVU RTQFWEGF D[ C TGEQIPK\GT 6JG[ CTG FGPQVGF CPF TGURGEVKXGN[ TGUWNVKPI KP
Æ
¼ ¾Æ ¾Æ
¼
6JKU CRRTQZKOCVKQP KU RCTVKEWNCTN[ GCU[ VQ KORNGOGPV HQT CTDKVTCT[ NQUU HWPEVKQPU *QYGXGT VJG WUG QH 0DGUV NKUVU OC[ KP UQOG ECUGU DG VQQ TGUVTKEVKXG CP CRRTQZKOC VKQP CPF UGCTEJ GTTQTU OC[ TGUWNV 6JGTGHQTG KV KU QH KPVGTGUV VQ KPETGCUG VJG UK\G QH VJGUG VYQ URCEGU VQ VJG TGEQIPKVKQP NCVVKEG KG VQ EQPUKFGT OQTG ECPFKFCVGU KP VJG UGCTEJ CPF VJG UWO
2.2.3 MBR Recognition with Lattices +P VJG HQNNQYKPI YG RTGUGPV C OWNVKUVCEM RTGſZ VTGG £ UGCTEJ CNIQTKVJO VJCV WUGU TGEQIPKVKQP NCVVKEGU CU VJG J[RQVJGUKU CPF GXKFGPEG URCEGU 6JG FGXGNQROGPV QH VJG CNIQTKVJO RTQEGGFU CU HQNNQYU 9G UVCTV D[ KPVTQFWEKPI UVCVKUVKECN SWCPVKVKGU FGTKXGF HTQO VJG NCVVKEG VJCV CTG PGGFGF D[ VJG UGCTEJ RTQEGFWTG 9G VJGP RTGUGPV C UKPING UVCEM £ UGCTEJ FKTGEVN[ QXGT VJG NCVVKEGU 6JKU UGCTEJ KU HWTVJGT TGſPGF D[ KPVTQFWE KPI C RTGſZ VTGG OWNVKUVCEM UVTCVGI[ (QT ENCTKV[ QH RTGUGPVCVKQP YG HQTOWNCVG VJG £ UGCTEJ HQT OKPKOK\CVKQP QH 9'4 6JKU KU TGCNK\GF CU C OKPKOWOTKUM RTQEGFWTG WPFGT C NQUU HWPEVKQP DCUGF QP .GXGPUJVGKP FKUVCPEG JGPEGHQTVJ TGHGTTGF VQ CU VJG .GXGPUJVGKP NQUU HWPEVKQP 9G GPF VJKU UGEVKQP D[ C FKUEWUUKQP QH VJG HGCUKDKNKV[ QH VJG £ UGCTEJ HQT QVJGT NQUU HWPEVKQPU 2.2.3.1 Lattice Definitions # TGEQIPKVKQP NCVVKEG KU C EQORCEV TGRTGUGPVCVKQP HQT C NCTIG UGV QH YQTF UVTKPIU CPF VJGKT VKOG DQWPFCTKGU £ +V KU CP CE[ENKE FKTGEVGF ITCRJ KU VJG UGV QH PQFGU KU VJG UGV QH GFIGU KU VJG WPKSWG NCVVKEG UVCTV PQFG KU VJG WPKSWG NCVVKEG GPF PQFG CPF URGEKſGU NCVVKEG EQPPGEVKXKV[ 'CEJ PQFG KP VJG NCVVKEG KU NCDGNGF D[ C YQTF CPF C VKOG 'CEJ GFIG JCU C UVCTV PQFG CPF CP GPF PQFG 'FIGU CTG CUUQEKCVGF YKVJ VJG YQTFU CV VJGKT GPF PQFGU CPF YKVJ VJG
.CVVKEGU CTG IGPGTCVGF WUKPI YQTF UVTKPIU CPF VJGKT UVCVG NGXGN CNKIPOGPV YKVJ VJG CEQWUVKE HTCOGU *QY GXGT YG EQPUKFGT NCVVKEGU KP YJKEJ VJG UVCVG CNKIPOGPV KPHQTOCVKQP KU FKUECTFGF CPF QPN[ VJG YQTF VKOG DQWPFCTKGU CTG MGRV
E\&5&3UHVV//&
VKOG KPVGTXCN HTQO VJGKT UVCTV PQFG VQ VJGKT GPF PQFG 6JG[ CTG CNUQ NCDGNGF D[ VJG LQKPV CEQWUVKE CPF NCPIWCIG OQFGN NQIRTQDCDKNKV[ VJCV VJGKT YQTF QEEWTU FWTKPI VJG CUUQEKCVGF KPVGTXCN 6JKU LQKPV NQIRTQDCDKNKV[ KU EQPFKVKQPGF QP VJG JKUVQT[ URGEKſGF D[ VJG UVCTV PQFG QH VJG GFIG (QT GZCORNG KP (KIWTG CP GFIG KFGPVKſGU VJG J[RQVJGUKU VJCV VJG YQTF 019 DGIKPU CV UGE CPF GPFU CV UGE 6JG PWODGT QP VJKU GFIG KU NQIRTQDCDKNKV[ VJCV VJG YQTF 019 QEEWTU DGVYGGP UGE CPF UGE IKXGP VJCV VJG YQTF *'..1 KU RTGUGPV HTQO VJG UVCTV QH VJG CEQWUVKE FCVC WPVKN UGE # path QT complete path KU C UGSWGPEG QH EQPPGEVGF PQFGU CPF NKPMU HTQO VQ VJTQWIJ VJG NCVVKEG # path segment KU C UGSWGPEG QH EQPPGEVGF PQFGU HTQO CP KPVGTPCN NCVVKEG PQFG VQ CPQVJGT KPVGTPCN NCVVKEG PQFG OC[ DG CPF OC[ DG # partial path KU C UGSWGPEG QH EQPPGEVGF PQFGU CPF NKPMU HTQO VQ CP KPVGTPCN NCVVKEG PQFG KV OC[ DG C EQORNGVG RCVJ KH KU 6JG CEQWUVKE UGIOGPV EQTTGURQPFKPI VQ C RCVJ UGIOGPV HTQO VQ UJCNN DG FGPQVGF .GV DG C RCTVKCN RCVJ HTQO NCVVKEG UVCTV VQ DG C RCVJ UGIOGPV HTQO VQ CPF DG C RCVJ UGIOGPV HTQO VQ VJG NCVVKEG GPF 6JG CEQWUVKE UGIOGPVU EQTTGURQPFKPI VQ VJGUG VJTGG RCVJ UGIOGPVU YKNN DG FGPQVGF CPF TGURGEVKXGN[ 6JG UWO QH NQIRTQDCDKNKVKGU QP VJG GFIGU CNQPI IKXGU VJG UWO QH NQIRTQDCDKNKVKGU CNQPI IKXGU CPF VJG UWO QH NQIRTQDCDKNKVKGU CNQPI IKXGU 9G KPVTQFWEG VJG partial path log-probability VJG lattice backward log-probability CPF VJG lattice total probability QH C RCTVKCN J[RQVJGUKU CU HQNNQYU 6JG RCTVKCN RCVJ NQIRTQDCDKNKV[ QH KU
6JG NCVVKEG DCEMYCTF NQIRTQDCDKNKV[ QH KU
¡ ¾Ï
YJGTG FGPQVGU VJG UGV QH CNN EQORNGVG RCVJU KP VJG NCVVKEG 6JG NCVVKEG VQVCN RTQDCDKNKV[ QH KU
5WDUVKVWVKPI VJG FGſPKVKQPU QH CPF KP 'SWCVKQP YG IGV
¡
¡
E\&5&3UHVV//&
¾Ï
¾Ï
9'..
PU
*19
*'..1
;17
P
*'..1
PG
#4'
019
VKOG UGE FIGURE 2.1 An example lattice. The time marks correspond to the node times and the word ending times. The numbers on the edges are logarithms of conditional joint probabilities as described in the text. The partial path log-probability of a partial hypothesis is the log of the probability of its path; the partial path (‘HELLO’,‘0.6’) in this lattice has value . The lattice backward log-probability of a partial hypothesis is the log of the sum of probabilities of all lattice paths from end node of to the lattice end node; for the partial path (‘HELLO’,‘0.6’) in this lattice these paths are indicated by dotted lines and the lattice backward log-probability of this is . The lattice total probability of a partial path is the exponentiated sum of its partial path log-probability and lattice backward log-probability; its value is for (‘HELLO’, ‘0.6’) in the lattice above.
¡ ¾Ï
¡
¾Ï
6JG NCVVKEG VQVCN RTQDCDKNKV[ QH C RCTVKCN J[RQVJGUKU EQWNF VJGTGHQTG DG KPVGTRTGVGF CU VJG LQKPV RTQDCDKNKV[ QH QDUGTXKPI VJG CEQWUVKEU CPF CNN RQUUKDNG EQORNGVG J[RQVJGUGU VJCV JCXG VJG RTGſZ 6JGUG RTQDCDKNKVKGU CTG KNNWUVTCVGF KP (KIWTG
E\&5&3UHVV//&
2.2.3.2
£ Search Under General Loss Functions
6JG UGV QH CNN EQORNGVG RCVJU KP VJG NCVVKEG EQPUVKVWVGU VJG J[RQVJGUKU URCEG QH 'SWCVKQP +V CNUQ HQTOU VJG GXKFGPEG URCEG 'SWCVKQP VJG CUUQEKCVGF LQKPV NQIRTQDCDKNKV[ ECP DG EQORWVGF D[ CFFKPI VJG NQI RTQDCDKNKVKGU QP NCVVKEG GFIGU CNQPI 6JGTGHQTG QP VJG NCVVKEG YG YQWNF KORNGOGPV
Æ
¼
¼
¾Ï ¾Ï
6JG IQCN KU VQ ſPF C EQORNGVG J[RQVJGUKU ¼ KG C RCVJ HTQO VQ VJTQWIJ VJG NCVVKEG UWEJ VJCV KVU GZRGEVGF NQUU
¼
¼
¾Ï
KU VJG NGCUV QH CNN EQORNGVG J[RQVJGUGU KP VJG NCVVKEG 6JKU UGCTEJ HQT ¼ ECP DG GHHGE VKXGN[ KORNGOGPVGF CU CP £ CNIQTKVJO = ? YJKEJ RTQEGGFU D[ GZVGPFKPI RCTVKCN J[RQVJGUGU HQTYCTF VJTQWIJ VJG NCVVKEG 6YQ EQUV HWPEVKQPU CTG TGSWKTGF HQT VJG UGCTEJ 6JG ſTUV EQUV HWPEVKQP KU CUUQEKCVGF YKVJ GCEJ J[RQVJGUKU YJGVJGT RCTVKCN QT EQORNGVG +VU XCNWG KU C NQYGT DQWPF QP VJG GZRGEVGF NQUU 'SWCVKQP VJCV ECP DG QDVCKPGF D[ GZVGPFKPI VJG J[RQVJGUKU VJTQWIJ VJG NCVVKEG VQ EQORNGVKQP
¡
¾Ï
¾Ï
6JG UGEQPF EQUV HWPEVKQP KU QPN[ CUUQEKCVGF YKVJ EQORNGVG J[RQVJGUGU +V KU CP QXGT GUVKOCVG QH VJG GZRGEVGF NQUU QH C EQORNGVG J[RQVJGUKU ¼
¼
¾Ï
¼
*[RQVJGUGU CTG MGRV KP C RTKQTKV[ SWGWG YJKEJ KU UQTVGF D[ EQUV YKVJ VJG UOCNNGUV EQUV J[RQVJGUKU CV VJG VQR 9G UJCNN WUG VJG VGTO őUVCEMŒ VQ TGHGT VQ VJG SWGWG UKPEG KP URGGEJ TGEQIPKVKQP VJG £ CNIQTKVJOU JCXG JKUVQTKECNN[ DGGP RTGUGPVGF KP VGTOU QH UVCEMU = ? #V GXGT[ KVGTCVKQP VJG J[RQVJGUKU CV VJG VQR QH VJG UVCEM KU GZVGPFGF 9JGP VJGTG KU C EQORNGVG J[RQVJGUKU CV VJG VQR KVU UGEQPF EQUV KU EQORWVGF +H VJKU QXGTGUVKOCVGF EQUV KU UOCNNGT VJCP VJG WPFGT GUVKOCVGF EQUV QH VJG PGZV UVCEM J[ RQVJGUKU QT KH VJGTG KU PQ RCTVKCN J[RQVJGUKU NGHV KP VJG UVCEM VJG CNIQTKVJO VGTOKPCVGU 9G PQVG VJCV £ RTQEGFWTGU WUWCNN[ GORNQ[ CP GZCEV GZRGEVGF NQUU 'SWCVKQP HQT EQORNGVG J[RQVJGUGU JQYGXGT VJKU KU RTQJKDKVKXGN[ GZRGPUKXG VQ ſPF KP QWT ECUG VJGTGHQTG YG WUG VJG QXGTGUVKOCVG 2.2.3.3 Single Stack Search Under Levenshtein Loss Function 9G PQY RTGUGPV WUCDNG EQUV HWPEVKQPU HQT VJG .GXGPUJVGKP FKUVCPEG ¼ 6JGUG EQUVU CTG PQV WPKSWG CPF VJG GHſEKGPE[ QH VJG UGCTEJ FGRGPFU QP VJG SWCNKV[ QH DQVJ VJG WPFGTGUVKOCVG CPF VJG QXGTGUVKOCVG
E\&5&3UHVV//&
#U C VGEJPKECN CUKFG YG PQVG VJCV VJG .GXGPUJVGKP NQUU HWPEVKQP KU PQV UGPUKVKXG VQ VJG YQTF VKOG DQWPFCTKGU RTGUGPV KP VJG NCVVKEG 6JGTGHQTG VJG YQTF VKOG DQWPFCTKGU YQWNF DG UWOOGF QXGT FWTKPI VJG UGCTEJ 6JWU VJKU £ UGCTEJ KORNKEKVN[ RTQXKFGU OCTIKPCNK\CVKQP QXGT FKHHGTGPV VKOG UGIOGPVU QH YQTF UVTKPIU RTGUGPV KP VJG NCVVKEG .GV FGPQVG VJG UGV QH CNN EQORNGVG CPF RCTVKCN J[RQVJGUGU KP VJG UVCEM 6JG WPFGTGUVKOCVG HQT RCTVKCN J[RQVJGUGU KU
Ï
¾Ï
¡ ¡ ¡ ¾Ï
Ï
¡ ¾Ï
YJGTG KU VJG UGV QH CNN RQUUKDNG YQTF UVTKPIU CPF VJGKT CNN RQUUKDNG VKOG DQWPFCTKGU VJCV ECP DG EQPUVTWEVGF D[ EQPECVGPCVKPI \GTQ QT OQTG YQTFU QH VJG XQECDWNCT[ 6JG FGTKXCVKQP UJQYKPI VJCV VJKU EQUV HWPEVKQP UCVKUſGU 'SWCVKQP KU RTGUGPVGF KP )QGN GVCN =? 6JG QXGT GUVKOCVG HQT C EQORNGVG J[RQVJGUKU ¼ ECP DG EQORWVGF CU HQNNQYU
¯ (QT C J[RQVJGUKU KP UVCEM Ï
NGV DG VJG NGPIVJ QH VJG NQPIGUV RCVJ HTQO KVU GPF PQFG VQ VJG NCVVKEG GPF PQFG
¯ #RRGPF GCEJ J[RQVJGUKU KP VJG UVCEM D[
KPUVCPEGU QH QWV QH XQECDW NCT[ OCTMGTU 6JGUG OCTMGTU FQ PQV OCVEJ CP[ YQTF KP VJG XQECDWNCT[
¯ %QORWVG VJG QXGTGUVKOCVG ¼
¾Ï
¡ ¼
# FGTKXCVKQP UJQYKPI VJCV VJKU GUVKOCVG UCVKUſGU 'SWCVKQP KU IKXGP KP )QGN GV CN =? 9KVJ VJG WPFGTGUVKOCVG 'SWCVKQP CPF VJG QXGTGUVKOCVG 'SWCVKQP VJG HQNNQYKPI UKPING UVCEM UGCTEJ CNIQTKVJO ECP DG WUGF VQ ſPF VJG FGUKTGF J[RQVJGUKU KP VJG TGEQIPKVKQP NCVVKEG /CTM VJG NCVVKEG PQFGU D[ VJG NCVVKEG DCEMYCTF NQIRTQDCDKNKV[ 'SWC VKQP #V GCEJ PQFG MGGR VJG NGPIVJ QH VJG NQPIGUV RCVJ VQ VJG GPF QH VJG NCVVKEG
Ï
/CKPVCKP C UVCEM QH RCTVKCN CPF EQORNGVG J[RQVJGUGU 'CEJ RCTVKCN UVCEM GPVT[ EQPVCKPU C J[RQVJGUKU 'SWCVKQP
'SWCVKQP CPF 'SWCVKQP 'CEJ EQORNGVG UVCEM GPVT[ EQPVCKPU C J[RQVJGUKU ¼ ¼ ¼ CPF ¼ 'SWC VKQP 6JG UVCEM QTFGTKPI KU FGſPGF ſTUV D[ KPETGCUKPI XCNWGU QH CPF UGEQPF D[ FGETGCUKPI XCNWGU QH KP ECUGU QH KFGPVKECN
¡ ¡
E\&5&3UHVV//&
¡
+PKVKCNK\G VJG UGCTEJ D[ KPUGTVKPI VJG UVCTV PQFG QH VJG NCVVKEG KG VJG 07.. J[RQVJGUKU KPVQ VJG UVCEM +H VJGTG CTG KPEQORNGVG J[RQVJGUGU KP VJG UVCEM GZVGPF VJG VQR KPEQO RNGVG J[RQVJGUKU D[ CNN NCVVKEG CTEU VJCV NGCXG KVU GPF PQFG %QORWVG HQT GCEJ QH VJG PGYN[ ETGCVGF RCTVKCN J[RQVJGUKU %QORWVG ¼ CPF ¼ HQT GCEJ PGYN[ ETGCVGF EQORNGVG J[RQVJGUKU ¼ 1VJGTYKUG KH VJGTG CTG PQ KPEQORNGVG UVCEM J[RQVJGUGU UGNGEV VJG J[ RQVJGUKU YKVJ NGCUV ¼ 6JKU KU VJG FGUKTGF ECPFKFCVG 7RFCVG VJG EQUV GUVKOCVGU 'SWCVKQPU CPF QH CNN QVJGT RCT VKCN CPF EQORNGVG UVCEM J[RQVJGUGU CHVGT CFFKPI VJGUG PGYN[ ETGCVGF J[RQVJGUGU VQ VJG GXKFGPEG URCEG +PUGTV VJG PGYN[ ETGCVGF J[RQVJG UGU CV VJGKT CRRTQRTKCVG RNCEGU UQTVGF ſTUV D[ ¡ CPF UGEQPF D[ ¡ KP ECUG QH VKGU KP VJG UVCEM 2TWPKPI OC[ DG CRRNKGF FWTKPI VJG KPUGTVKQP
UGG 5GEVKQP +H VJGTG KU C EQORNGVG J[RQVJGUKU CV VJG VQR QH VJG UVCEM CPF KH KVU QXGT GUVKOCVG KU UOCNNGT VJCP VJG WPFGTGUVKOCVG QH UGEQPF UVCEM J[RQVJGUKU
RCTVKCN QT EQORNGVG KV KU VJG FGUKTGF ECPFKFCVG CPF VJG UGCTEJ GPFU 1VJGTYKUG IQ VQ UVGR 2.2.3.4 Prefix Tree Search Under Levenshtein Loss Function +P QWT VTGCVOGPV UQ HCT VJG VKOG UGIOGPVCVKQP QH GCEJ J[RQVJGUKU KU TGVCKPGF UQ VJCV J[RQVJGUGU CTG FKUVKPEV KH VJG[ JCXG KFGPVKECN YQTF EQPVGPV DWV FKHHGTGPV VKOG UGIOGP VCVKQP 5KPEG VJG .GXGPUJVGKP FKUVCPEG FQGU PQV FGRGPF QP VJG VKOG UGIOGPVCVKQP QH J[RQVJGUGU YG ECP QDVCKP HWTVJGT UGCTEJ GHſEKGPE[ D[ TGOQXKPI VKOG KPHQTOCVKQP HTQO VJG NCVVKEGU CU HQNNQYU .GV DG VJG QRGTCVQT VJCV UVTKRU VJG VKOG UGIOGP VCVKQPU HTQO J[RQVJGUGU )KXGP C RCTVKCN J[RQVJGUKU HTQO VJG UVCEM Ï NGV DG VJG DG KVU YQTF EQPVGPVU .GV ¾Ï KPFWEGF VQVCN RTQDCDKNKV[ QH QXGT VJG EWTTGPV UVCEM 6JG EQUV HWPEVKQP QH 'SWC VKQP ECP DG TGCTTCPIGF WUKPI VJG QRGTCVQT CU
¾Ï ¡ ¾Ï ¡ ¾Ï
¡
¡
¡¾ Ï ¾ Ï ¡ ¾ Ï
¾ Ï ¡¾ Ï ¡ ¾ Ï
E\&5&3UHVV//&
¡
¡
¡
¡
¾ Ï ¡¾ Ï ¡¾ Ï
6JGTGHQTG VJG EQUV QH C RCTVKCN J[RQVJGUKU FGRGPFU QPN[ QP KVU YQTF EQPVGPVU 6JKU UWIIGUVU VJCV YG ECP KPVTQFWEG C prefix tree CU C EQORCEV TGRTGUGPVCVKQP QH VJG YQTF UGSWGPEGU CUUQEKCVGF YKVJ CNN RCTVKCN J[RQVJGUGU KP VJG UVCEM # PQFG KP VJG RTGſZ VTGG KFGPVKſGU C UGV QH J[RQVJGUGU CPF VJGKT GPF PQFGU KP VJG NCVVKEG 0QY VJG UGCTEJ ECP DG RGTHQTOGF QXGT VJG RTGſZ VTGG +V KU VJG UCOG CU VJG UKPING UVCEM UGCTEJ GZEGRV VJCV 6JG UVCEM EQPVCKPU RTGſZ VTGG PQFGU CPF KU QTFGTGF ſTUV D[
'SWCVKQP CPF VJGP D[ KP ECUG QH VKGU
6JG NCVVKEG RCVJU EQTTGURQPFKPI VQ VJG RTGſZ VTGG PQFG CV VJG VQR QH VJG UVCEM CTG GZVGPFGF D[ QPG YQTF 6JGUG GZVGPUKQPU [KGNF C PGY UGV QH RTGſZ VTGG PQFGU VQ DG KPUGTVGF KP VJG UVCEM 6JG QXGTGUVKOCVG KP VJG RTGſZ VTGG UGCTEJ KU UVKNN EQORWVGF CEEQTFKPI VQ 'SWC VKQP &WG VQ KVU FGRGPFGPEG QP VJG NQPIGUV EQORNGVKQP QH NCVVKEG RCVJU FKHHGTGPV RCVJU CV QPG RTGſZ VTGG PQFG YQWNF EQPVTKDWVG FKHHGTGPVN[ VQ VJKU QXGTGUVKOCVG 1VJGT QXGTGUVKOCVGU VJCV CTG FGRGPFGPV QPN[ QP VJG RTGſZ VTGG PQFG EQWNF DG FGTKXGF QPG UWEJ GUVKOCVG EQWNF DG DCUGF QP VCMKPI VJG OCZKOWO XCNWG COQPI VJG NQPIGUV EQO RNGVKQP NGPIVJU QH CNN VJG NCVVKEG RCVJU VJCV GPF CV QPG RTGſZ VTGG PQFG # UKIPKſECPV CFXCPVCIG QH WUKPI RTGſZ VTGGU HQT VJG .GXGPUJVGKP FKUVCPEG KU VJCV VJG[ HCEKNKVCVG UVQTCIG CPF EQORWVCVKQP QH
¡¾ Ï ¡¾ Ï
6JKU SWCPVKV[ PCOGF partial hypothesis comparison cost KU PGGFGF KP 'SWCVKQP CDQXG 'HſEKGPV EQORWVCVKQP QH VJG RCTVKCN J[RQVJGUKU EQORCTKUQP EQUV KU GUUGPVKCN HQT VJG HGCUKDKNKV[ CPF URGGF QH VJG £ UGCTEJ &WG VQ VJG TGEWTUKXG PCVWTG QH VJG .GXGPUJVGKP FKUVCPEG VJG RCTVKCN J[RQVJGUKU EQORCTKUQP EQUV ECP DG EQORWVGF RTQ ITGUUKXGN[ CU VJG UGCTEJ RTQEGGFU 2.2.3.5 Pruning and Multistack Organization of the Prefix Tree Search #NVJQWIJ VJG FGTKXCVKQPU QH WPFGT GUVKOCVGU CPF QXGT GUVKOCVGU QH EQUVU 'SWCVKQPU CPF FKF PQV VCMG UVCEM RTWPKPI KPVQ CEEQWPV RTWPKPI KU GUUGPVKCN HQT VJGUG CNIQ TKVJOU VQ DG HGCUKDNG = ? 9JGP GPVTKGU CTG RTWPGF HTQO VJG UVCEM 'SWCVKQP KU UVKNN C XCNKF WPFGT GUVKOCVG DWV 'SWCVKQP KU PQ NQPIGT C XCNKF QXGT GUVKOCVG +V KU JQYGXGT C XCNKF QXGT GUVKOCVG HQT VJG UWDNCVVKEG QH VJG QTKIKPCN NCVVKEG VJCV EQWNF DG EQPUVTWEVGF D[ EQORNGVKQP QH VJG RCTVKCN J[RQVJGUGU KP VJG RTWPGF UVCEM 6JGTG HQTG KP VJG UGCTEJ CNIQTKVJOU CDQXG YG ECP CV DGUV JQRG VQ ſPF VJG QRVKOCN UQNWVKQP YKVJKP VJKU UWDNCVVKEG
E\&5&3UHVV//&
6JG UKPING UVCEM UGCTEJ 5GEVKQP CPF VJG RTGſZ VTGG UGCTEJ 5GEVKQP JCXG VJG FKUCFXCPVCIG VJCV VJG EQUVU QH RCTVKCN J[RQVJGUGU QH FKHHGTGPV NGPIVJU CTG EQORCTGF 6JKU KU CEEGRVCDNG WPFGT VJG UGCTEJ HQTOWNCVKQP DWV KU PQV C IQQF EQORCT KUQP HQT WUG KP RTWPKPI UKPEG KV HCXQTU UJQTV J[RQVJGUGU 6JWU KV OC[ DG UWDQRVKOCN VQ RTWPG ECPFKFCVGU DCUGF QP VJGKT EQUV KP VJG UKPING UVCEM +P CP CVVGORV VQ CXQKF VJKU YG WUG C OWNVKUVCEM KORNGOGPVCVKQP YJKEJ KU C HCKTN[ UKORNG GZVGPUKQP QH VJG RTGſZ VTGG UGCTEJ VJCV OCKPVCKPU C UGRCTCVG UVCEM HQT GCEJ J[RQVJGUKU NGPIVJ 6JKU OWNVK UVCEM QTICPK\CVKQP JCU DGGP HQWPF VQ JCXG DGVVGT RTWPKPI EJCTCEVGTKUVKEU KP RTCEVKEG +V KU VJKU OWNVKUVCEM RTGſZ VTGG UGCTEJ VJCV YG TGRQTV VJG TGUWNVU QP 2.2.3.6 Loss Functions Other than Levenshtein Distance (TQO VJG UGCTEJ HQTOWNCVKQP QH 5GEVKQP KV KU ENGCT VJCV VJG HGCUKDKNKV[ QH VJG UGCTEJ FGRGPFU QP VJG CDKNKV[ VQ EQORWVG VJG VYQ EQUV HWPEVKQPU 'SWCVKQPU CPF VJCV RTQXKFG NQYGT CPF WRRGT DQWPFU QP VJG GZRGEVGF NQUU 1PG UWEJ RCKT QH EQUV HWPEVKQPU KU RTQXKFGF HQT VJG .GXGPUJVGKP NQUU HWPEVKQP KP 'SWCVKQPU CPF +V ECP DG UGGP HTQO VJG FGTKXCVKQP QH VJG WPFGT GUVKOCVG EQUV QH 'SWC VKQP VJCV KV KU FKTGEVN[ IGPGTCNK\CDNG VQ CP[ CTDKVTCT[ NQUU HWPEVKQP KH VJG GHſEKGPV EQORWVCVKQP QH VJG RTGſZ EQORCTKUQP EQUV HQT VJCV NQUU HWPEVKQP KU RQUUKDNG 6JG EQORWVCVKQP QH CP QXGT GUVKOCVG EQUV PGGFU VQ DG CFFTGUUGF QP C ECUG D[ ECUG DCUKU
2.3 Segmental MBR Procedures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high confidence regions YKVJKP VJG GXKFGPEG URCEG RTQFWEGF D[ VJG ſTUVRCUU #54 U[UVGO 9KVJKP VJGUG TGIKQPU VJG #54 U[UVGO YCU CDNG VQ RTQFWEG TGNKCDNG YQTF J[RQVJGUGU 5/$4 VJGP HQEWUGU QP VJG low confidence regions KP YJKEJ VJG ſTUVRCUU U[UVGO HCKNGF VQ RTQFWEG C J[RQVJGUKU YKVJ EQPſFGPEG 6JG XCNWG QH VJKU KU VJCV UGCTEJ URCEG KU GZRCPFGF YJGTG VJG ſTUVRCUU U[UVGO FKF PQV RGTHQTO YGNN CPF EQPVTCEVGF YJGTG VJG KPKVKCN J[RQVJGUKU KU CFGSWCVG 9G PQY RTGUGPV C IGPGTCN HQTOWNCVKQP QH VJGUG 5/$4 RTQEGFWTGU CHVGT YJKEJ UGXGTCN URGEKſE XCTKCPVU YKNN DG FGUETKDGF 9G ſTUV FGUETKDG VJG UGIOGPVCVKQP RTQEGUU .GV DG CP GXKFGPEG UGIOGPVCVKQP TWNG VJCV WPKSWGN[ UGIOGPVU GCEJ YQTF UVTKPI KP KPVQ UWDUVTKPIU QH \GTQ QT OQTG YQTFU #RRN[KPI VQ IGPGTCVGU segment sets 6JGUG UGIOGPV UGVU EQPUKUV QH UWDUVTKPIU HTQO VJG QTKIKPCN GXKFGPEG URCEG FGPQVGU VJG
Ï
E\&5&3UHVV//&
Ï
Ï
Ï
UGIOGPV QH VJG YQTF UGSWGPEG KG +P C UKOKNCT YC[ NGV DG C J[RQVJGUKU UGIOGPVCVKQP TWNG EQWRNGF YKVJ WPKSWGN[ FKXKFGU GCEJ UVTKPI KP VJG J[RQVJGUKU URCEG UGIOGPVU CPF KPVQ ¼ ¼ FGPQVGU VJG UGIOGPV QH VJG J[RQVJGUKU 6JG EQPUVTCKPV QP KU VJCV KV OWUV JCXG C conjunction rule HQT EQPECVGPCVKPI UVTKPIU HTQO UGIOGPV UGVU ¼ KU UGIOGPVGF CRRN[KPI 6JG EQPLWPEVKQP TWNG OWUV DG UWEJ VJCV YJGP ¼ VQ VJG UGIOGPVU TGRTQFWEGU 6Q UWOOCTK\G JQY VJG UGIOGPVCVKQP CPF EQPLWPEVKQP RTQEGUU YKNN DG WUGF KP FGEQF KPI VJG J[RQVJGUKU UGIOGPVCVKQP TWNG YKNN DG WUGF VQ FGſPG J[RQVJGUKU UGVU # UKPING J[RQVJGUKU ¼ YKNN DG EJQUGP HTQO GCEJ J[RQVJGUKU UGIOGPV UGV DCUGF QP VJG EQTTGURQPFKPI GXKFGPEG UGIOGPV UGV 6JG EQPLWPEVKQP TWNG YKNN DG VJGP ¼ ¼ ¼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
Ï
¼
¼
YJGTG KU C NQUU HWPEVKQP FGſPGF QP VJG UGIOGPV UGV +P GHHGEV YG CUUWOG VJCV GXGP VJQWIJ UGIOGPVCVKQP KPVTQFWEGU EQPUVTCKPVU KP VJG CNKIPOGPV DGVYGGP UG SWGPEGU VJG QXGTCNN NQUU HWPEVKQP KU PQV CHHGEVGF 9G ECP PQY UVCVG VJG HQNNQYKPI RTQRQUKVKQP YJKEJ HQNNQYU FKTGEVN[ D[ VJG UWDUVKVWVKQP QH 'SWCVKQP KPVQ 'SWCVKQP Proposition. #P WVVGTCPEG NGXGN /$4 TGEQIPK\GT QH 'SWCVKQP ECP DG KORNGOGPVGF CU C EQPECVGPCVKQP QH /$4 TGEQIPK\GTU =?
Æ Æ
YJGTG
Æ
CPF
¼
¾Ï
¾Ï
¼
KU VJG OCTIKPCN RTQDCDKNKV[ QXGT VJG GXKFGPEG UGIOGPV UGV
¾Ï
6JGTGHQTG WPFGT VJG CUUWORVKQP QH 'SWCVKQP WVVGTCPEG NGXGN /$4 TGEQIPKVKQP DGEQOGU C UGSWGPEG QH UOCNNGT /$4 TGEQIPKVKQP RTQDNGOU
E\&5&3UHVV//&
9G PQVG VJCV YJKNG VJG WVVGTCPEG NGXGN /$4 TGEQIPK\GT KU KORNGOGPVGF CU C UGSWGPEG QH UGIOGPVCN /$4 TGEQIPK\GTU VJG CEQWUVKE FCVC KU PQV UGIOGPVGF CV CNN #NN GXK FGPEG QTKIKPCNN[ CXCKNCDNG KU WUGF VQ EQORWVG VJG OCTIKPCN RTQDCDKNKVKGU #NUQ PQVG VJCV VJGTG KU PQ CUUWORVKQP QH NKPIWKUVKE KPFGRGPFGPEG DGVYGGP YQTF UVTKPIU DGNQPIKPI VQ CFLCEGPV GXKFGPEG UGIOGPV UGVU VJG NCPIWCIG OQFGN URCPU UGIOGPVU CPF EQWNF GXGP DG CRRNKGF CV VJG GPVKTG WVVGTCPEG NGXGN +P RTCEVKEG KV OC[ DG FKHſEWNV VQ UGIOGPV VJG GXKFGPEG CPF J[RQVJGUKU URCEGU UQ VJCV VJG NQUU HWPEVKQP FKUVTKDWVGU CEEQTFKPI VQ 'SWCVKQP *QYGXGT IKXGP CP[ UGIOGPVCVKQP YG ECP KFGPVKH[ CP CUUQEKCVGF WVVGTCPEG NGXGN induced NQUU HWPEVKQP FGſPGF CU
¼
¼
%NGCTN[ VJG UGIOGPVCN /$4 TGEQIPK\GTU CTG GSWKXCNGPV VQ CP WVVGTCPEG NGXGN /$4 TGEQIPK\GT WPFGT VJG NQUU HWPEVKQP 6JG QXGTCNN RGTHQTOCPEG WPFGT VJG FGUKTGF NQUU HWPEVKQP UJQWNF FGRGPF QP JQY YGNN CRRTQZKOCVGU
2.3.1 Segmental Voting # URGEKCN ECUG QH VJG UGIOGPVCN /$4 TGEQIPKVKQP CTKUGU WPFGT EGTVCKP EQPFKVKQPU 5WRRQUG GCEJ GXKFGPEG UGIOGPV UGV EQPVCKPU CV OQUV QPG YQTF HTQO GCEJ GXKFGPEG YQTF UVTKPI GCEJ J[RQVJGUKU UGIOGPV UGV EQPVCKPU CV OQUV QPG YQTF HTQO GCEJ J[ RQVJGUKU YQTF UVTKPI CPF VJGTG KU C NQUU HWPEVKQP 'SWCVKQP QP UGIOGPV UGVU 7PFGT VJGUG EQPFKVKQPU VJG UGIOGPVCN /$4 TGEQIPK\GT QH 'SWCVKQP DGEQOGU
Æ ¼
¼
¾Ï
YJGTG KU FGſPGF KP C OCPPGT UKOKNCT VQ VJCV QH 'SWCVKQP 'SWCVKQP KU PQPG QVJGT VJCP VJG OCZKOWO CRQUVGTKQTK RTQDCDKNKV[ FGEKUKQP QP GCEJ J[RQVJGUKU UGIOGPV UGV HQT GCEJ J[RQVJGUKU YQTF C OCTIKPCN RTQDCDKNKV[ KU EQORWVGF DCUGF QP VJG GXKFGPEG URCEG 6JG YQTF YKVJ JKIJGUV OCTIKPCN RTQDCDKNKV[ KU VJGP UGNGEVGF 6JKU KU VJG RTQEGFWTG QH UGIOGPVCN XQVKPI 6JG WVVGTCPEG NGXGN KPFWEGF NQUU 'SWCVKQP HQT UGIOGPVCN XQVKPI ECP DG YTKVVGP CU ¼
UGIXQVG
¼
¼
#U KU VJG ECUG YKVJ UGIOGPVCN /$4 TGEQIPKVKQP UGIOGPVCN XQVKPI KU GHHGEVKXG KH KU C IQQF CRRTQZKOCVKQP VQ VJG NQUU VJCV YG CTG VT[KPI VQ OKPKOK\G 5GIOGPVCN /$4 TGEQIPKVKQP FQGU PQV URGEKH[ JQY VQ ſPF VJG J[RQVJGUKU CPF GXK FGPEG UGIOGPV UGV UGIOGPVCVKQP RTQEGFWTGU CPF KV QPN[ URGEKſGU VJG EQP UVTCKPVU VJCV VJGUG RTQEGFWTGU OWUV QDG[ 6JG EQPUVTWEVKQP QH UGIOGPV UGVU VJGTGHQTG TGOCKPU C FGUKIP RTQDNGO VQ DG CFFTGUUGF KP CP CRRNKECVKQP URGEKſE OCPPGT 9G YKNN PQY FGUETKDG VYQ XGTUKQPU QH UGIOGPVCN /$4 TGEQIPKVKQP WUGF KP UVCVGQHVJGCTV
E\&5&3UHVV//&
#54 U[UVGOU $QVJ VJGUG RTQEGFWTGU CVVGORV VQ TGFWEG VJG YQTF GTTQT TCVG 9'4 CPF VJWU CTG DCUGF QP VJG .GXGPUJVGKP NQUU HWPEVKQP =?
2.3.2 ROVER 4GEQIPK\GT QWVRWV XQVKPI HQT GTTQT TGFWEVKQP 418'4 KU CP 0DGUV NKUV UGIOGPVCN XQVKPI RTQEGFWTG +V EQODKPGU VJG J[RQVJGUGU HTQO OWNVKRNG KPFGRGPFGPV TGEQIPK\GTU WPFGT VJG .GXGPUJVGKP NQUU +P KVU QTKIKPCN HQTOWNCVKQP =? GCEJ QH VJGUG QWVRWVU EQPUKUVU QH C UKPING YQTF UVTKPI CPF C YQTF NGXGN EQPſFGPEG UEQTG CUUQEKCVGF YKVJ GCEJ YQTF KP VJCV UVTKPI 2TQEGFWTGU HQT EQODKPKPI 0DGUVU NKUVU HTQO GCEJ U[UVGO JCXG UKPEG DGGP FGXGNQRGF = ? DG 0DGUV NKUVU RTQFWEGF D[ TGEQIPKVKQP U[UVGOU KP TG .GV URQPUG VQ CEQWUVKEU CPF NGV DG VJG RQUVGTKQT FKUVTKDWVKQP CUUQEKCVGF YKVJ .GV FGPQVG VJG WPKQP QH VJGUG 0DGUV NKUVU # RQUVGTKQT FKUVTKDWVKQP QP YQTF UVTKPIU KP KU FGTKXGF D[ ſTUV GZVGPFKPI GCEJ VQ CUUKIP \GTQ RTQDCDKNKV[ VQ YQTF UVTKPIU KP VJCV CTG PQV RTGUGPV KP CPF VJGP VCMKPI C EQPXGZ EQODKPCVKQP
6JG UGV CPF CTG VJG GXKFGPEG URCEG CPF VJG GXKFGPEG FKUVTKDWVKQP WUGF D[ 418'4 Ý 6JG YQTF UVTKPIU QH CTG CTTCPIGF KP C YQTF VTCPUKVKQP PGVYQTM 960 VJCV TGR TGUGPVU CP CRRTQZKOCVG simultaneous alignment QH VJGUG J[RQVJGUGU +V KU IGPGTCVGF D[ RKEMKPI VQR VYQ J[RQVJGUGU CNKIPKPI VJGO VQ RTQFWEG CP KPKVKCN 960 CPF VJGP KVGTCVKXGN[ CFFKPI GCEJ PGY J[RQVJGUKU D[ CNKIPKPI KV YKVJ VJG 960 EQPUVTWEVGF UQ HCT #P GZCORNG 960 RTQFWEGF D[ CNKIPKPI
ő1* 9'.. 9'Œ ő1 9'.. 9'ŏ4'Œ ő9'.. 9' 9'ŏ4'Œ KU IKXGP KP (KIWTG # UGV QH YQTFU VJCV CNKIP YKVJ GCEJ QVJGT KU ECNNGF C correspondence set 6JG 960 VTKXKCNN[ URGEKſGU CP GXKFGPEG UGIOGPVCVKQP TWNG HQT YQTF UVTKPIU QH 6JG J[RQVJGUKU URCEG QH 418'4 KU VJG UGV QH CNN VJG YQTF UVTKPIU VJCV ECP DG RTQFWEGF D[ RKEMKPI QPG YQTF HTQO GCEJ EQTTGURQPFGPEG UGV CPF EQPECVGPCVKPI VJGO 6JGTGHQTG VJG J[RQVJGUKU UGIOGPVCVKQP TWNG CPF VJG EQPLWPEVKQP TWNG CTG CNUQ VTKXKCNN[ URGEKſGF D[ VJG 960 *CXKPI UGIOGPVGF VJG GXKFGPEG CPF J[RQVJG UKU URCEGU C OCTIKPCN RTQDCDKNKV[ KU EQORWVGF HQT GCEJ YQTF KP GCEJ EQTTGURQPFGPEG UGV CEEQTFKPI VQ 'SWCVKQP CPF VJG YQTF YKVJ VJG NCTIGUV OCTIKPCN RTQDCDKNKV[ KU EJQUGP HTQO GCEJ EQTTGURQPFGPEG UGV 6JGUG YQTFU CTG EQPECVGPCVGF VQ HQTO VJG ſPCN QWVRWV QH 418'4
Ý 418'4 QTKIKPCNN[ KPEQTRQTCVGF C YQTF NGXGN EQPſFGPEG UEQTG KPUVGCF QH
VJKU KU FKUEWUUGF D[ )QGN GV CN =?
E\&5&3UHVV//&
CU KP 'SWCVKQP
9'
1*
1
9'ŏ4'
9'ŏ4'
9'..
07..
07..
FIGURE 2.2 An example word transition network.
6JG WVVGTCPEG NGXGN KPFWEGF NQUU 'SWCVKQP KP 418'4 KU FGTKXGF HTQO 'SWC VKQP YJGTG VJG UWO KU QXGT VJG EQTTGURQPFGPEG UGVU
¼
¼
6JKU NQUU KU UKOKNCT VQ VJG .GXGPUJVGKP FKUVCPEG DGVYGGP UVTKPIU CPF ¼ YJGP VJGKT CNKIPOGPV KU URGEKſGF D[ VJG 960 5KPEG VJG 960 EQPUVTWEVKQP RTQEGUU CFFU GCEJ PGY YQTF UVTKPI VQ VJG 960 UQ CU VQ OKPKOK\G VJG CNKIPOGPV EQUV DGVYGGP VJCV UVTKPI CPF VJG 960 YG EQWNF GZRGEV ¼ VQ CRRTQZKOCVG VJG .GXGPUJVGKP FKUVCPEG DGVYGGP CPF ¼
2.3.3 e-ROVER 6JG UKOWNVCPGQWU CNKIPOGPV RTQFWEGF KP 418'4 OC[ UQOGVKOGU DG UWDQRVKOCN HQT UQOG GXKFGPEG J[RQVJGUGU RCKTU 6JG PCVWTCN TGOGF[ KU VQ CNNQY OWNVKRNG EQPUGEW VKXG YQTFU KP GCEJ EQTTGURQPFGPEG UGV %QPUKFGTKPI VJG UGCTEJ HQT UGIOGPV UGVU CU C ENWUVGTKPI RTQDNGO VJGTG CTG VYQ FKHHGTGPV CRRTQCEJGU VJCV EQWNF DG VCMGP 9G EQWNF VCMG C ŎVQRFQYPŏ CRRTQCEJ VJCV UVCTVU YKVJ C UKPING EQTTGURQPFGPEG UGV VJCV EQPVCKPU CP GPVKTG 0DGUV NKUV CPF UGIOGPVU KV KPVQ UGVU VJCV EQPVCKP UJQTVGT YQTF UVTKPIU #NVGT PCVKXGN[ YG EQWNF VCMG C ŎDQVVQOWRŏ CRRTQCEJ YJGTG YG ſTUV EQPUVTWEV C 960 VJCV EQPVCKPU PQ OQTG VJCP QPG EQPUGEWVKXG YQTF KP GCEJ EQTTGURQPFGPEG UGV 9G EQWNF VJGP LQKP EQPUGEWVKXG UGVU VQ QDVCKP UGVU YKVJ NQPIGT YQTF UVTKPIU 6JG RTQEGFWTG QH GZVGPFGF418'4 G418'4 UVCTVU YKVJ VJG 418'4 960 CPF VCMGU VJG NCVVGT DQVVQOWR CRRTQCEJ 6JG RTQEGUU QH joining VYQ EQTTGURQPFGPEG UGVU [KGNFU QPG expanded UGV VJCV EQP VCKPU CNN VJG RCVJU HTQO VJG QTKIKPCN RCKT QH EQTTGURQPFGPEG UGVU 6JKU KU ITCRJKECNN[ KNNWUVTCVGF KP (KIWTG 6JG WVVGTCPEG NGXGN NQUU HWPEVKQP QH G418'4 KU IKXGP CU HQNNQYU 5VCTVKPI HTQO VJG KPKVKCN 960 NGV VYQ EQPUGEWVKXG EQTTGURQPFGPEG UGVU UC[ UGVU CPF DG LQKPGF CPF NGV VJG NQUU HWPEVKQP QP VJG GZRCPFGF UGV DG VJG .GXGPUJVGKP FKUVCPEG 6JG NQUU HWPEVKQP QP EQTTGURQPFGPEG UGVU VJCV FKF PQV GZRCPF TGOCKPU VJG NQUU
E\&5&3UHVV//&
1*
9'..
1
9'
9'ŏ4'
9'ŏ4'
07..
07..
9'9'ŏ4'
1*
9'ŏ4' 1
9'.. 9'
07.. 9'ŏ4'9'ŏ4'
FIGURE 2.3 Joining two correspondence sets. 6JG WVVGTCPEG NGXGN NQUU KU VJGP
¼
¼ *GTG CPF CTG YQTF UWDUGSWGPEGU HTQO VJG LQKPGF UGIOGPV UGVU +V HQNNQYU HTQO VJG FGſPKVKQP QH .GXGPUJVGKP FKUVCPEG VJCV
6JKU HQNNQYU DGECWUG VJG G418'4 CNKIPOGPVU YKNN GXGPVWCNN[ CEJKGXG VJG .GXGP UJVGKP CNKIPOGPV CU VJG CNKIPOGPV EQPUVTCKPVU CTG TGFWEGF 6JG LQKPKPI RTQEGFWTG ECP DG ECTTKGF QWV OCP[ VKOGU VQ [KGNF UWEEGUUKXGN[ DGVVGT CRRTQZKOCVKQPU VQ VJG .GXGPUJVGKP FKUVCPEG 6JG 960 QDVCKPGF CHVGT GCEJ LQKPKPI QRGTCVKQP URGEKſGU C PGY UGIOGPVCVKQP QH VJG GXKFGPEG CPF J[RQVJGUKU URCEGU +P EQORCTKPI G418'4 VQ 418'4 KV KU KORQTVCPV VQ PQVG VJCV QPN[ VJG segmentation QH VJG J[RQVJGUKU CPF GXKFGPEG URCEGU EJCPIGU YKVJ LQKPKPI QRGTCVKQP VJG CEVWCN URCEGU TGOCKP VJG UCOG CU VJG[ YGTG KP 418'4 6JGTG CTG VYQ EQPUGSWGPEGU QH LQKPKPI EQTTGURQPFGPEG UGVU (KTUV CHVGT VJG LQKPKPI QRGTCVKQP VJG NQUU HWPEVKQP QP VJG GZRCPFGF UGV KU PQ NQPIGT VJG NQUU DWV KU KPUVGCF VJG .GXGPUJVGKP FKUVCPEG *GPEG VJG /$4 J[RQVJGUKU UGNGEVKQP QP VJKU UGV PGGFU VQ HQNNQY 'SWCVKQP 5GEQPF VJG UK\G QH VJG GZRCPFGF UGV ITQYU GZRQPGPVKCNN[ YKVJ VJG PWODGT QH LQKPKPI QRGTCVKQPU OCMKPI 'SWCVKQP RTQITGUUKXGN[ FKHſEWNV VQ KORNGOGPV +V KU VJGTGHQTG KORQTVCPV VQ FGVGTOKPG VJG UGVU VQ DG LQKPGF ECTGHWNN[ UQ CU VQ [KGNF OCZKOWO ICKP KP .GXGPUJVGKP FKUVCPEG CRRTQZKOCVKQP YKVJ OKPKOWO EQODKPCVKQPU QH VJG EQTTGURQPFGPEG UGVU # JGWTKUVKE RTQEGFWTG HQT LQKPKPI UGVU =? KU DCUGF QP ſTUV KFGPVKH[KPI EQTTGURQPFGPEG UGVU KP YJKEJ VJG NCTIGUV XCNWG QH VJG
E\&5&3UHVV//&
OCTIKPCN RTQDCDKNKV[ 'SWCVKQP KU DGNQY C VJTGUJQNF 'CEJ EQPUGEWVKXG UVTGVEJ QH UWEJ UGVU KU LQKPGF VQ HQTO QPG GZRCPFGF UGV 5GVU KP YJKEJ VJG NCTIGUV XCNWG QH VJG OCTIKPCN RTQDCDKNKV[ KU CDQXG VJG VJTGUJQNF CTG MGRV ŎRKPEJGFŏ VJG[ CTG PQV LQKPGF YKVJ CP[ QVJGT UGV (QT FGVCKNU QH VJKU RTQEGFWTG TGCFGTU CTG TGHGTTGF VQ )QGN GV CN =? #U PQVGF CDQXG VJG J[RQVJGUKU CPF VJG GXKFGPEG URCEGU KP G418'4 CTG KFGPVKECN VQ VJQUG KP 418'4 *QYGXGT VJG NQUU HWPEVKQP KP G418'4 RTQXKFGU C DGVVGT CR RTQZKOCVKQP VQ VJG YQTF GTTQT TCVG VJCP 418'4 5KPEG VJG[ CTG DQVJ KPUVCPVKCVKQPU QH 'SWCVKQP G418'4 YQWNF DG GZRGEVGF KP VJGQT[ VQ [KGNF C NQYGT YQTF GTTQT TCVG VJCP 418'4
2.4 Experimental Results 6JG OKPKOWO $C[GUTKUM RTQEGFWTGU [KGNF C VJGQTGVKECNN[ NQYGT GZRGEVGF GTTQT TCVG VJCP VJG /#2 TGEQIPK\GT *QYGXGT VJGKT RTCEVKECN OGTKV ECP QPN[ DG ICWIGF KP TGCN ENCUUKſECVKQP VCUMU +P VJKU UGEVKQP YG RTGUGPV GZRGTKOGPVU VJCV EQORCTG /$4 CPF UGIOGPVCN /$4 RTQEGFWTGU YKVJ /#2 TGEQIPKVKQP CPF YKVJ GCEJ QVJGT 9G ſTUV CFFTGUU C RTCEVKECN RTQDNGO CUUQEKCVGF YKVJ KPEQTRQTCVKQP QH VJG *// CPF /CTMQX EJCKP OQFGNU KPVQ VJG OKPKOWOTKUM UGCTEJ RTQEGFWTGU FGUETKDGF CDQXG
2.4.1 Parameter Tuning within the MBR Classification Rule 6JG LQKPV FKUVTKDWVKQP VQ DG WUGF KP VJG /$4 TGEQIPK\GTU KU FGTKXGF D[ EQODKPKPI RTQDCDKNKVKGU HTQO CEQWUVKE CPF NCPIWCIG OQFGNU +V KU QHVGP HQWPF WUGHWN KP RTCEVKEG VQ KPVTQFWEG UQOG VWPKPI RCTCOGVGTU VQ JGNR OCVEJ VJGUG OQFGNU DGVVGT +P VJG HQNNQYKPI YG FKUEWUU C RCTCOGVGTK\CVKQP QH VJCV KU UWKVCDNG HQT WUG KP /$4 TGEQIPK\GTU 9G VJGP RTGUGPV UVTCVGIKGU VQ QRVKOK\G VJGUG RCTCOGVGTU YKVJKP VJG /$4 ENCUUKſECVKQP TWNG +V KU EWUVQOCT[ KP #54 VQ WUG VYQ VWPKPI RCTCOGVGTU KP VJG EQORWVCVKQP QH LQKPV RTQDCDKNKV[
YJGTG KU VJG PWODGT QH YQTFU KP YQTF UVTKPI 6JG RCTCOGVGT WUWCNN[ C PGICVKXG EQPUVCPV ECWUGU C FGETGCUG QH RTQDCDKNKV[ YKVJ KPETGCUKPI (QT VJKU TGCUQP KV KU ECNNGF word insertion penalty 6JG QVJGT RCTCOGVGT UECNGU VJG NCPIWCIG OQFGN RTQDCDKNKV[ TGNCVKXG VQ VJG CEQWUVKE OQFGN RTQDCDKNKV[ KV KU VGTOGF language model scale factor 9G JCXG HQWPF KV WUGHWN VQ KPVTQFWEG CP CFFKVKQPCN likelihood scale factor =?
½
6JG NKMGNKJQQF UECNG HCEVQT TGUVTKEVU VJG F[PCOKE TCPIG QH VJG RTQDCDKNKVKGU (QT GZCO RNG EQPUKFGT VJG DGUV NKUV QH 6CDNG 6JGUG CTG VGP OQUV NKMGN[ YQTF UVTKPIU RTQ
E\&5&3UHVV//&
TABLE 2.1 'ZCORNG VGP OQUV NKMGN[ J[RQVJGUGU CPF VJG RQUVGTKQT RTQDCDKNKV[ QH VJGUG J[RQVJGUGU WPFGT VYQ FKHHGTGPV RCTCOGVGTK\CVKQPU 'SWCVKQPU CPF QH VJG RQUVGTKQT FKUVTKDWVKQP CPF
' ' ' ' ' ' ' ' '
5GPVGPEG + *#8' # 474#. #4'# + *#8' # 4'#. 474#. #4'# #.6*17)* +0 # 474#. #4'# + .+8' +0 # 474#. #4'# #.6*17)* +6 9+.. #4'# 51 + *#8' # 474#. #4'# *#8' # 474#. #4'# +ŏ/ # 474#. #4'# + *#8' # .+66.' 474#. #4'# + *#8' # 41.' #4'#
FWEGF D[ QWT #54 U[UVGO HQT CP WVVGTCPEG ő+ .+8' +0 # 474#. #4'#Œ 6JG NQI NKMGNKJQQFU CNQPI YKVJ VJG RQUVGTKQT FKUVTKDWVKQPU EQORWVGF YKVJ CPF YKVJQWV VJG NKMGNKJQQF UECNG HCEVQT CTG UJQYP KP 6CDNG 6JG RQUVGTKQT FKU VTKDWVKQP KU EQORWVGF D[ GZRQPGPVKCVKPI VJG NQINKMGNKJQQFU CPF VJGP PQTOCNK\KPI VJGO QXGT VJG VGPDGUV NKUV (TQO 6CDNG KV ECP DG UGGP VJCV KU JGCXKN[ YGKIJVGF VQYCTFU VJG OQUV NKMGN[ ECPFKFCVG QYKPI VQ C NCTIG XCTKCVKQP KP NQINKMGNKJQQF XCNWGU 6JKU NGCFU VQ C FGIGPGTCVKQP QH VJG GXKFGPEG URCEG +V KU RTGXGPVGF D[ VJG KPVTQFWEVKQP QH VJG NKMGNK JQQF UECNG HCEVQT YJKEJ ƀCVVGPU VJG FKUVTKDWVKQP CPF [KGNFU OQTG TGCUQPCDNG RQUVGTKQT RTQDCDKNKVKGU # XCNWG QH KU WUGF KP QWT GZCORNG QH 6CDNG 2.4.1.1 Optimization of Likelihood Parameters .GV Æ DG VJG OKPKOWOTKUM TGEQIPK\GT 'SWCVKQP KPEQTRQTCVKPI VJG RCTCO GVGTK\GF FKUVTKDWVKQP QH 'SWCVKQP 9G QRVKOK\G CPF VQ OKPKOK\G VJG GORKTKECN TKUM =? QH Æ
¾Ì
Æ
QXGT C FCVCDCUG QH NCDGNGF WVVGTCPEGU 5KPEG VJG WVVGTCPEG NCDGNU CTG MPQYP VJKU KU supervised optimization (QT UQOG RTQDNGOU KV OC[ DG FGUKTCDNG VQ VWPG ENCUUKſECVKQP TWNG RCTCOGVGTU YKVJQWV WUKPI C UGRCTCVG VTCKPKPI UGV 9G CRRTQCEJ VJKU unsupervised optimization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
E\&5&3UHVV//&
+P QTFGT VQ TGFWEG VJG PWODGT QH RCTCOGVGTU VQ DG VTCKPGF YG MGRV VJG YQTF KPUGTVKQP RGPCNV[ CPF NCPIWCIG OQFGN UECNG HCEVQT ſZGF CV VJGKT XCNWGU QDVCKPGF HTQO VTCKPKPI YKVJ /#2 ENCUUKſGT +P CNN QWT GZRGTKOGPVU TGRQTVGF KP VJKU VJGUKU VJGUG XCNWGU YGTG CPF # ITKF UGCTEJ HQT QRVKOCN YCU RGTHQTOGF KP DQVJ UWRGTXKUGF CPF WPUWRGTXKUGF QRVKOK\CVKQP #P CNVGTPCVKXG VQ VTCKPKPI KU VQ WUG =? 9G EQORCTG CNN VJTGG OGVJQFU HQT QDVCKPKPI FGUETKDGF JGTG KP VJG GZRGTKOGPVU VQ HQNNQY
2.4.2 Utterance Level MBR Word and Keyword Recognition 9G PQY GXCNWCVG WVVGTCPEG NGXGN 0DGUV NKUV TGUEQTKPI 5GEVKQP CPF RTGſZ VTGG DCUGF UGCTEJ 5GEVKQP HQT VCUMU QH VTCPUETKRVKQP CPF MG[YQTF URQVVKPI 6TCPUETKRVKQP KU VJG VCUM QH KFGPVKH[KPI YQTF EQPVGPV QH URQMGP CEQWUVKEU +VU GTTQT TCVG KU OGCUWTGF D[ VJG .GXGPUJVGKP FKUVCPEG ¼ DGVYGGP VJG CEVWCNN[ URQ MGP WVVGTCPEG CPF VJG TGEQIPK\GTŏU QWVRWV 6JG IQCN QH MG[YQTF URQVVKPI KU VQ KFGPVKH[ VJG RTGUGPEG CPF UQOGVKOGU VJG VKOG NQECVKQP QH C RTGURGEKſGF UGV QH MG[YQTFU # NQUU HWPEVKQP UWKVCDNG HQT UWEJ C VCUM YQWNF RC[ CVVGPVKQP QPN[ VQ VJG MG[YQTFU QVJGT URQMGP YQTFU UJQWNF DG KIPQTGF 9G EJQUG VQ GZRGTKOGPV YKVJ C NQUU HWPEVKQP DCUGF QP C XCTKCPV QH .GXGPUJVGKP FKUVCPEG VJCV CUUKIPU C EQUV QH QPG YJGP VJGTG KU CP GTTQT QP C MG[YQTF CPF CUUKIPU PQ EQUV VQ GTTQTU QP QVJGT YQTFU 6JG FGſPKVKQP WUGF YCU
¼
¼
YJGTG KU VJG .GXGPUJVGKP FKUVCPEG CPF KU FGTKXGF HTQO D[ FGNGVKPI CNN KVU PQPMG[YQTFU #54 RGTHQTOCPEG OGCUWTGF WPFGT ¼ YKNN DG TGHGTTGF VQ CU MG[YQTF GTTQT TCVG -'4 'ZRGTKOGPVU YGTG EQPFWEVGF QP VJG 5YKVEJDQCTF =? EQTRWU VJCV EQPUKUVU QH URQP VCPGQWU VGNGRJQP[ EQPXGTUCVKQPU DGVYGGP KPFKXKFWCNU 6JG VGUV UGV YCU C NKPIWKUVK ECNN[ UGIOGPVGF UWDUGV QH VJKU EQTRWU WUGF HQT VJG ,QJPU *QRMKPU 7PKXGTUKV[ .8%54 9QTMUJQR =? 6JKU VGUV UGV EQPVCKPGF WVVGTCPEGU HTQO EQPXGTUC VKQP UKFGU VJG EQORNGVG VGUV UGV FGſPKVKQP CPF QVJGT FGVCKNU ECP DG HQWPF KP VJG YQTM UJQR RTQEGGFKPIU 9QTF NCVVKEGU YGTG IGPGTCVGF WPFGT C VTKITCO NCPIWCIG OQFGN WU KPI URGCMGT CPF IGPFGT KPFGRGPFGPV *6-DCUGF EQORQPGPV )CWUUKCP OKZVWTG ETQUUYQTF VTKRJQPG U[UVGO =? YKVJ VTKRJQPG UVCVGU (QT WUG KP VJG 0DGUV NKUV TGUEQTKPI RTQEGFWTG QH 'SWCVKQP GNGOGPV 0DGUV NKUVU YGTG IGPGTCVGF HTQO VJG VTKITCO NCVVKEGU 6JGUG YGTG WUGF CU VJG GXKFGPEG URCEG CPF HTQO VJGO VJG VQR GNGOGPVU YGTG MGRV CU VJG J[RQVJGUKU URCEG 6JG /#2 ECPFKFCVG KP VJGUG 0DGUV NKUVU CPF JGPEG KP VJGUG NCVVKEGU UGTXGF CU VJG DCUGNKPG YKVJ C YQTF GTTQT TCVG QH CPF C UGPVGPEG GTTQT TCVG 5'4 QH 9QTFU KP VJG VCUM XQECDWNCT[ YGTG OCTMGF CU MG[YQTFU KH VJG[ QEEWTTGF TGNCVKXGN[ KPHTGSWGPVN[ KP C NCTIG EQTRWU =? 'ZCORNGU QH VJG VYQ MKPFU QH YQTFU CTG -G[YQTFU 0QPMG[YQTFU
E\&5&3UHVV//&
abilities, bartenders, calculation, databases a, and, the, besides, collaboration, distribution
6JG PWODGTU KP RCTGPVJGUGU CDQXG FGPQVG VJG VQVCN PWODGT QH FKUVKPEV YQTFU QH VJCV MKPF KP VJG U[UVGO XQECDWNCT[ QH UK\G 'XGP VJQWIJ VJG PQPMG[YQTFU EQP UVKVWVG C UOCNN HTCEVKQP QH VJG XQECDWNCT[ VJG[ CTG SWKVG CDWPFCPV CPF CEEQWPV HQT OQTG VJCP QH VJG YQTF VQMGPU 6JG HWNN XQECDWNCT[ YKVJ OCTMGF MG[YQTFU ECP DG HQWPF CV QWT YGD UKVG =? 2.4.2.1 Likelihood Scale Factor Tuning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
2.4.2.2 N-best List Rescoring and Search 6JG 0DGUV NKUV TGUEQTKPI RTQEGFWTG QH 5GEVKQP YCU KORNGOGPVGF YKVJ C GNGOGPV GXKFGPEG URCEG CPF C GNGOGPV J[RQVJGUKU URCEG HQT DQVJ 9'4 CPF -'4 4GUWNVU QH VJKU TGUEQTKPI WPFGT 9'4 CTG NKUVGF KP 5GEVKQP # QH 6CDNG CPF VJQUG WPFGT -'4 CTG NKUVGF KP 5GEVKQP $ QH 6CDNG .QQMKPI WPFGT VJG 9'4 EQNWOP KP 5GEVKQP # CPF WPFGT VJG -'4 EQNWOP KP 5GEVKQP $ YG PQVG VJCV 0DGUV NKUV TGUEQTKPI [KGNFU C UOCNN [GV UKIPKſECPV KORTQXGOGPV QXGT EQTTGURQPFKPI /#2 DCUGNKPGU (WT VJGTOQTG TGUEQTKPI HQT 9'4 KU PQV CHHGEVGF D[ VJG NKMGNKJQQF UECNG HCEVQT UGNGEVKQP OGVJQF YJGTGCU HQT -'4 VJG WPUWRGTXKUGF QRVKOK\CVKQP OGVJQF QWVRGTHQTOU VJG QVJGT VYQ OGVJQFU 6JG OWNVKUVCEM RTGſ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ſZ VTGG UGCTEJ VQQM CRRTQZKOCVGN[ VYKEG CU NQPI CU VJG 0DGUV NKUV TGUEQTKPI RTQEGFWTG .QQMKPI CV VJG 9'4 RGTHQTOCPEG QH 9'4 QRVKOK\GF UGCTEJ CPF -'4 RGTHQTOCPEG QH -'4 QRVKOK\GF UGCTEJ KP 6CDNG YG PQVG VJCV VJG UGCTEJ [KGNFU UKIPKſECPV GTTQT TCVG TGFWEVKQP QXGT VJG EQTTGURQPFKPI 0DGUV NKUV TGUEQTKPI RTQEGFWTGU 6JG KORQTVCPEG QH WPUWRGTXKUGF QRVKOK\CVKQP OGVJQF QH NKMGNKJQQF UECNG RCTCOGVGT KU CNUQ OQTG RTQOKPGPV KP VJKU ECUG #P QXGTCNN KPETGCUG KP VJG 9'4 HQT J[RQVJGUGU QRVKOK\GF HQT -'4 CPF VJG -'4
E\&5&3UHVV//&
TABLE 2.2
'XCNWCVKQP QH RCTCOGVGT VWPKPI CPF TGEQIPKVKQP RTQEGFWTGU HQT OKPKOK\CVKQP QH 9'4 CPF -'4
# $
4GEQIPKVKQP 6WPKPI %TKVGTKQP 9'4 9'4 9'4 -'4 -'4 -'4
$CUGNKPG /#2 9'4 -'4 2CTCOGVGT .KMGNKJQQF 4GEQIPKVKQP 5VTCVGI[ 6WPKPI 5ECNG 0DGUV 5VTCVGI[ (CEVQT 9'4 -'4 9'4 -'4 ./ 5ECNG 5WRGTXKUGF 7PUWRGTXKUGF ./ 5ECNG 0# 5WRGTXKUGF 0# 7PUWRGTXKUGF 0#
RGTHQTOCPEG QH J[RQVJGUGU QRVKOK\GF HQT 9'4 TGKPHQTEGU VJCV CU FGUKTGF C VCUM URGEKſE OKPKOWOTKUM ENCUUKſGT QWVRGTHQTOU ENCUUKſGTU QRVKOK\GF HQT QVJGT VCUMU
2.4.3 ROVER and e-ROVER for Multilingual ASR +P VJKU UGEVKQP YG GXCNWCVG VJG 0DGUV NKUV DCUGF UGIOGPVCN /$4 RTQEGFWTGU QH 418'4 5GEVKQP CPF G418'4 5GEVKQP 9G YKNN CRRN[ VJGUG OGVJQFU VQ OWNVKNKPIWCN NCPIWCIG KPFGRGPFGPV CEQWUVKE OQFGNKPI =? 6JG QDLGEVKXG JGTG KU VQ VTCKP C OQPQNKPIWCN U[UVGO QP C UOCNN COQWPV QH VTCPUETKDGF URGGEJ CPF VJGP VQ KORTQXG KVU RGTHQTOCPEG WUKPI CEQWUVKE OQFGN VTCKPGF KP QVJGT NCPIWCIGU 1PG QH VJG ENCKOGF CFXCPVCIGU QH 418'4 VGEJPKSWGU KU VJG CDKNKV[ VQ EQODKPG OWNVKRNG #54 U[UVGOU VQ IGPGTCVG C UKPING J[RQVJGUKU 9G YKNN UJQY VJCV 418'4 FQGU KPFGGF KORTQXG QXGT VJG RGTHQTOCPEG QH C OQPQNKPIWCN U[UVGO CPF VJCV G418'4 ECP DG WUGF HQT HWTVJGT KORTQXGOGPVU 6JTGG U[UVGOU YGTG EQODKPGF C VTKRJQPG U[UVGO VTCKPGF QP QPG JQWT QH %\GEJ XQKEG QH #OGTKEC %<81# FCVCDCUG Þ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ſTUVRCUU NCVVKEGU RTQFWEGF D[ VJG %\GEJ OQPQNKPIWCN U[UVGO (QT GCEJ U[UVGO VJG YQTF KPUGTVKQP RGPCNV[ CPF VJG NCPIWCIG OQFGN UECNG HCEVQT
5GEVKQP YGTG EJQUGP VQ [KGNF QRVKOCN RGTHQTOCPEG D[ VJG /#2 FGEKUKQP Þ #XCKNCDNG HTQO VJG .KPIWKUVKE &CVC %QPUQTVKWO .&%5 8QKEG QH #OGTKEC 81# %\GEJ $TQCF ECUV 0GYU #WFKQ
E\&5&3UHVV//&
TCVKQ
GŦ418'4 418'4
RKPEJKPIVJTGUJQNF
9'4
GŦ418'4 418'4
RKPEJKPIVJTGUJQNF
FIGURE 2.4 Top panel shows the ratio of total number of e-ROVER correspondence sets to that of ROVER correspondence sets, as a function of the pinching threshold. Bottom panel shows the WER performance of e-ROVER for these thresholds. TWNG 6JG NKMGNKJQQF UECNG HCEVQT YCU QDVCKPGF D[ EQPFWEKPI CP WPUWRGTXKUGF QR VKOK\CVKQP 5GEVKQP UGRCTCVGN[ HQT GCEJ U[UVGO 418'4 CPF G418'4 YGTG KORNGOGPVGF D[ EQODKPKPI VJGUG VJTGG UGVU QH J[RQVJGUGU 6JG RQUVGTKQT FKU VTKDWVKQP QXGT VJG TGUWNVKPI DGUV NKUV YCU FGTKXGF D[ UKORN[ TGPQTOCNK\KPI VJG NQINKMGNKJQQFU QH VJG UECNGF KPFKXKFWCN J[RQVJGUGU 2.4.3.1 Correspondence Set Pinching +P G418'4 VJG EQTTGURQPFGPEG UGVU YGTG LQKPGF WUKPI VJG JGWTKUVKE RTQEGFWTG FG UETKDGF KP 5GEVKQP 6JKU RTQEGFWTG LQKPU VJG EQTTGURQPFGPEG UGVU DCUGF QP C ŒRKPEJKPI VJTGUJQNFŒ VJCV EQPUKFGTU VJG NCTIGUV RQUVGTKQT RTQDCDKNKV[ QH CP[ YQTF UVTKPI KP GCEJ EQTTGURQPFGPEG UGV # VJTGUJQNF QH TGUWNVU KP PQ LQKPKPI CV CNN YJKEJ KU GSWKXCNGPV VQ 418'4 YJKNG CP[ VJTGUJQNF CDQXG OGTIGU CNN VJG EQTTG URQPFGPEG UGVU 1WT KORNGOGPVCVKQP QH 418'4 TGUWNVGF KP C 9'4 YJKEJ KU C CD UQNWVG KORTQXGOGPV QXGT VJG DGUV /#2 YQTF GTTQT TCVG QH VJG VJTGG U[UVGOU DGKPI EQODKPGF (KIWTG UJQYU VJCV CFFKVKQPCN ICKPU ECP DG QDVCKPGF WUKPI G418'4 6JG VQR RCPGN UJQYU VJG TCVKQ QH VQVCN PWODGT QH G418'4 EQTTGURQPFGPEG UGVU VQ VQVCN PWODGT QH 418'4 EQTTGURQPFGPEG UGVU CU C HWPEVKQP QH VJG RKPEJKPI VJTGUJ QNF 6JKU TCVKQ KU HQT VJTGUJQNF XCNWG QH CPF FGETGCUGU OQPQVQPKECNN[ CU VJG VJTGUJQNF KPETGCUGU +V KU PQV CV KVU OKPKOWO HQT C VJTGUJQNF QH FWG VQ VJG RTGU GPEG QH EQTTGURQPFGPEG UGVU YJKEJ EQPVCKP QPN[ QPG YQTF VJGUG UGVU JCXG C YQTF YKVJ OCTIKPCN RTQDCDKNKV[ QH CPF TGOCKPGF RKPEJGF HQT C VJTGUJQNF XCNWG QH
E\&5&3UHVV//&
6JG DQVVQO RCPGN KP (KIWTG UJQYU VJG GHHGEV QH RKPEJKPI QP 9'4 9G PQVG VJCV CNN VJTGUJQNFU TGUWNV KP DGVVGT VJCP 418'4 YQTF GTTQT TCVG 6JG VJTGUJQNF QH [KGNFU VJG DGUV RGTHQTOCPEG QH CDUQNWVG KORTQXGOGPV QXGT 418'4 CPF JGPEG C VQVCN QH CDUQNWVG QXGT VJG DGUV DCUGNKPG GTTQT TCVG 9G UGG C FGITCFCVKQP KP RGTHQTOCPEG HQT VJTGUJQNFU NCTIGT VJCP 1PG RQUUKDNG GZRNCPCVKQP KU VJG PGGF HQT JGCXKGT RTWPKPI FWG VQ VJG ITGCVN[ GPNCTIGF UGCTEJ URCEG VJCV TGUWNVU HTQO GZRCPFKPI CNN VJG UGIOGPV UGVU #PQVJGT RQUUKDKNKV[ KU VJCV VJG DGUV UVTCVGI[ KU VQ TGVCKP VJG YQTF UGIOGPVU VJCV YGTG TGEQIPK\GF YKVJ CDUQNWVG EGTVCKPV[ D[ VJG ſTUVRCUU U[UVGO
2.5 Summary 9G JCXG FGUETKDGF CWVQOCVKE URGGEJ TGEQIPKVKQP CNIQTKVJOU VJCV CVVGORV VQ OKPKOK\G VJG CXGTCIG OKUTGEQIPKVKQP EQUV WPFGT VCUM URGEKſE NQUU HWPEVKQPU 6JGUG TGEQIPK\ GTU CNVJQWIJ IGPGTCNN[ OQTG EQORWVCVKQPCNN[ EQORNGZ VJCP OQTG YKFGN[ WUGF /#2 CNIQTKVJOU ECP DG GHſ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ſ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
E\&5&3UHVV//&
2.6 Acknowledgements 9G VJCPM &KOKVTC 8GTI[TK HQT RTQXKFKPI VJG NCVVKEGU VJCV YGTG WUGF KP QWT GZRGTKOGPVU CPF -WOCT 5JCPMCT HQT JGNR YKVJ VJG GZRGTKOGPVU 9G CNUQ VJCPM #PFTGCU 5VQNEMG CPF .KFKC /CPIW HQT WUGHWN FKUEWUUKQPU
References =? 2 , $KEMGN CPF - # &QMUWO Mathematical Statistics: Basic Ideas and Selected topics *QNFGP&C[ +PE 1CMNCPF %# =? 9 $[TPG 2 $G[GTNGKP , *WGTVC 5 -JWFCPRWT $ /CTVJK , /QTICP 0 2G VGTGM , 2KEQPG & 8GTI[TK CPF 9 9CPI 6QYCTFU NCPIWCIG KPFGRGPFGPV CEQWUVKE OQFGNKPI +P IEEE Conference on Acoustics, Speech, and Signal Processing RCIGU Ō +UVCPDWN 6WTMG[ =? 0 %JKPEJQT 2 4QDKPUQP CPF ' $TQYP *WD 0COGF 'PVKV[ 6CUM &GſPK VKQP 8GTUKQP +P Hub-5 Conversational Speech Recognition Workshop #XCKNCDNG CV YYYPKUVIQXURGGEJJWD =? ) 'XGTOCPP CPF 2 9QQFNCPF 2QUVGTKQT 2TQDCDKNKV[ &GEQFKPI %QPſFGPEG 'UVKOCVKQP CPF 5[UVGO %QODKPCVKQP +P In Proceedings of the NIST and NSA Speech Transcription Workshop %QNNGIG 2CTM /& =? , (KUEWU # 2QUVRTQEGUUKPI 5[UVGO VQ ;KGNF 4GFWEGF 9QTF 'TTQT 4CVGU 4GE QIPK\GT 1WVRWV 8QVKPI 'TTQT 4GFWEVKQP 418'4 +P IEEE Workshop on Automatic Speech Recognition and Understanding RCIGU Ō =? 4CFW (NQTKCP CPF &CXKF ;CTQYUM[ &[PCOKE 0QPNQECN .CPIWCIG /QFGNKPI XKC *KGTCTEJKECN 6QRKE$CUGF #FCRVCVKQP +P ACL99 RCIGU Ō =? , , )QFHTG[ ' % *QNNKOCP CPF , /E&CPKGN 5YKVEJDQCTF 6GNGRJQPG 5RGGEJ %QTRWU HQT 4GUGCTEJ CPF &GXGNQROGPV +P IEEE Conference on Acoustics, Speech, and Signal Processing XQNWOG RCIGU Ō 5CP (TCPEKUEQ %# =? 8 )QGN Word List With Content Word Marks JVVRYYYENURLJWGFWRGQRNGXIQGN
#XCKNCDNG CV
=? 8 )QGN Minimum Bayes-Risk Automatic Speech Recognition 2J& &KUUGT VCVKQP ,QJPU *QRMKPU 7PKXGTUKV[ $CNVKOQTG /& =? 8 )QGN CPF 9 $[TPG 6CUM &GRGPFGPV .QUU (WPEVKQPU KP 5RGGEJ 4GEQIPK VKQP 5GCTEJ QXGT 4GEQIPKVKQP .CVVKEGU +P Eurospeech-99 RCIGU Ō $WFCRGUV *WPICT[
E\&5&3UHVV//&
=? 8 )QGN CPF 9 $[TPG #RRNKECVKQPU QH /KPKOWO $C[GU4KUM &GEQFKPI VQ .8%54 +P In Proceedings of the NIST and NSA Speech Transcription Workshop %QNNGIG 2CTM /& =? 8 )QGN CPF 9 $[TPG /KPKOWO $C[GU4KUM #WVQOCVKE 5RGGEJ 4GEQIPKVKQP Computer Speech and Language Ō =? 8 )QGN CPF 9 $[TPG 4GEQIPK\GT 1WVRWV 8QVKPI CPF &/% KP /KPKOWO $C[GU4KUM (TCOGYQTM +P Research Notes No. 40, Center for Language and Speech Processing =? 8 )QGN 9 $[TPG CPF 5 -JWFCPRWT .8%54 4GUEQTKPI 9KVJ /QFKſGF .QUU (WPEVKQPU # &GEKUKQP 6JGQTGVKE 2GTURGEVKXG +P IEEE Conference on Acoustics, Speech, and Signal Processing XQNWOG RCIGU Ō =? 8 )QGN 5 -WOCT CPF 9 $[TPG 5GIOGPVCN /KPKOWO $C[GU4KUM #54 8QVKPI 5VTCVGIKGU +P International Conference on Spoken Language Processing XQNWOG RCIGU Ō $GKLKPI %JKPC =? 2 5 )QRCNCMTKUJPCP . 4 $CJN CPF 4 . /GTEGT # 6TGG 5GCTEJ 5VTCVGI[ HQT .CTIG 8QECDWNCT[ %QPVKPWQWU 5RGGEJ 4GEQIPKVKQP +P IEEE Conference on Acoustics, Speech, and Signal Processing XQNWOG RCIGU Ō =? 2 ' *CTV 0 , 0KNUUQP CPF $ 4CRJCGN # (QTOCN $CUKU HQT VJG *GWTKUVKE &GVGTOKPCVKQP QH /KPKOWO %QUV 2CVJU IEEE Transactions on Systems Science and Cybernetics 55% Ō =? 2 ' *CTV 0 , 0KNUUQP CPF $ 4CRJCGN %QTTGEVKQP VQ Ŏ# (QTOCN $CUKU HQT VJG *GWTKUVKE &GVGTOKPCVKQP QH OKPKOWO %QUV 2CVJUŏ SIGART Newsletter Ō =? ( ,GNKPGM # (CUV 5GSWGPVKCN &GEQFKPI #NIQTKVJO 7UKPI C 5VCEM IBM Journal of Research Development Ō =? ( ,GNKPGM Statistical Methods for Speech Recognition 6JG /+6 2TGUU %CO DTKFIG /CUUCEJWUGVVU =? Proceedings of the 1997 Large Vocabulary Continuous Speech Recognition Workshop #XCKNCDNG CV JVVRYYYENURLJWGFWYU =? $* ,WCPI CPF 5 -CVCIKTK &KUETKOKPCVKXG .GCTPKPI HQT /KPKOWO 'TTQT %NCU UKſECVKQP IEEE Transactions on Signal Processing 52 Ō =? , -CKUGT $ *QTXCV CPF < -CEKE # 0QXGN .QUU (WPEVKQP HQT VJG 1XGTCNN 4KUM %TKVGTKQP $CUGF &KUETKOKPCVKXG 6TCKPKPI QH *// /QFGNU +P International Conference on Spoken Language Processing XQNWOG RCIGU Ō $GK LKPI %JKPC =? 6 -CYCJCTC %* .GG CPF $* ,WCPI %QODKPKPI -G[ 2JTCUG &GVGEVKQP CPF 5WDYQTF $CUGF 8GTKſECVKQP HQT (NGZKDNG 5RGGEJ 7PFGTUVCPFKPI +P IEEE
E\&5&3UHVV//&
Conference on Acoustics, Speech, and Signal Processing XQNWOG RCIGU Ō =? / 9 -QQ %* .GG CPF $* ,WCPI # 0GY *[DTKF &GEQFKPI #NIQTKVJO HQT 5RGGEJ 4GEQIPKVKQP CPF 7VVGTCPEG 8GTKſECVKQP +P 1997 IEEE Workshop on Automatic Speech Recognition and Understanding RCIGU Ō =? 8 + .GXGPUJVGKP $KPCT[ %QFGU %CRCDNG QH %QTTGEVKPI &GNGVKQPU +PUGTVKQPU CPF 4GXGTUCNU Soviet Phys. Dokl. Ō =? . /CPIW ' $TKNN CPF # 5VQNEMG (KPFKPI %QPUGPUWU #OQPI 9QTFU .CVVKEG $CUGF 9QTF 'TTQT /KPKOK\CVKQP +P Eurospeech-99 RCIGU Ō $WFCRGUV *WPICT[ =? # /CTVKP , (KUEWU / 2T\[DQEMK CPF $ (KUJGT *WD 9QTMUJQR +P HQTOCVKQP 4GVTKGXCN +P 9th Hub-5 Conversational Speech Recognition Workshop =? # /CTVKP , (KUEWU / 2T\[DQEMK CPF $ (KUJGT *WD 9QTMUJQR 9GKIJVGF 9QTF 4GUWNVU +P 9th Hub-5 Conversational Speech Recognition Workshop =? - 0C $ ,GQP & %JCPI 5 %JCG CPF 5 #PP &KUETKOKPCVKXG 6TCKPKPI QH *KFFGP /CTMQX /QFGNU 7UKPI 1XGTCNN 4KUM %TKVGTKQP CPF 4GFWEGF )TCFKGPV /GVJQF +P Eurospeech-95 RCIGU Ō /CFTKF 5RCKP =? # 0CFCU # &GEKUKQP 6JGQTGVKE (QTOWNCVKQP QH VJG 6TCKPKPI 2TQDNGO KP 5RGGEJ 4GEQIPKVKQP CPF C %QORCTKUQP QH 6TCKPKPI D[ 7PEQPFKVKQPCN 8GTUWU %QPFKVKQPCN /CZKOWO .KMGNKJQQF IEEE Transactions on Acoustics, Speech, and Signal Processing #552 Ō =? # 0CFCU 1RVKOCN 5QNWVKQP QH C 6TCKPKPI 2TQDNGO KP 5RGGEJ 4GEQIPK VKQP IEEE Transactions on Acoustics, Speech, and Signal Processing #552 Ō =? & $ 2CWN #P 'HſEKGPV 5VCEM &GEQFGT #NIQTKVJO HQT %QPVKPWQWU 5RGGEJ 4GEQIPKVKQP YKVJ C 5VQEJCUVKE .CPIWCIG /QFGN +P IEEE Conference on Acoustics, Speech, and Signal Processing XQNWOG RCIGU Ō =? ' 4KUVCF CPF 2 ;KCPKNQU .GCTPKPI 5VTKPI 'FKV &KUVCPEG IEEE Trans. PAMI Ō =? 4 % 4QUG CPF & $ 2CWN # *KFFGP /CTMQX /QFGN $CUGF -G[YQTF 4GEQIPK VKQP 5[UVGO +P IEEE Conference on Acoustics, Speech, and Signal Processing XQNWOG RCIGU Ō =? # 5VQNEMG ; -QPKI CPF / 9GKPVTCWD 'ZRNKEKV 9QTF 'TTQT /KPKOK\CVKQP KP 0$GUV .KUV 4GUEQTKPI +P Eurospeech-97 XQNWOG RCIGU Ō 4JQFGU )TGGEG =? 8 8CRPKM Estimation of Dependences Based on Empirical Data 5RTKPIGT 8GTNCI 0GY ;QTM
E\&5&3UHVV//&
=? ( 9GUUGN 4 5EJNWVGT CPF * 0G[ 7UKPI 2QUVGTKQT 9QTF 2TQDCDKNKVKGU (QT +ORTQXGF 5RGGEJ 4GEQIPKVKQP +P IEEE Conference on Acoustics, Speech, and Signal Processing XQNWOG RCIGU Ō +UVCPDWN 6WTMG[ =? , ) 9KNRQP . 4 4CDKPGT %* .GG CPF ' 4 )QNFOCP #WVQOCVKE 4GEQI PKVKQP QH -G[YQTFU KP 7PEQPUVTCKPGF 5RGGEJ 7UKPI *KFFGP /CTMQX /QF GNU IEEE Transactions on Acoustics, Speech, and Signal Processing #552 Ō =? 5 ;QWPI HTK 2.1 'PVTQRKE %CODTKFIG 4GUGCTEJ .CDQTCVQT[ .VF %CO DTKFIG 7-
E\&5&3UHVV//&
3 A Decision Theoretic Formulation for Robust Automatic Speech Recognition Qiang Huo The University of Hong Kong, Hong Kong, China
CONTENTS
+PVTQFWEVKQP 1RVKOCN $C[GUŏ &GEKUKQP 4WNG HQT #54 #FCRVKXG &GEKUKQP 4WNGU %QPUVTWEVGF HTQO 6TCKPKPI 5CORNGU 8KQNCVKQPU QH /QFGNKPI #UUWORVKQPU KP #54 +ORTQXKPI #FCRVKXG &GEKUKQP 4WNGU XKC &GEKUKQP 2CTCOGVGT #FCRVCVKQP 4QDWUV &GEKUKQP 4WNGU 5WOOCT[ #EMPQYNGFIGOGPV 4GHGTGPEGU
3.1 Introduction /QFGTP CWVQOCVKE URGGEJ TGEQIPKVKQP #54 VGEJPQNQI[ = ? KU DCUGF QP C EQOOWPKECVKQP VJGQTGVKE XKGY QH VJG IGPGTCVKQP CESWKUKVKQP CPF VTCPUOKUUKQP CPF RGTEGRVKQP QH URGGEJ =? (KIWTG CFCRVGF HTQO ,WCPIŏU MG[PQVG URGGEJ KP 0052ŏ =? UJQYU C EQPEGRVWCN OQFGN QH C PQKU[ EJCPPGN HQT URGGEJ IGPGTCVKQP CPF UKIPCN ECRVWTKPI 6JG IQCN QH URGGEJ TGEQIPKVKQP KU VJGP FGſPGF CU TGEQXGTKPI VJG YQTF UGSWGPEG HTQO VJG CEQWUVKE UKIPCN 6JKU ECP CNUQ DG XKGYGF CU C decision problem KG DCUGF QP VJG KPHQTOCVKQP KP CPF VJG QVJGT TGNGXCPV CURGEVU QH VJG RTQDNGO YG CVVGORV VQ OCMG VJG DGUV FGEKUKQP KP UQOG VJCV JCU DGGP GODGFFGF KP (QT VJG UKORNKEKV[ QH FKUEWUUKQP YG UGPUG QH VJG ECP XKGY GCEJ RQUUKDNG YQTF UGSWGPEG CU C class .GV WU CUUWOG VJGTG CTG VQVCN WPKSWG ENCUUGU 5Q URGGEJ TGEQIPKVKQP EQPUKUVU KP ſPFKPI QRVKOCN KP UQOG UGPUG FGEKUKQP TWNGU HQT ENCUUKſECVKQP QH VJG QDUGTXCVKQP KPVQ QPG QH ſZGF ENCUUGU &GRGPFKPI QP FKHHGTGPV ETKVGTKC VJGTG GZKUV OCP[ FGEKUKQP TWNGU 0QV CNN QH VJGO CTG QH GSWCN XCNWG KP RTCEVKEG $GECWUG QH VJG FKHHGTGPV UQWTEGU QH XCTKCDKNKV[ CU UJQYP KU WUWCNN[ HGCVWTGF D[ WPEGTVCKPV[ XCTKCDKNKV[ KP (KIWTG VJG URGGEJ UKIPCN NCEM QH FGVGTOKPKUO CPF UVQEJCUVKEKV[ 6JKU OCMGU VJG statistical pattern recognition CRRTQCEJ = ? C PCVWTCN EJQKEG HQT HQTOWNCVKPI CPF UQNXKPI VJG
Ï
Ï
Ï
E\&5&3UHVV//&
Å
Å
FIGURE 3.1 Communication Theoretic View of ASR: Noisy Channel for Speech Generation and Signal Capturing (adapted from [68]). #54 RTQDNGO CU FGUETKDGF DTKGƀ[ KP VJG HQNNQYKPI (KTUV VJG UVCVKUVKECN OQFGNU HQT VJG EJCPPGNU KP (KIWTG CTG UKORNKſGF CU HQNNQYU
# YQTF UGSWGPEG CPF VJG CUUQEKCVGF CEQWUVKE QDUGTXCVKQP CTG XKGYGF CU C LQKPVN[ FKUVTKDWVGF TCPFQO RCKT (QT PQVCVKQPCN UKORNKEKV[ YG YKNN WUG VJG UCOG U[ODQN VQ FGPQVG DQVJ VJG TCPFQO XCTKCDNG CPF VJG XCNWG KV OC[ CUUWOG
6JG LQKPV FKUVTKDWVKQP QH KU OQFGNGF D[ C parametric family QH 2&(
RTQDCDKNKV[ FGPUKV[ HWPEVKQP KU MPQYP CU VJG CEQWUVKE OQFGN YKVJ RCTCOGVGTU CPF CU VJG NCPIWCIG OQFGN YKVJ RCTCOGVGTU
6JG RCTCOGVGTU QH VJG CDQXG FKUVTKDWVKQPU CTG VQ DG GUVKOCVGF HTQO UQOG training data D[ WUKPI RCTVKEWNCT RCTCOGVGT GUVKOCVKQP VGEJPKSWGU 9KVJ VJGUG UKORNKſECVKQPU VJG OQUV RQRWNCT YC[ VQ UQNXG VJG #54 RTQDNGO KU VQ WUG VJG YGNNMPQYP plug-in MAP OCZKOWO a posteriori decision rule = ?
Ï Ï
CPF CTG VJG GUVKOCVGF RCTCOGVGTU QDVCKPGF FWTKPI VTCKPKPI CPF KU VJG YJGTG TGEQIPK\GF UGPVGPEG FWTKPI VGUVKPI 6JKU FGEKUKQP TWNG FGTKXGF HTQO VJG QRVKOCN $C[GUŏ FGEKUKQP TWNG KU CNUQ YKFGN[ WUGF KP OCP[ QVJGT RCVVGTP TGEQIPKVKQP CRRNKEC VKQPU
E\&5&3UHVV//&
6JKU EJCRVGT CVVGORVU VQ GZRNCKP HTQO C statistical decision RQKPV QH XKGY YJ[ VJG CDQXG CRRTQCEJ YQTMU UQ YGNN KP EGTVCKP EQPFKVKQPU CPF OQTG KORQTVCPVN[ YJ[ KV FQGU PQV YQTM KP OCP[ QVJGT UKVWCVKQPU 6Q FQ VJKU 5GEVKQP ſTUV GZRNCKPU VJG FGEKUKQP VJGQTGVKE HQTOWNCVKQP QH VJG #54 RTQDNGO CPF VJG QRVKOCN FGEKUKQP TWNG VJCV ECP DG EQPUVTWEVGF KH GXGT[VJKPI CDQWV VJG RTQDNGO KU MPQYP 6JGP 5GEVKQP GZRNCKPU JQY VQ EQPUVTWEV VJG CFCRVKXG FGEKUKQP TWNGU YJGP NGCTPKPI HTQO C VTCKP KPI UCORNG UGV 6JG TCVKQPCNG QH VYQ RQRWNCT FGUKIP RTKPEKRNGU KP EQPUVTWEVKPI UWEJ CFCRVKXG FGEKUKQP TWNGU KU CNUQ FKUEWUUGF 5GEVKQP FKUEWUUGU VJG ENCUUKſECVKQP QH RQUUKDNG FKUVQTVKQPU QH J[RQVJGVKECN OQFGNU CPF FCVC CPF VJG RQUUKDNG YC[U QH CEJKGX KPI RGTHQTOCPEG TQDWUVPGUU 5GEVKQP TGXKGYU CPF FKUEWUUGU UQOG QH VJG TGEGPV RCTCOGVGT CFCRVCVKQP VGEJPKSWGU HQT KORTQXKPI CFCRVKXG FGEKUKQP TWNGU 5GEVKQP GZRNCKPU VJG DCUKE PQVCVKQP QH VJG FGEKUKQP TWNG TQDWUVPGUU CPF UJQYU VYQ GZCORNGU QH JQY VQ EQPUVTWEV TQDWUV FGEKUKQP TWNGU PCOGN[ VJG OKPKOCZ FGEKUKQP TWNG CPF VJG $C[GUKCP RTGFKEVKXG ENCUUKſECVKQP TWNG 5GEVKQP UWOOCTK\GU VJG KFGCU FKUEWUUGF KP VJG EJCRVGT
3.2 Optimal Bayes’ Decision Rule for ASR +P KVU UKORNGUV HQTO NGV WU CUUWOG VJCV QWT #54 RTQDNGO KU VQ ENCUUKH[ C URGGEJ QD UGTXCVKQP KP RTCEVKEG WUWCNN[ C HGCVWTG XGEVQT UGSWGPEG GZVTCEVGF HTQO VJG URGGEJ ENCUUGU Ï YJGTG Ï FG UKIPCN KPVQ QPG QH ENCUUGU &GRGPFKPI QP VJG RTQDNGO QH KPVGTGUV C ENCUU PQVGU VJG UGV QH OC[ DG QH CP[ NKPIWKUVKE WPKV GI C RJQPGOG C U[NNCDNG C YQTF C RJTCUG C UGP VGPEG C UGOCPVKE EQPEGRV QT CVVTKDWVG GVE .GV WU CUUWOG VJCV VJG URGGEJ QDUGTXCVKQP DGNQPIU VQ C UWKVCDNG URCEG 6JG RTQDNGO QH EQPUVTWEVKPI C URGGEJ TGEQIPK\GT KU VJGP GSWKXCNGPV VQ ſPFKPI C decision rule KP C UGV QH RQUUKDNG FGEKUKQP TWNGU UWEJ VJCV QT UKORN[
HQT CPF
YKVJ DGKPI QPG QH VJG RQUUKDNG ENCUU NCDGNU KP +P VJKU ECUG VJG decision space QH VJG FGEKUKQP TWNG KU VJG UCOG CU VJG # FGEKUKQP TWNG KORNKGU C OCRRKPI HTQO VJG UCORNG URCEG VQ VJG ENCUU NCDGN URCEG 6JKU OCRRKPI KU MPQYP CU C nonrandomized decision rule =? &GſPG VQ DG C UWDUGV QH EQTTG URQPFKPI VQ VJG TGIKQP QH DGKPI OCRRGF CU ENCUU YKVJ VJG FGEKUKQP TWNG VJGP VJG EQPUVTWEVKQP QH C FGEKUKQP TWNG COQWPVU VQ ſPFKPI C RCTVKVKQP QH VJG QDUGTXCVKQP URCEG WPFGT VJG HQNNQY KPI EQPUVTCKPVU
E\&5&3UHVV//&
6JGTG OC[ GZKUV CP KPſPKVG UGV QH FGEKUKQP TWNGU HQT VJG UCOG IKXGP ENCUUKſECVKQP RTQD NGO 0QV CNN QH VJGO CTG QH GSWCN XCNWG KP RTCEVKEG VJQWIJ 6Q FGVGTOKPG YJGVJGT C FGEKUKQP TWNG KU őIQQFŒ QPG JCU VQ CITGG QP C TGCUQPCDNG UGV QH ETKVGTKC HQT CUUGUUKPI VJG őIQQFPGUUŒ .GV WU UJQY QPG RQUUKDNG HQTOWNCVKQP D[ WUKPI VJG ENCUUKECN UVCVKUVK ECN FGEKUKQP VJGQT[ RKQPGGTGF D[ 9CNF CPF FGXGNQRGF D[ OCP[ QVJGTU = ? CPF CP QDUGTXCVKQP CU C LQKPVN[ FKUVTKDWVGF TCPFQO RCKT .GV WU XKGY YJQUG LQKPV 2&( KU FGPQVGF D[ +P VJG UQECNNGF sampling paradigm YG ECP FGEQORQUG KPVQ C RTQFWEV QH VJG ENCUU RTKQT RTQDCDKNKV[ CPF VJG ENCUU 1PG YC[ QH HQTOCNK\KPI EQPFKVKQPCN 2&( KG C IQQFPGUU ETKVGTKQP KU VQ WUG VJG MPQYNGFIG QH VJG RQUUKDNG EQPUGSWGPEGU QH VJG FGEKUKQPU 1HVGP VJKU MPQYNGFIG ECP DG SWCPVKſGF D[ CUUKIPKPI C loss VJCV YQWNF DG DG VJG loss function CUUQEKCVGF KPEWTTGF HQT GCEJ RQUUKDNG FGEKUKQP .GV YKVJ OCMKPI C FGEKUKQP KH VJG VTWG ENCUU KU 1PG YQWNF NKMG VJG NQUU HWPEVKQP VQ JCXG VJG HQNNQYKPI RTQRGTV[
+H YG CUUWOG VJG true distribution KU MPQYP VJGP VJG EQPFKVKQPCN CPF OCTIKPCN FKUVTKDWVKQPU PCOGN[ CPF ECP DG ECNEWNCVGF 0QY YG ECP FGſPG VJG total risk HQT C FGEKUKQP TWNG CU CP GZRGEVGF XCNWG QH
VJG NQUU HWPEVKQP KG
Ï
´
Ü Ï
µ
ªÏ ª
ª
ªÏ
Ü
ª
ªÜ
YJGTG ´µ FGPQVGU OCVJGOCVKECN GZRGEVCVKQP YKVJ TGURGEV VQ VJG FKUVTKDWVKQP QH 6JG CDQXG VQVCN TKUM ECP DG WUGF CU C OGCUWTG QH VJG SWCNKV[ QH FGEKUKQP TWNGU 7UWCNN[ VJG NGUU VJG VQVCN TKUM VJG DGVVGT KU VJG FGEKUKQP TWNG +P VJKU HTCOG YQTM VJG KUUWG QH EQPUVTWEVKPI CP QRVKOCN FGEKUKQP TWNG DGEQOGU VJG HQNNQYKPI TKUM OKPKOK\CVKQP RTQDNGO
´
µ
´
µ
ªÜ
ªÏ
6JKU QRVKOK\CVKQP ECP DG UQNXGF D[ OKPKOK\KPI VJG GZRTGUUKQP KP VJG USWCTG DTCEMGVU KP VJG CDQXG GSWCVKQP +V KU ENGCT VJCV VJG UQNWVKQP NGCFU VQ VJG HQNNQYKPI QRVKOCN FGEKUKQP TWNG
E\&5&3UHVV//&
´
µªÏ ªÏ
YJKEJ KU CNUQ MPQYP CU VJG Bayes’ decision rule 6JG TGUWNVKPI OKPKOWO VQVCN TKUM
Ó
Ü
Ï Ï
Ó
KU ECNNGF VJG Bayes’ risk 6JKU TKUM XCNWG KU VJG DGUV VJCV ECP DG CEJKGXGF KH VJG FKUVTKDWVKQP KU MPQYP +P URGGEJ TGEQIPKVKQP C TGCUQPCDNG QRVKQP KU VQ CUUWOG VJCV GXGT[ OKUENCUUKſECVKQP QH KU GSWCNN[ UGTKQWU VJGTGD[ TGUWNVKPI KP VJG UQECNNGF 0-1 loss function
HQT
Ï
Ï
KH KH
EQTTGEV FGEKUKQP
YTQPI FGEKUKQP
5WDUVKVWVKPI KPVQ YG QDVCKP
Ï Ï
Ï Ï
Ü Ï
Ü Ï
6JGTGHQTG KP VJG ECUG QH VJG NQUU HWPEVKQP VJG VQVCN TKUM KU VJG WPEQPFKVKQPCN GTTQT RTQDCDKNKV[ YJKEJ KU CRRCTGPVN[ C IQQF OGCUWTG QH VJG SWCNKV[ QH FGEKUKQP TWNGU HQT VJG #54 VCUM 6JG QRVKOCN FGEKUKQP TWNG WPFGT VJG minimum classification
UWEJ VJCV error ETKVGTKQP YKVJ VJG NQUU HWPEVKQP KU VJGP UQNXGF CU
Ï
Ï
YJKEJ KU CNUQ MPQYP CU VJG MAP decision rule +P UWOOCT[ KP EQPUVTWEVKPI VJGUG QRVKOCN FGEKUKQP TWNGU KV YCU CUUWOGF VJCV EQO RNGVG RTKQT KPHQTOCVKQP CDQWV VJG ENCUUGU KU MPQYP KG VJG QDUGTXCVKQP URCEG Ü KU IKXGP
KU IKXGP CPF VJG VTWG 2&( QT CPF CTG MPQYP
VJG NQUU HWPEVKQP
7PFGT VJGUG CUUWORVKQPU VJG QRVKOCNKV[ ETKVGTKQP KU VJG OKPKOK\CVKQP QH VJG TKUM HWPEVKQPCN CPF VJG QRVKOCN FGEKUKQP TWNG KU VJG $C[GUŏ FGEKUKQP TWNG
3.3 Adaptive Decision Rules Constructed from Training Samples
+P RTCEVKEG YG MPQY PGKVJGT VJG true RCTCOGVTKE HQTO QH VJG LQKPV FKUVTKDWVKQP PQT KVU true RCTCOGVGTU 9G UJCNN UC[ VJCV YG JCXG prior uncertainty =? KP VJKU ECUG
E\&5&3UHVV//&
+H YG JCXG UQOG NCDGNGF independent VTCKPKPI UCORNG UGV QDVCKPGF D[ C UGTKGU QH independent GZRGTKOGPVU UWEJ VJCV HQT VJG #54 VCUM CV JCPF QT KP OKPF YG ECP TGFWEG VJG RTKQT WPEGTVCKPV[ D[ EQPUVTWEVKPI C FGEKUKQP TWNG HTQO 6JG FGEKUKQP TWNG DCUGF QP VJG VTCKPKPI UGV CPF WUGF VQ ENCUUKH[ C TCPFQO QDUGTXCVKQP X VJCV KU independent QH KU ECNNGF CP adaptive decision rule =? 6JGTG CTG UGXGTCN RTKPEKRNGU VJCV ECP DG WUGF HQT VJG EQPUVTWEVKQP QH UWEJ TWNGU6YQ QH VJGO CTG DTKGƀ[ FKUEWUUGF KP VJG HQNNQYKPI
3.3.1 Plug-in Bayes’ Decision Rules with Maximum-likelihood Density Estimate 3.3.1.1 What are Plug-in Bayes’ Decision Rules? 6JG OQUV RQRWNCT HCOKN[ QH CFCRVKXG FGEKUKQP TWNGU OKIJV DG VJG UQECNNGF plug-in decision rules (QT VJKU CRRTQCEJ NGV DG CP[ UVCVKUVKECN GUVKOCVQTU QH VTWG FKUVTKDWVKQPU DCUGF QP VJG VTCKPKPI UCORNG 6JG plug-in decision rule =? KU VJG CFCRVKXG FGEKUKQP TWNG FGTKXGF HTQO VJG $C[GUKCP FGEKUKQP TWNG D[ UWDUVKVWVKQP QH VJG GUVKOCVQTU HQT WPMPQYP VTWG FKUVTKDWVKQPU
Ï
YJGTG
´
µ ª
ªÏ
$[ XCT[KPI VJG NQUU HWPEVKQP CPF D[ WUKPI VJG FKHHGTGPV MKPFU QH GUVKOCVQTU C HCKTN[ TKEJ HCOKN[ QH RNWIKP FGEKUKQP TWNGU ECP DG QDVCKPGF (QT GZCORNG CFQRVKPI VJG NQUU HWPEVKQP YKNN NGCF VQ VJG HQNNQYKPI RNWIKP FGEKUKQP TWNG UWEJ VJCV
YJKEJ KU CNUQ MPQYP CU VJG plug-in MAP decision rule +V ECP DG UJQYP =? VJCV VJG RNWIKP FGEKUKQP TWNG KP 'S OKPKOK\GU VJG plug-in risk
ªÏ
ªÜ
YJKEJ KU CP GUVKOCVG QH VJG VQVCN TKUM WUKPI VJG density plug-in estimator KG
6JG OKPKOWO RNWIKP TKUM KU VJGP
´ µ
E\&5&3UHVV//&
3.3.1.2 Why Could Plug-in Bayes’ Decision Rules Work? #U PQVGF KP =? VJG RNWIKP TKUM QH VJG RNWIKP $C[GUŏ FGEKUKQP TWNG KP 'S
KU QHVGP NGUU VJCP KVU VQVCN TKUM CPF KU GXGP QRVKOKUVKECNN[ DKCUGF CU CP GUVKOCVQT QH VJG $C[GUŏ TKUM Property: +H VJG GUVKOCVQTU
CTG RQKPVYKUG WPDKCUGF VJGP
*QYGXGT VJG WUGHWNPGUU QH VJG RNWIKP $C[GUŏ FGEKUKQP TWNG KP 'S ECP DG LWUVKſGF D[ VJG HQNNQYKPI VJGQTGO QH Bayes’ risk consistency =? Theorem: (Bayes’ risk consistency): +H VJG GUVKOCVQTU CTG UVTQPIN[ EQPUKUVGPV KG EQPXGTIG VQ VJG VTWG FKUVTKDWVKQPU CNOQUV UWTGN[ CU VJG VTCKPKPI UCORNG UK\G KPETGCUGU
CPF
VJGP VJG RNWIKP TKUM HQT VJG RNWIKP FGEKUKQP TWNG KP 'S KU C UVTQPIN[ EQPUKUVGPV GUVKOCVQT QH VJG $C[GUŏ TKUM KG
HQT
3.3.1.3 Implications on Parametric Models and Parameter Estimation +P RTCEVKEG DGECWUG QH VJG EQPUVTCKPVU QH VJG NKOKVGF EQORWVCVKQPCN TGUQWTEGU CPF VTCKPKPI FCVC YG CNYC[U JCXG VQ assume UQOG RCTCOGVTKE HQTO HQT GI XKC CPF 6JG RCTCOGVGT UGV JCU VQ DG estimated HTQO VJG IKXGP VTCKPKPI UGV D[ WUKPI EGTVCKP RCTCOGVGT GUVKOCVKQP VGEJPKSWGU 6JG CDQXG $C[GUŏ TKUM EQPUKUVGPE[ VJGQTGO VGNNU WU VJCV KV KU QHVGP RQUUKDNG VQ EQPUVTWEV RNWIKP RTQEGFWTGU VJCV CTG Bayes’ risk consistent KP VJG UGPUG VJCV VJG UGSWGPEG QH RNWIKP TKUMU EQPXGTIGU VQ VJG $C[GUŏ TKUM CU VJG VTCKPKPI UGVU KPETGCUG KP UK\G *QYGXGT VJGTG KU CP KORQTVCPV CUUWORVKQP DGJKPF VJKU CTIWOGPV VJCV KU VJG CUUWOGF FKUVTKDWVKQPU CPF QDG[ VJG RCTCOGVTKE UVTWEVWTG KP SWGUVKQP +P QTFGT VQ CEJKGXG C IQQF CRRTQZKOCVKQP VQ TGCNKV[ UQOG ƀ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representative VTCKPKPI URGGEJ CPF VGZV FCVC IQQF UVCVKUVKECN OQFGNU QH URGGEJ CPF NCPIWCIG ECP DG EQPUVTWEVGF VQ CEJKGXG C JKIJ RGTHQTOCPEG HQT C YKFG TCPIG QH #54 VCUMU 6JKU JCU IKXGP VJG URGGEJ TGUGCTEJ EQOOWPKV[ C EGTVCKP NGXGN QH EQPſFGPEG KP
E\&5&3UHVV//&
DGNKGXKPI VJCV VJG Discrete HMM &*// =? CPF VJG /KZVWTG )CWUUKCP Continuous Density HMM %&*// = ? VQIGVJGT YKVJ ITCO OQFGNU =? RTQXKFG C IQQF CRRTQZKOCVG RCTCOGVTKE HQTO HQT CPF TGURGE VKXGN[ #NVJQWIJ VJGUG OQFGNU CTG CRRCTGPVN[ KORGTHGEV = ? VJG[ CTG OCVJGOCVKECNN[ YGNNFGſPGF CPF ECRCDNG QH UKOWNVCPGQWUN[ OQFGNKPI DQVJ VJG URGE VTCN CPF VGORQTCN XCTKCVKQP KP URGGEJ 6JG[ CTG CNUQ YGNN VJQWIJV QH DGECWUG VJG[ DQVJ ſV KPVQ VJG HTCOGYQTM QH finite state TGRTGUGPVCVKQPU = ? QH knowledge sources UQ VJCV VJG URGGEJ TGEQIPKVKQP RTQDNGO ECP DG UQNXGF CU C network search RTQDNGO QXGT C EQORNGZ PGVYQTM TGRTGUGPVCVKQP QH URGGEJ CPF NCPIWCIG =? $CUGF QP VJG DG NKGH VJCV VJGUG CEQWUVKE CPF NCPIWCIG OQFGNU CTG IQQF CRRTQZKOCVGU VJG maximum likelihood /. GUVKOCVG HQT VJG *// RCTCOGVGTU = ? CPF ITCO OQFGN RCTCOGVGTU = ? JCU DGGP VJG OQUV RQRWNCT RCTCOGVGT GUVKOCVKQP OGVJQF 6JG YKFGURTGCF WUG QH VJG RNWIKP /#2 FGEKUKQP TWNG YKVJ VJG /. GUVKOC VQT ECP DG LWUVKſGF D[ WUKPI VJG CDQXG $C[GUŏ TKUM EQPUKUVGPE[ VJGQTGO FWG VQ VJG HQNNQYKPI HCEVU
6JG /. GUVKOCVQTU QH
CTG UVTQPIN[ EQPUKUVGPV WPDKCUGF CPF GHſEKGPV
6JKU ECP VJGP DG VTCPUNCVGF KPVQ VJG FKUVTKDWVKQP EQPUKUVGPE[ KH VJG RCTCOGVTKE HQTOU QH VJG CPF CTG KPFGGF EQTTGEV
#EEQTFKPI VQ QWT MPQYNGFIG KV YCU 0CFCU =? YJQ ſTUV RTQXKFGF UWEJ CP KPUKIJV HQT VJG URGGEJ TGEQIPKVKQP EQOOWPKV[ CPF OC[ DG 1H EQWTUG QPG ECP CNYC[U CTIWG VJCV CNVJQWIJ VJG /. GUVKOCVQTU GZEGNNGPV GUVKOCVQTU QH CPF VJGTG KU PQ IWCTCPVGG VJCV CPF CTG IQQF IWGUUGU HQT CPF DGECWUG QH VJG KPEQTTGEV OQFGN CUUWORVKQPU 0QT KU PGEGUUCTKN[ C IQQF CRRTQZKOCVKQP VQ 6JG RGTHQTOCPEG QH VJG RNWI KP TWNGU CPF QVJGT RTQEGFWTGU UJQWNF TGCNN[ DG LWFIGF D[ VJG ETKVGTKQP QH VQVCN TKUM QT D[ QVJGT ETKVGTKC VKGF OQTG FKTGEVN[ VQ VJG ENCUUKſECVKQP CEEWTCE[ VJCP VQ VJG DGJCXKQT CU C point estimator HQT 6JKU JCU OQVKXCVGF OCP[ UVWFKGU KP VJG RCUV QH VYQ FGECFGU CKOKPI CV C IQQF CNVGTPCVKXG VQ /. VTCKPKPI 1PG OGVJQF KU minimum discrimination information /&+ VTCKPKPI =? YJKEJ CFLWUVU VJG *// RCTCOGVGTU VQ OKPKOK\G VJG discrimination information QT directed divergence DGVYGGP VJG CU UWOGF *// FKUVTKDWVKQP CPF VJG DGUV RQUUKDNG FKUVTKDWVKQP FGTKXGF HTQO VJG VTCKPKPI FCVC WPFGT EGTVCKP EQPUVTCKPVU GODGFFGF KP VJG VTCKPKPI FCVC 7PHQTVWPCVGN[ PQ UKI PKſECPV GZRGTKOGPVCN TGUWNVU JCXG DGGP TGRQTVGF VQ UJQY JQY /&+ YQTMU KP C URGGEJ TGEQIPKVKQP VCUM #PQVJGT ENCUU QH CRRTQCEJGU KU VJG UQECNNGF discriminative training OGVJQF 5QOG QH VJGO UWEJ CU maximum mutual information //+ VTCKPKPI =? conditional maximum likelihood estimate %/.' =? CPF H-criteria =? CKO KP FKTGEVN[ CV TGFWEKPI VJG GTTQT TCVG QH VJG URGGEJ TGEQIPK\GT QP VJG VTCKPKPI UGV 1VJGT OGVJQFU UWEJ CU corrective training =? CPF minimum empirical classification error VTCKPKPI = ? VT[ VQ TGFWEG VJG TGEQIPKVKQP GTTQT TCVG QP VTCKPKPI UCORNG UGV KP C OQTG FKTGEV YC[ #OQPI VJGUG CRRTQCEJGU VJG OKPKOWO GORKTKECN ENCUUKſ ECVKQP GTTQT MPQYP CU /%' HQTOWNCVKQP RTQRQUGF KP =? KU KP O[ QRKPKQP OQTG VJGQTGVKECNN[ UQWPF VJWU YKNN DG FKUEWUUGF DTKGƀ[ KP VJG HQNNQYKPI
E\&5&3UHVV//&
3.3.2 Maximum-Discriminant Decision Rules Minimizing the Empirical Classification Error 3.3.2.1 What are Maximum-Discriminant Decision Rules?
5WRRQUG QPG ECP FGſPG C discriminant function £ HQT GCEJ ENCUU VJCV EJCTCEVGTK\GU VJG UKOKNCTKV[ DGVYGGP CP QDUGTXCVKQP CPF VJG ENCUU YJGTG KU VJG UGV QH ENCUUKſGT RCTCOGVGTU VQ DG GUVKOCVGF HTQO VJG VTCKPKPI FCVC UGV 0CVWTCNN[ VJG HQNNQYKPI maximum-discriminant decision rule
ECP DG WUGF VQ ENCUUKH[ CP WPMPQYP QDUGTXCVKQP KPVQ QPG QH VJG ENCUUGU KP Ï
£
Ï
6JG QDXKQWU ETKVGTKQP HQT GUVKOCVKPI VJG ENCUUKſGT RCTCOGVGTU KU VQ OKPKOK\G VJG GORKTKECN ENCUUKſECVKQP GTTQT QP VJG VTCKPKPI UCORNG UGV FGſPGF CU HQNNQYU
PWODGT QH EQTTGEV ENCUUKſECVKQPU D[ VQVCN PWODGT QH UCORNG QDUGTXCVKQPU QP
0QY NGV FGPQVG CP CTDKVTCT[ DWV EQORNGVGN[ URGEKſGF EQNNGEVKQP QH FKUETKOKPCPV DCUGF FGEKUKQP TWNGU # UCORNGDCUGF FKUETKOKPCPV FGEKUKQP TWNG YKNN DG ECNNGF C minimum misclassification QT best-count FKUETKOKPCPV FGEKUKQP TWNG KH KV OKP KOK\GU VJG UCORNG GTTQT TCVG COQPI CNN FKUETKOKPCPV FGEKUKQP TWNGU VJCV KU C DGUVEQWPV FKUETKOKPCPV FGEKUKQP TWNG UCVKUſGU
´µ
5KOKNCT VQ VJG ECUG QH VJG density estimator KV ECP DG UJQYP =? VJCV
´µ
¼½
¼½
5Q KU CP QRVKOKUVKECNN[ DKCUGF GUVKOCVQT QH VJG CEVWCN GTTQT TCVG QH CPF VJG NGCUV RQUUKDNG GTTQT TCVG 3.3.2.2 Why Could Discriminant Approach Work? 6JG WUGHWNPGUU QH VJG DGUVEQWPV FKUETKOKPCPV CRRTQCEJ ECP DG LWUVKſGF D[ VJG HQN NQYKPI VJGQTGO UKOKNCT VQ VJG QPG KP UGEVKQP CPF CNUQ RTQXGF D[ )NKEM =? Theorem: (Uniform Convergence) # FKUETKOKPCPV FGEKUKQP TWNG YKNN DG ECNNGF mconvex KH KVU RCTVKVKQP TGIKQPU CTG UGVU KP VJG ſPKVG ſGNF IGPGTCVGF D[ UQOG OGCUWTCDNG EQPXGZ UGVU #U VJG UCORNG UK\G VJG GUVKOCVQT EQPXGTIGU VQ ¼½ uniformly QXGT CNN FKUETKOKPCPV FGEKUKQP TWNGU KP CP[ EQNNGEVKQP QH EQPXGZ FKUETKOKPCPV FGEKUKQP TWNGU VJCV KU VJG EQP XGTIGPEG KU CNOQUV UWTGN[ CU
´µ
E\&5&3UHVV//&
¼½
6JKU WPKHQTO EQPXGTIGPEG KORNKGU VJCV VJG best-count FKUETKOKPCPV CU[ORVQVKECNN[ QRVKOCN KP VJG UGPUG QH
¼½
KU
¼½ YKVJ RTQDCDKNKV[ QPG CPF
´µ
¼½ YKVJ RTQDCDKNKV[ QPG
´µ
+H EQNNGEVKQP EQPVCKPU CP[ QRVKOCN FKUETKOKPCPV FGEKUKQP TWNG
¼½ ´µ
VJGP KU CU[ORVQVKECNN[ QRVKOCN KP VJG WPTGUVTKEVGF UGPUG XK\ UVTQPIN[ EQPUKU VGPV KP $C[GUŏ TKUM #U RQKPVGF QWV KP =? VJKU TGUWNV KU PCTTQYGT VJCP KVU RCTCNNGN TGUWNV HQT FGPUKV[ GUVKOCVGU UVCVGF KP VJG VJGQTGO KP 5GEVKQP +V YKNN DG KPVGT GUVKPI VQ KPXGUVKICVG JQY HCT VJG CDQXG TGUWNV ECP DG IGPGTCNK\GF VQ C YKFGT TCPIG QH FKUETKOKPCPV HWPEVKQPU 3.3.2.3 Implications on the Choice of Discriminant Functions and the Practical Training Algorithms 6JG CDQXG VJGQTGVKECN TGUWNV IKXGU QPG EQPſFGPEG VJCV KH C RTQRGT HQTO HQT VJG FKU ETKOKPCPV HWPEVKQPU ECP DG URGEKſGF HQT VJG IKXGP RCVVGTP TGEQIPKVKQP RTQDNGO KV KU QHVGP RQUUKDNG VQ EQPUVTWEV OCZKOWOFKUETKOKPCPV FGEKUKQP TWNGU D[ GUVKOCVKPI VJG ENCUUKſGT RCTCOGVGTU WPFGT VJG ETKVGTKQP QH OKPKOWO GORKTKECN ENCUUKſECVKQP GTTQT 5WEJ FGEKUKQP TWNGU CTG $C[GUŏ TKUM EQPUKUVGPV KP VJG UGPUG VJCV VJG UGSWGPEG QH GO RKTKECN TKUMU EQPXGTIGU VQ VJG $C[GUŏ TKUM CU VJG VTCKPKPI UGVU KPETGCUG KP UK\G 1H EQWTUG JQY VQ FGſPG CP QRVKOCN HQTO HQT VJG FKUETKOKPCPV HWPEVKQPU KU CRRNKECVKQP FGRGPFGPV CPF TGOCKPU NCTIGN[ CP QRGP TGUGCTEJ RTQDNGO 1P VJG QVJGT JCPF VJG IQQF PGYU KU VJCV VJG UOQQVJ /%' QDLGEVKXG HWPEVKQP RTQRQUGF KP =? ECP CRRTQZK OCVG VJG GORKTKECN GTTQT TCVG HQT VJG FGUKIP UCORNG UGV CTDKVTCTKN[ ENQUGN[ +V ECP VJWU DG WUGF CU VJG FGUKIP ETKVGTKQP VQ DG QRVKOK\GF D[ CP[ ITCFKGPVDCUGF QRVKOK\CVKQP OGVJQFU +P VJG RCUV FGECFG VJKU /%' HQTOWNCVKQP JCU DGGP GZVGPUKXGN[ UVWFKGF TGſPGF CPF UWEEGUUHWNN[ CRRNKGF VQ UQNXKPI OCP[ RCVVGTP TGEQIPKVKQP CRRNKECVKQPU UGG HQT GZCORNG = ? CPF VJG TGHGTGPEGU VJGTGKP
3.3.3 Discussion 5Q HCT YG JCXG EQPUKFGTGF VJG HQNNQYKPI VYQ UVTCVGIKGU VJCV JCXG DGGP WUGF VQ EQP UVTWEV C OQFGTP #54 U[UVGO 7UKPI plug-in MAP CU C FGEKUKQP TWNG HQT TGEQIPKVKQP FGEKUKQP CPF /. CU C ETKVGTKQP HQT VJG GUVKOCVKQP QH FGEKUKQP RCTCOGVGTU 7UKPI maximum discriminant CU C FGEKUKQP TWNG HQT TGEQIPKVKQP FGEKUKQP CPF minimum empirical classification error /%' CU C ETKVGTKQP HQT VJG GUVKOCVKQP QH FGEKUKQP RCTCOGVGTU
E\&5&3UHVV//&
6JG HQNNQYKPI EQPENWUKQPU OC[ DG FTCYP EQPEGTPKPI VJGUG VYQ UVTCVGIKGU 6JG CU[ORVQVKE DGJCXKQT QH VJG ſTUV CRRTQCEJ YKNN FGRGPF QP VJG CRRTQRTK CVGPGUU KP VJG UGPUG QH GUVKOCVQT EQPUKUVGPE[ QH VJG RCTCOGVTKE HQTOU QH VJG CUUWOGF FKUVTKDWVKQPU 6JG CU[ORVQVKE DGJCXKQT QH VJG UGEQPF CRRTQCEJ YKNN FGRGPF QP VJG EJQKEG QH VJG FKUETKOKPCPV HWPEVKQP 6JGQTGVKECNN[ URGCMKPI KV KU PQV UQ ENGCT [GV YJKEJ UVTCVGI[ KU DGVVGT HQT C OQFGT CVGN[ UK\GF VTCKPKPI UGV *QYGXGT KP VJG RCUV FGECFG KV JCU DGGP FGOQPUVTCVGF D[ OCP[ TGUGCTEJ ITQWRU VJCV YJGP UWHſEKGPV COQWPV QH representative VTCKPKPI FCVC CTG CXCKNCDNG CP #54 U[UVGO EQPUVTWEVGF WPFGT VJG UGEQPF RTKPEKRNG ECP QWVRGTHQTO KVU EQWPVGTRCTV EQPUVTWEVGF WPFGT VJG ſTUV RTKPEKRNG HQT OCP[ #54 CRRNKECVKQPU
3.4 Violations of Modeling Assumptions in ASR 3.4.1 Types of Distortions 6JG RTKPEKRNGU QH VJG EQPUVTWEVKQP QH VJG CDQXGOGPVKQPGF QRVKOCN FGEKUKQP TWNG CPF CFCRVKXG FGEKUKQP TWNGU CTG DCUGF QP UQOG CUUWORVKQPU YJKEJ OC[ DG XKQNCVGF KP RTCEVKEG (TQO VJG EQORWVCVKQPCN OQFGNKPI RQKPV QH XKGY VJGTG CTG VJTGG OCKP FKU VQTVKQP V[RGU VJCV RTQFWEG XKQNCVKQPU QH CUUWORVKQPU UWOOCTK\GF CU HQNNQYU =? FKUVQTVKQPU ECWUGF D[ UOCNNUCORNG GHHGEVU FKUVQTVKQPU QH OQFGNU QT FKUETKOKPCPV HWPEVKQPU HQT VTCKPKPI UCORNGU CPF FKUVQTVKQPU QH VTCKPGF OQFGNU QT FKUETKOKPCPV HWPEVKQPU HQT QDUGTXCVKQPU VQ DG ENCUUKſGF 6JG FKUVQTVKQPU ECWUGF D[ UOCNNUCORNG GHHGEVU CTG V[RKECN HQT CNN UVCVKUVKECN RNWIKP RTQEGFWTGU 6JG[ CTKUG HTQO VJG PQPEQKPEKFGPEG QH VJG UVCVKUVKECN GUVKOCVGU QH RTQDCDKNKV[ EJCTCEVGTKUVKEU CPF VJGKT VTWG XCNWGU 9G YCPV VQ GORJCUK\G CICKP VJCV VJG RNWIKP FGEKUKQP TWNGU FGUETKDGF KP RTGXKQWU UGEVKQP CTG CU[ORVQVKECNN[ QRVKOCN QPN[ YJGP
VJG VTCKPKPI UCORNGU CTG EQNNGEVGF D[ C UGTKGU QH independent GZRGTKOGPVU UWEJ VJCV QT OQTG KPVWKVKXGN[ URGCMKPI UJQWNF DG representative GPQWIJ YKVJ TGURGEV VQ VJG VTWG FKUVTKDWVKQP QH VJG VGUVKPI FCVC CPF
VTCKPKPI UCORNG UK\G CXCKNCDNG
E\&5&3UHVV//&
KG VJGTG KU UWHſEKGPV COQWPV QH VTCKPKPI FCVC
+P RTCEVKEG VJG VTCKPKPI UCORNG UGV CNYC[U JCU C ſPKVG UK\G KG CPF KP OCP[ ECUGU KU RQUUKDN[ CNUQ PQV TGRTGUGPVCVKXG GPQWIJ 6JG TCPFQO FGXKCVKQPU QH UVCVKUVKECN GUVKOCVGU ECP VJGP RTQFWEG UKIPKſ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ƀGZKDNG RCTCOGVTKE HQTOU QH VJG OQFGN QT FKUETKOKPCPV HWPEVKQP VJG OKUENCUUKſECVKQP QH VTCKPKPI UCORNGU QWV NKGTU KP VTCKPKPI UCORNGU GVE 6JG[ YKNN ECWUG DQVJ modeling error CPF estimation error 6Q EQRG YKVJ VJGUG RTQDNGOU DGVVGT OQFGNU QT FKUETKOKPCPV HWPEVKQPU PGGF VQ DG HQWPF CPF VGEJPKSWGU PGGF VQ DG FGUKIPGF HQT TQDWUV NGCTPKPI HTQO FCVC 6JG DKIIGUV RTQDNGO HQT #54 OKIJV DG ECWUGF D[ VJG VJKTF V[RG QH FKUVQTVKQP +P OQUV TGCN CRRNKECVKQPU VJGTG CNYC[U GZKUVU UQOG HQTO QH OKUOCVEJ YJKEJ ECWUGU C FKUVQTVKQP DGVYGGP VJG VTCKPGF OQFGNU QT FKUETKOKPCPV HWPEVKQPU CPF VJG VGUV FCVC 6JGUG OKUOCVEJGU UQOG QH VJGO KFGPVKſGF KP (KIWTG OC[ CTKUG HTQO KPVGT CPF KPVTCURGCMGT XCTKCDKNKVKGU VTCPUFWEGT EJCPPGN CPF QVJGT GPXKTQPOGPVCN XCTKCDKNKVKGU CPF OCP[ QVJGT RJQPGVKE CPF NKPIWKUVKE GHHGEVU ECWUGF D[ OKUOCVEJ KP VTCKPKPI CPF VGUVKPI VCUM FGſPKVKQPU *QY VQ CEJKGXG VJG RGTHQTOCPEG TQDWUVPGUU KP VJKU EQPVGZV JCU DGEQOG QPG QH VJG OQUV CEVKXG TGUGCTEJ CTGCU KP #54 KP VJG RCUV FGECFG
3.4.2 Towards Adaptive and Robust ASR (TQO VJG CDQXG CPCN[UKU CPF FKUEWUUKQP KV KU SWKVG ENGCT VJCV KP QTFGT VQ FGUKIP CP CWVQOCVKE URGGEJ TGEQIPK\GT VJCV YQTMU YGNN HQT FKHHGTGPV VCUMU CPF URGCMGTU QXGT WP GZRGEVGF CPF RQUUKDN[ CFXGTUG EQPFKVKQPU CNN QH VJG CDQXG VJTGG FKUVQTVKQP V[RGU PGGF VQ DG CRRTQRTKCVGN[ VTGCVGF 1PG QH VJG GHHGEVKXG YC[U VQ KORTQXG #54 TQDWUVPGUU KU VQ ſPF KPXCTKCPV QT TQDWUV HGCVWTGU UQ CU VQ OKPKOK\G VJG QDUGTXCVKQP XCTKCDKNKV[ ECWUGF D[ VJG FKHHGTGPV V[RGU QH KPVGTHGTKPI HCEVQTU CPF VJG RQUUKDNG OKUOCVEJ DG VYGGP VTCKPKPI CPF VGUVKPI EQPFKVKQPU 'XGP VJQWIJ UQOG HGCVWTGU JCXG DGGP UJQYP NGUU CHHGEVGF D[ C EGTVCKP V[RG QH FKUVQTVKQP UWEJ CU NKPGCT OKETQRJQPG QT EJCPPGN GHHGEV PQ HGCVWTG JCU [GV DGGP FKUEQXGTGF VJCV KU KPXCTKCPV CETQUU CNN CFXGTUG CEQWUVKE EQPFKVKQPU (WTVJGT TGUGCTEJ KP HTQPVGPF UKIPCN RTQEGUUKPI CPF HGCVWTG GZVTCEVKQP KU FGſPKVGN[ PGGFGF VQ KORTQXG QP VJG EWTTGPVN[ őUVCPFCTFŒ CEQWUVKE CPCN[UKU HQT #54 =? 1PEG VJG HGCVWTG GZVTCEVKQP OGVJQF KU ſZGF CPQVJGT VTCFKVKQPCN CRRTQCEJ VQ TQDWUV URGGEJ TGEQIPKVKQP KU VQ FGXGNQR DGVVGT OQFGNKPI CPF NGCTPKPI VGEJPKSWGU VJCV JCXG C IQQF IGPGTCNK\CVKQP ECRCDKNKV[ +P CFFKVKQP HQWT OCLQT ENCUUGU QH UVCVKUVKECN VGEJPKSWGU VQ KORTQXG #54 TQDWUVPGUU ECP DG FGſPGF CFCRVKPI TGEQIPK\GT RCTCOGVGTU VQ PGY QRGTCVKPI EQPFKVKQPU WUKPI CFCRVCVKQP CPFQT VGUVKPI FCVC
E\&5&3UHVV//&
OQFKH[KPI UKIPCN HGCVWTG QT TGEQIPK\GT RCTCOGVGTU WUKPI QPN[ VJG WVVGTCPEG VQ DG TGEQIPK\GF VQ TGFWEG VJG OKUOCVEJ DGVYGGP VJG VTCKPKPI CPF VGUVKPI EQPFK VKQPU WUKPI TQDWUV FGEKUKQP UVTCVGIKGU CPF RQUUKDNG EQODKPCVKQPU QH VJG CDQXG VGEJPKSWGU #NQPI VJGUG NKPGU OCP[ VGEJPKSWGU JCXG DGGP FGXGNQRGF CPF CTG TGXKGYGF HTQO FKHHGTGPV RGTURGEVKXGU KP HQT GZCORNG = ? 4GCFGTU CTG TGHGTTGF VQ VJGUG TGXKGYU HQT C TKEJ RKEVWTG QH VJG ſGNF CPF VJG TGHGTGPEGU VJGTGKP HQT VJG FGVCKNU QH VJG FKHHGTGPV VGEJPKSWGU +P VJG TGOCKPKPI RCTV QH VJG EJCRVGT + YKNN DTKGƀ[ TGXKGY VYQ VGEJPQNQIKGU PCOGN[ TGEQIPK\GT RCTCOGVGT CFCRVCVKQP CPF TQDWUV FGEKUKQP TWNGU VJCV YGTG FGXGNQRGF KP VJG RCUV FGECFG VQ EQRG YKVJ VJG CDQXG RTQDNGOU 6JG UGNGEVKQP QH OCVGTKCNU KU IWKFGF D[ VJG EQPUKFGTCVKQP VJCV FKUEWUUKQPU ECP DG OCFG KP C TGNCVKXGN[ OQTG TKIQTQWU YC[ HTQO VJG XKGYRQKPV QH RTGXKQWUN[ FKUEWUUGF FGEKUKQP VJGQTGVKE HQTOWNCVKQPU HQT VJG #54 RTQDNGO
3.5 Improving Adaptive Decision Rules via Decision Parameter Adaptation 3.5.1 Decision Parameter Adaptation for Stationary Operating Conditions +H VJG QRGTCVKPI EQPFKVKQP QH C URGGEJ TGEQIPK\GT KU UVCVKQPCT[ VJGP VJGTG OWUV GZKUV C VTWG FKUVTKDWVKQP 5WRRQUG VJG VTCKPKPI FCVC KU PQV TGRTGUGPVCVKXG GPQWIJ UQ VJCV VJG TGEQIPK\GT EQPUVTWEVGF WUKPI VJG FGUKIP RTKPEKRNGU FKUEWUUGF RTGXKQWUN[ FQGU PQV YQTM UQ YGNN HQT VJG VGUVKPI FCVC HTQO +H VJG CRRNKECVKQP UEG PCTKQ CNNQYU C UVTCKIJVHQTYCTF UQNWVKQP VQ KORTQXKPI VJG CFCRVKXG FGEKUKQP TWNGU KU VQ EQNNGEV CFFKVKQPCN VTCKPKPI FCVC MPQYP CU CFCRVCVKQP FCVC KP C URGEKſE VGUVKPI EQPFKVKQP UWEJ VJCV CPF VJGP VQ CFCRV VJG TGEQIPK\GT RCTCOGVGTU CEEQTFKPIN[ VQ YQTM DGVVGT KP VJG RTGUETKDGF UEGPCTKQ &GRGPFKPI QP YJKEJ FGUKIP RTKPEKRNG YCU WUGF VQ EQPUVTWEV VJG URGGEJ TGEQIPK\GT HTQO VJG VTCKPKPI UCORNG VJGTG CTG PCVWTCNN[ VYQ goals of adaptation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
E\&5&3UHVV//&
3.5.1.1 Adaptation for Plug-in Decision Rules %QPUKUVGPV YKVJ VJG ſTUV FGUKIP RTKPEKRNG FKUEWUUGF KP 5GEVKQP OCP[ UWEEGUUHWN CFCRVCVKQP VGEJPKSWGU JCXG DGGP FGXGNQRGF KP VJG RCUV FGECFG VQ EQRG YKVJ VJG RQU UKDNG RTQDNGO QH OKUOCVEJGU DGVYGGP VTCKPKPI CPF VGUVKPI EQPFKVKQPU $GECWUG YG JCXG CNTGCF[ IKXGP CP QXGTXKGY QH VJGUG VGEJPKSWGU KP =? TGEGPVN[ + LWUV CFF VYQ OQTG TGOCTMU JGTG VQ UWRRNGOGPV VJG FGVCKNGF FKUEWUUKQPU KP =? Remark 1: 6Q FGCN YKVJ VJG URCTUG FCVC RTQDNGO VYQ UVTCVGIKGU JCXG DGGP UWEEGUU HWNN[ WUGF 1PG KU VJG CRRTQCEJ QH regularization CPF CPQVJGT KU VJG CRRTQCEJ QH imposing constraints VQ TGFWEG VJG FGITGGU QH HTGGFQO HQT RCTCOGVGT GU VKOCVKQP 6JG RQRWNCT $C[GUKCP RQKPV GUVKOCVG UWEJ CU VJG /#2 GUVKOCVG KU CP GZCORNG QH VJG HQTOGT YJKNG VJG VTCPUHQTOCVKQPDCUGF CRRTQCEJ KU CP GZ CORNG QH VJG NCVVGT 5Q VJG /#2 GUVKOCVG KU UQOGVKOGU CNUQ TGHGTTGF VQ CU maximum penalized likelihood GUVKOCVG 1H EQWTUG VJG CDQXG VYQ UVTCVGIKGU ECP DG UKOWNVCPGQWUN[ WUGF VQ FGCN YKVJ VJG URCTUG FCVC RTQDNGO Remark 2: 7PUWRGTXKUGF CFCRVCVKQP TGOCKPU NCTIGN[ CP WPUQNXGF TGUGCTEJ RTQDNGO 6TCPUHQTOCVKQPDCUGF WPUWRGTXKUGF CFCRVCVKQP YQTMU UQOGVKOGU LWUV DGECWUG VJG VTCPUHQTOCVKQPU CTG RQUUKDN[ UJCTGF D[ FKHHGTGPV URGGEJ WPKVU VJWU VJG EQP UGSWGPEG QH VJG YTQPI UWRGTXKUKQP KU PQV CU UGXGTG CU KP QVJGT CRRTQCEJGU YKVJQWV WUKPI VJG OGEJCPKUO QH RCTCOGVGT V[KPI QT UJCTKPI 3.5.1.2 Adaptation for Maximum-Discriminant Decision Rules +P EQPVTCUV YKVJ VJG GZVGPUKXG TGUGCTEJGU WPFGT VJG ſTUV FGUKIP RTKPEKRNG NGUU GHHQTVU JCXG DGGP FGXQVGF VQ FGXGNQR VGEJPKSWGU HQT FGEKUKQP RCTCOGVGT CFCRVCVKQP YJKEJ KU EQPUKUVGPV YKVJ VJG UGEQPF FGUKIP RTKPEKRNG FKUEWUUGF KP 5GEVKQP # UVWF[ QP /%' CFCRVCVKQP QH %&*// RCTCOGVGTU YCU ſTUV ECTTKGF QWV D[ CWVJQTU QH =? 5GXGTCN HQNNQYWR UVWFKGU YGTG CNUQ TGRQTVGF D[ QVJGT TGUGCTEJ ITQWRU = ? # OQTG TGEGPV QPG YCU TGRQTVGF KP =? CPF FGOQPUVTCVGF VJCV FKTGEV /%' CFCRVCVKQP HQT /%'VTCKPGF *// RCTCOGVGTU YQTMU YGNN YJGP UWHſEKGPV w.r.t. VJG PWODGT QH RCTCOGVGTU DGKPI CFCRVGF COQWPV QH CFCRVCVKQP FCVC CTG CXCKNCDNG *QYGXGT YJGP QPN[ UOCNN COQWPV QH CFCRVCVKQP FCVC CTG CXCKNCDNG FKTGEV /%' CFCRVCVKQP QH *// RCTCOGVGTU FQGU PQV YQTM UQ YGNN 6JG NCEM QH CP GHſEKGPV CFCRVCVKQP CNIQTKVJO HQT /%'VTCKPGF UGGF OQFGNU OKIJV DG QPG QH VJG OCKP TGCUQPU YJ[ VJG /%' VTCKPKPI JCU PQV DGGP YKFGN[ WUGF [GV VQ EQPUVTWEV CP #54 U[UVGO HQT CRRNKECVKQPU KP YJKEJ FGEKUKQP RCTCOGVGT CFCRVCVKQP KU TGSWKTGF +P VJG RCUV UGXGTCN [GCTU VJGTG JCXG DGGP UQOG GHHQTVU VQ FGXGNQR FKUETKOKPCVKXG NKPGCT TGITGUUKQP CFCRVCVKQP VGEJPKSWGU WPFGT FKHHGTGPV ETKVGTKC CPF PQVKQPU UWEJ CU /%' =? OCZKOWO UECNGF NKMGNKJQQF =? OCZKOCN TCPM NKMGNKJQQF =? //+
OCZKOWO OWVWCN KPHQTOCVKQP =? CPF %/. EQPFKVKQPCN OCZKOWO NKMGNKJQQF =? +PVGTGUVKPIN[ CNVJQWIJ CNN QH VJGO CTG FGXGNQRGF YKVJ VJG CKO QH CP GHſEKGPV FKUETKOKPCVKXG CFCRVCVKQP VJG[ JCXG QPN[ DGGP CRRNKGF VQ CFCRVKPI VJG /.VTCKPGF UGGF OQFGNU 0Q TGUWNVU JCXG DGGP TGRQTVGF [GV JQY VJG[ YQTM HQT VJG CFCRVCVKQP QH VJG FKUETKOKPCVKXGN[ VTCKPGF UGGF OQFGNU +P O[ QRKPKQP VJKU KU C OQTG FGUKTCDNG UEGPCTKQ VQ CRRN[ FKUETKOKPCVKXG CFCRVCVKQP DGECWUG VJG EQPUKUVGPV ETKVGTKC CTG WUGF
E\&5&3UHVV//&
KP DQVJ UGGF OQFGN VTCKPKPI CPF VJG UWEEGGFKPI CFCRVCVKQP CPF JQRGHWNN[ C DGVVGT RGTHQTOCPEG ECP DG CEJKGXGF KP VJKU YC[ +V KU VJKU HCEV VJCV OQVKXCVGU WU VQ RGTHQTO C UVWF[ CU TGRQTVGF KP =? +P =? YG JCXG RTGUGPVGF C HQTOWNCVKQP QH OKPKOWO ENCUUKſECVKQP GTTQT NKPGCT TGITGUUKQP /%'.4 HQT CFCRVCVKQP QH )CWUUKCP OKZVWTG %&*// RCTCOGVGTU 9G FGOQPUVTCVG VJCV VJG /%'.4 ECP DG WUGF VQ CFCRV VJG /%'VTCKPGF *// RCTCOGVGTU WPFGT C EQPUKUVGPV ETKVGTKQP +P C UWRGTXKUGF URGCMGT CFCRVCVKQP CRRNKECVKQP YG QDUGTXG VJCV UWEJ CFCRVGF OQFGNU RGTHQTO DGVVGT VJCP VJG QPGU CFCRVGF WUKPI OCZKOWO NKMGNKJQQF NKPGCT TGITGUUKQP /..4 HTQO VJG /. VTCKPGF UGGF OQFGNU (WTVJGT UVWFKGU CTG PGGFGF VQ GZRNQTG /%'.4ŏU DGJCXKQT HQT NQPIVGTO CFCRVCVKQP WUKPI KPETGCUKPI COQWPV QH CFCRVCVKQP FCVC +P CFFKVKQP VQ WUKPI VJG CDQXG OKPKOWO empirical ENCUUKſECVKQP GTTQT ETKVGTKQP HQT FGEKUKQP RCTCOGVGTU CFCRVCVKQP QPG ECP CNUQ CFQRV CPQVJGT ETKVGTKQP ECNNGF minimum expected classification error CU FGſPGF KP VJG HQNNQYKPI
Ï ¾ªÏ ¾ªÜ
£
YJGTG £ KU C NQUU HWPEVKQP EJCTCEVGTK\GF D[ VJG FGEKUKQP TWNG RCTCOGVGTU #RRCTGPVN[ VJG CDQXG QDLGEVKXG HWPEVKQP KU CP WPFGTFGſPGF HWPEVKQPCN DGECWUG VJG VTWG FKUVTKDWVKQP KU WPMPQYP *QYGXGT D[ WUKPI VJG stochastic approximation OGVJQF UWIIGUVGF KP VJG U D[ 4QDDKPU CPF /QPTQG =? VJG HWPEVKQPCN ECP DG OKPKOK\GF YKVJ TGURGEV VQ VJG RCTCOGVGTU D[ WUKPI VJG VGUVKPI FCVC FTCYP HTQO CU HQNNQYU
·½
£
+V ECP DG RTQXGP VJCV VJKU OGVJQF KU EQPUKUVGPV WPFGT XGT[ IGPGTCN EQPFKVKQPU QP VJG ITCFKGPV £ CPF VJG UEJGFWNG QH VJG NGCTPKPI TCVG *KUVQTKECNN[ VJKU CRRTQCEJ YCU KPFGRGPFGPVN[ RTQRQUGF CPF FGXGNQRGF HQT RCVVGTP TGEQIPKVKQP CRRNKECVKQP D[ #OCTK =? CPF 6U[RMKP = ? TGURGEVKXGN[ #ICKP C UOQQVJ NQUU HWPEVKQP £ YCU RTQRQUGF KP =? VQ OCMG VJG CDQXG RTQEGFWTG RTCEVKECNN[ WUGHWN #RRCTGPVN[ VJG CDQXG IGPGTCN NGCTPKPI RTKPEKRNG ECP DG WUGF HQT UWRGTXKUGF QPNKPG CFCRVCVKQP QH VJG FGEKUKQP RCTCOGVGTU *QYGXGT KV EQPXGTIGU KP RTQDCDKNKV[ YJKEJ OGCPU VJCV VJG CNIQTKVJO EQPXGTIGU QPN[ CHVGT C NCTIG COQWPV QH UCORNGU CTG WUGF 6JKU OCMGU VJG CRRTQCEJ OQTG UWKVCDNG HQT NQPIVGTO CFCRVCVKQP
3.5.2 Decision Parameter Adaptation for Slowly Changing Operating Conditions
/QUV QH VJG GZKUVKPI CFCRVCVKQP CNIQTKVJOU VTGCV VJG KPFKXKFWCN FCVC DNQEM QH VJG CXCKNCDNG CFCRVCVKQP FCVC CU GSWCNN[ KORQTVCPV VJWU CTG XCNKF QPN[ KP C UVC VKQPCT[ QRGTCVKPI EQPFKVKQP HQT GUVKOCVKPI UVCVKQPCT[ RCTCOGVGTU *QYGXGT KP OCP[ TGCN URGGEJ TGEQIPKVKQP CRRNKECVKQPU VJG UVCVKUVKECN EJCTCEVGTKUVKEU QH VJG QDUGTXCVKQP FCVC WPFGTIQ ITCFWCN EJCPIGU FWG VQ OCP[ RQUUKDNG HCEVQTU UWEJ CU VJG EJCPIKPI URGCMKPI DGJCXKQT QH C URGCMGT VJG EJCPIKPI QRGTCVKPI GPXKTQPOGPV VJG EJCPIKPI VTCPUOKUUKQP EJCPPGN GVE 6JG RTQDNGO QH RCTCOGVTKE NGCTPKPI YKVJ UWEJ UNQYN[
E\&5&3UHVV//&
EJCPIKPI QRGTCVKPI EQPFKVKQPU KU VQ GUVKOCVG VKOGXCT[KPI FGEKUKQP TWNG RCTCOGVGTU +P UWEJ ECUGU FKHHGTGPV FCVC UGIOGPVU QHVGP EQTTGURQPF VQ FKHHGTGPV RCTCOGVGT XCNWGU +P QTFGT VQ EQPVKPWQWUN[ VTCEM VJG XCTKCVKQPU QH VJG OQFGN RCTCOGVGTU EQTTGURQPFKPI VQ VJG PGY FCVC UQOG forgetting mechanisms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ſPCN TGOCTM VJG CDQXG CNIQTKVJOU CTG FGXGNQRGF VQ CFCRV VJG RCTCOGVGTU QH VJG RNWIKP /#2 FGEKUKQP TWNG *QY VQ CFCRV VJG RCTCOGVGTU QH VJG OCZKOWOFKUETKOKPCPV FGEKUKQP TWNG KP C PQPUVCVKQPCT[ QRGTCVKPI EQPFKVKQP TGOCKPU CP KPVGTGUVKPI QRGP RTQDNGO
3.5.3 Decision Parameter Adaptation for Switching Operating Conditions 9JGP CP #54 U[UVGO JCU VQ DG QRGTCVGF WPFGT TCRKFN[ UYKVEJKPI EQPFKVKQPU VJG CDQXG CFCRVCVKQP CNIQTKVJOU ECP PQV DG CRRNKGF +H VJG PQPUVCVKQPCT[ QRGTCVKPI EQP FKVKQP ECP DG CRRTQZKOCVGF D[ C ſPKVG PWODGT QH FKHHGTGPV UVCVKQPCT[ EQPFKVKQPU VJGP C UKORNG UQNWVKQP EQWNF DG KOCIKPGF #P QHƀKPG EQPFKVKQPENWUVGTKPI ECP DG RGT HQTOGF ſTUV CPF CP KPFKXKFWCN TGEQIPK\GT KU VJGP EQPUVTWEVGF HQT GCEJ ENWUVGT )KXGP CP WPMPQYP WVVGTCPEG VQ DG TGEQIPK\GF VJG OQUV UKOKNCT EQPFKVKQPENWUVGT ECP DG KFGPVKſGF CPF VJG CUUQEKCVGF TGEQIPK\GT ECP DG WUGF VQ TGEQIPK\G VJG WPMPQYP WV VGTCPEG 6JG VTCFKVKQPCN VGEJPKSWG QH URGCMGT CFCRVCVKQP XKC URGCMGT ENWUVGTKPI CPF UGNGEVKQP KU C IQQF GZCORNG QH VJKU UVTCVGI[ = ? *QYGXGT KH VJG EWTTGPV QR GTCVKPI EQPFKVKQP KU PQV UKOKNCT VQ CP[ UKPING VTCKPKPI EQPFKVKQP [GV UJCTGU EGTVCKP EJCTCEVGTKUVKEU YKVJ UQOG VTCKPKPI EQPFKVKQPU VJGP C UVTCVGI[ QH adaptive model fusion ECP DG CFQRVGF CU HQT GZCORNG KP = ? #NN QH VJGUG YQTMU UJCTG VJG UKOKNCTKV[ KP VJG IGPGTCN UGPUG VJCV VJG[
¯ ſTUV RTGRCTG QHƀKPG C UGV QH OQFGNU HTQO VTCKPKPI FCVC CPF VJGP ¯ HWUG CFCRVKXGN[ D[ WUKPI VJG KPHQTOCVKQP GODGFFGF KP VJG WVVGTCPEG VQ DG TGE QIPK\GF C UGV QH PGY OQFGNU YJKEJ JQRGHWNN[ KU OQTG őCRRTQRTKCVGŒ VQ VJG VGUVKPI WVVGTCPEG CPF ſPCNN[ ¯ TGTGEQIPK\G VJG VGUVKPI WVVGTCPEG CICKP #NVJQWIJ VJG CDQXG CRRTQCEJGU JCXG OCKPN[ DGGP FGXGNQRGF CPF UVWFKGF HQT FGCNKPI YKVJ VJG URGCMGT XCTKCDKNKV[ VJG UCOG KFGC QH QHƀKPG XCTKCDKNKV[ FGEQORQUKVKQP CPF
E\&5&3UHVV//&
QPNKPG CFCRVKXG OQFGN HWUKQP ECP DG HWTVJGT GZRNQTGF VQ FGCN YKVJ QVJGT CURGEVU QH TQDWUV #54
3.5.4 Discussion 5Q HCT + JCXG DTKGƀ[ FKUEWUUGF UGXGTCN UVTCVGIKGU HQT FGEKUKQP RCTCOGVGT CFCRVCVKQP KP VJTGG V[RGU QH QRGTCVKPI EQPFKVKQPU +V KU ENGCT VJCV VJG ITGCVGUV EJCNNGPIG EQOGU HTQO VJQUG CRRNKECVKQPU YJKEJ QPN[ KPXQNXG C EQWRNG QH WVVGTCPEGU DWV GXGT[ WVVGTCPEG KPXQNXGU C FKUVKPEV [GV EQORNKECVGF őFKUVQTVKQP EJCPPGNŒ HTQO VJG KPVGPFGF OGUUCIG C URGCMGT YCPV VQ EQPXG[ VQ VJG TGEGKXGF UKIPCN QH C URGGEJ TGEQIPK\GT #HVGT CNN QH VJG CDQXG CFCRVCVKQP VGEJPKSWGU JCXG DGGP EQPUKFGTGF CPQVJGT UVTCVGI[ PCOGN[ robust decision rule ECP CNYC[U DG VTKGF QWV VQ UGG YJGVJGT RGTHQTOCPEG TQDWUVPGUU ECP DG HWTVJGT KORTQXGF +P VJG TGOCKPKPI RCTV QH VJG EJCRVGT 6JKU UVTCVGI[ KU GZRNCKPGF KP FGVCKN
3.6 Robust Decision Rules 3.6.1 Decision Rule Robustness +PVWKVKXGN[ URGCMKPI C FGEKUKQP UVTCVGI[ TWNG KU ECNNGF TQDWUV KH KV KU PQV XGT[ UGPUK VKXG VQ VJG RTGXKQWUN[ FKUEWUUGF RTKQT WPEGTVCKPV[ QT FKUVQTVKQPU /QTG HQTOCNN[ NGV DG CP CTDKVTCT[ FGEKUKQP TWNG EQPUVTWEVGF WPFGT UQOG J[RQVJGVKECN OQFGN ¼ YJGTG Ï KU VJG ENCUU VQ YJKEJ VJG QDUGTXCVKQP Ü YKNN DG CUUKIPGF CPF KU C VTCKPKPI UCORNG UGV WUGF HQT VJG EQPUVTWEVKQP QH VJG FGEKUKQP TWNG .GV ¯ FGPQVG CP CTDKVTCT[ CFOKUUKDNG FKUVQTVGF FCVC OQFGN HQT VJG FKUVQTVKQP V[RGU FKUEWUUGF KP 5GEVKQP YJGTG KU WUGF VQ EJCTCEVGTK\G VJG FKUVQTVKQP NGXGN .GV £¯ FGPQVG VJG UGV QH CFOKUUKDNG FKUVQTVGF FCVC OQFGNU 6JG ENCUUKſEC VKQP RGTHQTOCPEG QH VJG FGEKUKQP TWNG KP C UKVWCVKQP YJGTG FCVC CTG ſVVGF VQ VJG FKUVQTVGF OQFGN ¯ £¯ YKNN DG EJCTCEVGTK\GF D[ VJG TKUM HWPEVKQPCN
¯
YJGTG FGPQVGU VJG GZRGEVCVKQP YKVJ TGURGEV VQ VJG RTQDCDKNKV[ FKUVTKDWVKQP QH EQTTGURQPFKPI VQ VJG FKUVQTVGF OQFGN ¯ £ ¯ .GV WU ECNN VJG HWPEVKQPCN
·
·
ů ¾Å¯
¯
VJG guaranteed (upper) risk =? HQT VJG FGEKUKQP TWNG KP VJG RTGUGPEG QH FKUVQT VKQPU ¯ £¯ +H YG MPQY VJG FKUVTKDWVKQP QH ¯ QP £¯ YG ECP HWTVJGT FGſPG VJG HQNNQYKPI HWPEVKQPCN
¯
YJGTG FGPQVGU VJG GZRGEVCVKQP YKVJ TGURGEV VQ VJG FKUVTKDWVKQP QH QP £ 9G ECNN VJG overall risk #RRCTGPVN[ DQVJ · CPF ECP DG WUGF CU QRVKOCNKV[ ETKVGTKC KP UGCTEJKPI HQT robust (with respect to distortions £ ) decision rules # FGEKUKQP TWNG £ YKVJ VJG OKPKOCN XCNWG QH VJG IWCTCPVGGF TKUM HQT CNN CFOKUUKDNG FKUVQTVKQPU £ ·
´ ¡µ
KU TGHGTTGF VQ CU C minimax decision rule # FGEKUKQP TWNG XCNWG QH VJG QXGTCNN TKUM HQT CNN CFOKUUKDNG FKUVQTVKQPU
YKVJ VJG OKPKOCN
´ ¡µ
KU TGHGTTGF VQ CU C predictive decision rule 6JG EQPUVTWEVKQP QH VJGUG TQDWUV FGEKUKQP TWNGU YKNN FGRGPF QP JQY VJG CFOKUUKDNG FKUVQTVKQPU £ CTG FGſPGF CPF CNUQ HQT VJG ECUG QH VJG RTGFKEVKXG FGEKUKQP TWNG VJG FKUVTKDWVKQP QH VJG FKUVQTVKQP QP £ +P VJG HQNNQYKPI VYQ UWDUGEVKQPU + UJQY VYQ GZCORNGU QH UWEJ TQDWUV FGEKUKQP TWNGU PCOGN[ minimax decision rule CPF Bayesian predictive decision rule TGURGEVKXGN[ $QVJ QH VJGO CUUWOG VJCV
VJG FKUVTKDWVKQPU CPF CTG MPQYP WR VQ UQOG URGEKſCDNG RCTCO GVGTU KP VJG HQTOU QH £ CPF
VJG VTWG RCTCOGVGTU QH VJGUG FKUVTKDWVKQPU CPF NKG KP C PGKIJDQTJQQF QH VJG GUVKOCVGF QT J[RQVJGVKECN QPGU VJGTGHQTG VJG RTKQT WPEGTVCKPV[ ECP DG OQFGNGF D[ FGſPKPI CP uncertainty neighborhood QH VJG OQFGN RCTCOGVGTU CPFQT RQUUKDN[ C FKUVTKDWVKQP QH OQFGN RCTCO GVGTU QP VJKU WPEGTVCKPV[ PGKIJDQTJQQF
9KVJ VJGUG CUUWORVKQPU VJG URGEKſE OKPKOCZ FGEKUKQP TWNG CPF RTGFKEVKXG FGEKUKQP TWNG ECP DG EQPUVTWEVGF CEEQTFKPIN[ VQ UCVKUH[ UQOG FGUKTGF TQDWUVPGUU RTQRGTVKGU
3.6.2 Minimax Classification Rule .GV ¼ ¼ FGPQVG VJG uncertainty neighborhood QH VJG VTWG OQFGN RCTCOGVGTU KG ¼ ¼ YJGTG ¼ ¼ CTG OQFGN RCTCOGVGTU GUVKOCVGF HTQO VJG VTCKPKPI FCVC CPF ECP DG XKGYGF CU C IGPGTKE RCTCOGVGT VQ EJCTCEVGTK\G VJG FGITGG QH VJG FKUVQTVKQP 6JGP YG JCXG £
£
YJGTG £ KU VJG UGV QH FKUVQTVGF OQFGNU CPF ·
·
´£ µ¾¯ ´£¼ ¼ µ ¾ªÏ
¼ ¼
¾ªÜ
£
6Q EQPUVTWEV C OKPKOCZ FGEKUKQP TWNG YJKEJ OKPKOK\GU VJG CDQXG IWCTCPVGGF TKUM · KU PQV C GCU[ VCUM +P RTCEVKEG UQOG OQTG TGNCZGF ETKVGTKC JCXG VQ DG
E\&5&3UHVV//&
CFQRVGF 1PG RQUUKDKNKV[ KU VQ WUG VJG WRRGT DQWPF QH · YJKEJ YG FGPQVG ·· ·· ··
Ï ¾ªÏ ¾ªÜ ´£ µ¾ ¯ ´£¼ ¼ µ
£
6Q UKORNKH[ QWT FKUEWUUKQP YG CUUWOG VJCV YG FQ PQV EQPUKFGT VJG WPEGTVCKPV[ QH VJGTGCHVGT CPF WUG ¼ CU VJG NCPIWCIG OQFGN YKVJ ¼ DGKPI VJG UGV QH NCPIWCIG OQFGN RCTCOGVGTU GUVKOCVGF HTQO VJG VTCKPKPI VGZV FCVC $[ WUKPI VJG NQUU HWPEVKQP YG VJGP JCXG
·· ··
¾ª Ï
¼
ªÜ ´ µ £ ¯ ´£¼ µ
£
# FGEKUKQP TWNG YJKEJ OKPKOK\GU VJG CDQXG ·· KU CU HQNNQYU
·· ¼
£ ¯ ´£¼ µ £
6JKU KU VJG UQECNNGF minimax decision rule YJKEJ YCU ſTUV UVWFKGF D[ /GTJCX CPF .GG KP =? +V ECP DG UQNXGF KP VYQ UVGRU (KTUV YG GUVKOCVG VJG WPFGTN[KPI RCTCOG ´ µ VGTU WUKPI VJG /. CRRTQCEJ YKVJKP GCEJ PGKIJDQTJQQF ¼ KG
£ £ ¯ ´£´¼Ï µ µ
´ µ
YJGTG ¼ FGPQVGU RTGVTCKPGF OQFGN RCTCOGVGTU HQT YQTF 6JGP YG CRRN[ ´ µ 6JGTGHQTG VJG RNWIKP /#2 FGEKUKQP TWNG YKVJ TGRNCEKPI VJG QTKIKPCN ¼ EQPEGRVWCNN[ VJG OKPKOCZ FGEKUKQP TWNG FGUETKDGF KP 'S ECP DG XKGYGF CU C RTQEGFWTG YJKEJ OQFKſGU VJG RNWIKP /#2 FGEQFGT UJQYP KP 'S YKVJ CP GZVTC UVGR CU KP 'S VQ ſPF C OQFKſGF RQKPV GUVKOCVG KP VJG PGKIJDQTJQQF ´ µ ´ µ ¼ ¼ QH VJG QTKIKPCN ENCUUKſGT RCTCOGVGTU ¼ ¼ 6JG CDQXG TQDWUV OKPKOCZ ENCUUKſECVKQP TWNG OCMGU PQ CUUWORVKQP CDQWV VJG HQTO QH VJG FKUVQTVKQP *QYGXGT KVU GHſECE[ FQGU FGRGPF QP CP CRRTQRTKCVG URGEKſECVKQP QH ´ µ VJG RCTCOGVGT WPEGTVCKPV[ PGKIJDQTJQQF ¼ ¼ +P VJG RCUV UGXGTCN [GCTU UQOG QVJGT URGEKſE VGEJPKSWGU JCXG CNUQ DGGP FGXGNQRGF VQ KORNGOGPV VJG CDQXG OKPKOCZ FGEKUKQP TWNG KP *//DCUGF #54 U[UVGOU = ? 6JG[ CTG UJQYP VQ DG GHHGEVKXG KP FGCNKPI YKVJ PQKU[ URGGEJ TGEQIPKVKQP CPF VJG OKUOCVEJ ECWUGF D[ FKHHGTGPV TGEQTFKPI EQPFKVKQPU 6JGTG CTG JQYGXGT QVJGT RQUUKDKNKVKGU VQ OQFGN VJG CFOKUUKDNG FKUVQTVKQPU (QT GZCORNG KH YG WUG
£
¼ ¼ FGPQVGU C URGEKſE VTCPUHQTOCVKQP QH ¼ YKVJ RCTCOGVGTU +P VJKU
YJGTG YC[ VJG WPEGTVCKPV[ QH ECP DG EJCTCEVGTK\GF D[ VJG WPEGTVCKPV[ QH 6JGP VJG minimax decision rule YKVJ TGURGEV VQ VJG CDQXG YKNN DG
·· ¼
E\&5&3UHVV//&
¼
6JG UQECNNGF model-space stochastic matching OGVJQF FGUETKDGF KP = ? ECP DG VJGQTGVKECNN[ LWUVKſGF KP VJKU YC[
3.6.3 Bayesian Predictive Classification Rule #U + FKUEWUUGF DGHQTG OKPKOCZ ENCUUKſECVKQP VTKGU VQ JCPFNG VJG YQTUV ECUG OKUOCVEJ D[ CUUWOKPI C WPKHQTO FKUVTKDWVKQP KP VJG WPEGTVCKPV[ PGKIJDQTJQQF HQT CNN RQUUKDNG FGXKCVKQPU HTQO VJG PQOKPCN RCTCOGVGTU ¼ +PUVGCF QH CUUKIPKPI CPQVJGT RQKPV GU CU FQPG KP VJG OKPKOCZ ENCUUKſECVKQP TWNG FKUEWUUGF CDQXG QPG ECP CNUQ VKOCVG average out VJG GHHGEV QH VJG RQUUKDNG OQFGNKPI CPF GUVKOCVKQP GTTQTU D[ CUUWOKPI C IGPGTCN RTKQT 2&( HQT VQ EJCTCEVGTK\G VJG RCTCOGVGT XCTKCDKNKV[ YJKNG OCMKPI ENCU UKſECVKQP FGEKUKQPU +P VJKU YC[ C PGY TQDWUV FGEKUKQP UVTCVGI[ ECP DG FGTKXGF CPF KU QHVGP TGHGTTGF VQ CU C Bayesian predictive classification $2% TWNG = ? .GV WU EQPUKFGT VJG WPEGTVCKPV[ QH VJG OQFGN RCTCOGVGTU D[ VTGCVKPI VJGO CU KH VJG[ YGTG TCPFQO 1WT RTKQT MPQYNGFIG CDQWV KU CUUWOGF VQ DG UWOOCTK\GF ´¼µ ´¼µ KP C MPQYP LQKPV a priori FGPUKV[ £ YKVJ £ CPF ´¼µ ´¼µ YJGTG £ CPF FGPQVG VJG CFOKUUKDNG TGIKQPU QH CPF CPF £ CTG VJG UGVU QH RCTCOGVGTU QH VJG RTKQT 2&( QHVGP TGHGTTGF VQ CU hyperparameters YJKEJ CTG CUUKIPGF XCNWGU D[ VJG KPXGUVKICVQT 5WEJ RTKQT KPHQTOCVKQP OC[ HQT GZCORNG EQOG HTQO UWDLGEV OCVVGT EQPUKFGTCVKQPU CPFQT HTQO RTGXKQWU GZRGTKGPEGU (QT VJG UKORNKEKV[ QH VJG HQNNQYKPI FKUEWUUKQP NGV WU HWTVJGT CUUWOG VJCV
´¼µ ´¼µ £
´¼µ £
´¼µ
)KXGP C VTCKPKPI UGV CU FGUETKDGF CV VJG DGIKPPKPI QH 5GEVKQP VJG WPEGT ´¼µ ´¼µ VCKPV[ CDQWV ECP DG TGFWEGF D[ GXQNXKPI £ #RRCTGPVN[ VJGTG CTG OCP[ YC[U VQ GXQNXG YJKEJ FGRGPF QP VJG RWTRQUG QH VJG OQFGNKPI KP OKPF VJG MPQYNGFIGKPHQTOCVKQP UQWTEGU WUGF CPF VJG RQUUKDNG EQPUVTCKPVU KORQUGF = ? +V KU CV VJKU RQKPV VJCV QWT RTQRQUCN FGRCTVU HTQO VJG EQPXGPVKQPCN VTGCV ´¼µ ´¼µ OGPV KP UVCVKUVKEU %QPXGPVKQPCNN[ £ KU GXQNXGF D[ EQPUVTWEVKPI VJG HQNNQYKPI RQUVGTKQT 2&(
´¼µ £ ´¼µ £ ª ª
Ê
Ê
´¼µ ´¼µ
YJGTG
ª
Ê
Ê ª
´¼µ £ ´¼µ £ ´¼µ ´¼µ
6JKU RQUVGTKQT 2&( KPENWFGU CNN QH VJG KPHQTOCVKQP KPJGTKVGF HTQO VJG RTKQT ´¼µ ´¼µ MPQYNGFIG £ CPF NGCTPGF HTQO VJG VTCKPKPI FCVC %QPXGPVKQPCNN[
E\&5&3UHVV//&
HTQO GI /#2 GUVKOCVG CPF VJGP QPG FGTKXGU C point estimate WUG VJG RNWIKP /#2 FGEKUKQP TWNG HQT TGEQIPKVKQP 6JG EQPXGPVKQPCN RNWIKP /#2 FGEKUKQP TWNG DCUGF QP VJG /. GUVKOCVG QH VJG OQFGN RCTCOGVGTU ECP DG VTGCVGF CU C URGEKCN ECUG QH VJG CDQXG /#2 GUVKOCVG YKVJ C PQPKPHQTOCVKXG RTKQT *QYGXGT KP QWT RTQRQUCN YG FQ PQV WUG OGEJCPKECNN[ +PUVGCF YG CFQRV C OQTG ƀGZKDNG empirical Bayes CRRTQCEJ KP YJKEJ C URGEKſE RCTCOGVTKE 2&(
£ £
KU WUGF VQ TGRTGUGPV QWT WPEGTVCKPV[ CDQWV CHVGT QDUGTXKPI 6JG KPVTCEVCDKNKV[ QH FKTGEVN[ ECNEWNCVKPI HQT VJG RQRWNCT OQFGNU KP #54 UWEJ CU *//U KU PQV VJG QPN[ TGCUQP HQT VJG CDQXG RTQRQUCN # OQTG KORQTVCPV TGCUQP KU VJCV WUKPI £ KPUVGCF QH VQ TGRTGUGPV VJG prior uncertainty YKVJ TG URGEV VQ TGEQIPK\KPI RTQXKFGU C ƀGZKDNG YC[ VQ KPEQTRQTCVG CPF OCMG WUG QH QVJGT ´¼µ ´¼µ MPQYNGFIG UQWTEGU YJKEJ OC[ DG CXCKNCDNG KP CFFKVKQP VQ CPF £ CPFQT GXGP VQ EQPUKFGT VJG OQFGNKPI KPVGPVKQP (QT GZCORNG VJG UGV QH J[RGTRC TCOGVGTU £ EQWNF DG GUVKOCVGF HTQO VTCKPKPI FCVC QT URGEKſGF WUKPI UQOG GORKTKECN TGCUQPKPI QT VJGKT EQODKPCVKQP = ? (WTVJGTOQTG KH VJG VTCKP KPI FCVC KU PQV representative GPQWIJ VJGP UWEJ NGCTPGF OKIJV PQV DG KPHQTOCVKXG GPQWIJ VQ JGNR TGEQIPK\G +P VJKU ECUG DGHQTG KPXQMKPI VJG TGEQIPK VKQP RTQEGUU YG ECP ſTUV KORTQXG D[ WUKPI VJG KPHQTOCVKQP GODGFFGF KP VJG QDUGTXCVKQP KVUGNH +P CP[ ECUG YG WUG £ IGPGTKECNN[ VQ TGRTGUGPV QWT prior uncertainty CDQWV LWUV DGHQTG OCMKPI VJG FGEKUKQP QP +P VJKU YC[ YG CTG GUUGPVKCNN[ EQPUKFGTKPI VJG HQNNQYKPI CFOKUUKDNG FKUVQTVGF UGV QH FCVC OQFGN £
£
£
£ £
YJGTG YG ECP XKGY VJG CU C RCTCOGVGT VQ EJCTCEVGTK\G VJG DTQCFPGUU QH VJG FKUVTK DWVKQP £ QT GSWKXCNGPVN[ VJG FGITGG QH VJG FKUVQTVKQP $CUGF QP VJG CDQXG £ VJG overall risk KU
Ï Ü
´µ ´£ µ
¾
ª
¾
ªÏ
¾ª
ª ª
¾ªÜ
YJGTG
ª
ª
£
£
CTG ECNNGF predictive densities = ? DGECWUG YG ECP XKGY £ CU C HWPEVKQP QH VTCKPKPI UCORNGU 6JGP WPFGT VJG NQUU HWPEVKQP VJG RTGFKEVKXG
E\&5&3UHVV//&
FGEKUKQP TWNG YJKEJ OKPKOK\GU VJG CDQXG KU CU HQNNQYU
Ï
Ï
6JKU FGEKUKQP TWNG YKNN DG TGHGTTGF VQ CU VJG Bayesian predictive classification (BPC) rule 6JTGG MG[ KUUWGU VJWU CTKUG KP $2% PCOGN[ VJG FGſPKVKQP QH VJG RTKQT FGPUKV[ £ HQT OQFGNKPI VJG WPEGTVCKPV[ QH VJG OQFGN RCTCOGVGTU VJG URGEKſECVKQP QH VJG J[RGTRCTCOGVGTU £ CPF VJG GXCNWCVKQP QH VJG RTGFKEVKXG FGPUKV[ +P VJG RCUV UGXGTCN [GCTU UQOG URGEKſE VGEJPKSWGU JCXG DGGP FGXGNQRGF VQ CFFTGUU VJG CDQXG KUUWGU CPF UQOG GPEQWTCIKPI TGUWNVU JCXG DGGP QDVCKPGF 4GCFGTU CTG TGHGTTGF VQ = ? HQT FGVCKNU +P GZVGPFKPI VJG CDQXGHQTOWNCVGF $2% CRRTQCEJ VJGTG CTG VYQ QRRQUKVG FKTGEVKQPU YJKEJ ECP DG RWTUWGF 1PG FKTGEVKQP KU VQ WUG C UVTWEVWTG OQFGN HQT OQFGNKPI RCTCO GVGT WPEGTVCKPV[ (QT GZCORNG YG ECP WUG
£
£ ¼ ¼ £ £
YJGTG £ ¼ CPF ¼ FGPQVG URGEKſE VTCPUHQTOCVKQPU QH ¼ CPF ¼ YKVJ RC TCOGVGTU £ CPF TGURGEVKXGN[ +P VJKU YC[ VJG WPEGTVCKPV[ QH ECP DG EJCT CEVGTK\GF D[ VJG WPEGTVCKPV[ QH £ 6JG $2% FGEKUKQP TWNG KP ECP VJGP DG OQFKſGF WUKPI VJG HQNNQYKPI RTGFKEVKXG 2&(U
ª£ ª
£ ¼ £ £
¼
6JG CDQXG KUUWG QH RTKQT URGEKſECVKQP YKNN VJGP DG VTCPUNCVGF KPVQ VJG URGEKſECVKQP QH £ CPF 4GCFGTU CTG TGHGTTGF VQ =? HQT C TGEGPV YQTM CNQPI VJKU NKPG QH VJQWIJV #PQVJGT RQUUKDNG FKTGEVKQP VQ RWTUWG KU VQ IQ DG[QPF VJG model parameter uncertainty D[ EQPUKFGTKPI VJG CFOKUUKDNG FKUVQTVGF FGPUKVKGU CV VJG FKUVTKDWVKQP NGXGN
½ £¼ ½ ½ ¾ ¼ ¾ ¾ ½ ¾
6JWU VJG FKUVQTVGF FGPUKV[ CPF KU C OKZVWTG QH VJG J[RQVJGVKECN FKUVTKDWVKQP £¼ CPF ¼ CPF CP CTDKVTCT[ FKUVTKDWVKQP ½ CPF ¾ FGUETKDKPI VJG RQUUKDNG FKUVQTVKQPU 6JKU V[RG QH FKUVQTVKQP OQFGN KU VJG OQUV RQRWNCT QPG KP TQDWUV UVCVKUVKEU = ? *QY VQ FGTKXG VJG TGNGXCPV TQDWUV FGEKUKQP TWNG WPFGT VJKU FKUVQTVKQP OQFGN TGOCKPU CP KPVGTGUVKPI RTQDNGO HQT HWVWTG TGUGCTEJ
E\&5&3UHVV//&
3.6.4 Discussion 6JG ETWEKCN FKHHGTGPEG DGVYGGP VJG RNWIKP CPF RTGFKEVKXG ENCUUKſGTU KU VJCV VJG HQTOGT CEVU CU KH VJG GUVKOCVGF OQFGN RCTCOGVGTU YGTG VJG VTWG QPGU YJGTGCU VJG RTGFKEVKXG OGVJQFU CXGTCIG QXGT VJG WPEGTVCKPV[ KP RCTCOGVGTU *QYGXGT KH YG WUG VJG RQUVGTKQT 2&( KP 'S FGTKXGF HTQO VJG training set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ſECVKQP CRRTQCEJ TGEGKXGU NKVVNG CVVGPVKQP KP OCP[ ENCUUKECN UVCVKUVKEU VGZVDQQMU FGURKVG VJG GZKUVGPEG QH OCP[ IQQF YQTMU = ? #U RQKPVGF QWV D[ 4KRNG[ =? VJKU OC[ DG DGECWUG KV WUWCNN[ OCMGU NKVVNG FKHHGTGPEG HTQO RNWIKP CRRTQCEJGU YKVJKP VJG RTQDNGOU CPF VJG VKIJVN[EQPUVTCKPGF RCTCOGVTKE HCOKNKGU OCP[ UVCVKUVKEKCPU WUG QT EQPUKFGT 0QPGVJGNGUU KV YKNN DGEQOG KORQTVCPV YJGP YG EQPUKFGT OWEJ NCTIGT HCOKNKGU CPF HQTOWNCVG VJG RTQDNGO CRRTQRTKCVGN[ CU UJQYP DGHQTG 6Q QWT MPQYNGFIG KV YCU 0CFCU YJQ ſTUV
CFQRVGF C $2% HQTOWNCVKQP CPF RQKPVGF QWV KVU RQVGPVKCN KP URGGEJ TGEQIPKVKQP CRRNKECVKQPU =? CPF GZRNKEKVN[ UVCVGF CPF RTQXGF VJG QRVKOCNKV[ QH $2% KP VJG UGPUG QH OKPKOK\KPI QXGTCNN TKUM 4KRNG[ CNUQ FKUEWUUGF VJG RTGFKEVKXG ENCUUKſECVKQP CRRTQCEJ KP VJKU YC[ KP JKU TGEGPV RCVVGTP TGEQIPKVKQP VGZVDQQM =?
*QYGXGT NKMG QVJGT UVCVKUVKEKCPU 0CFCU YCU FKTGEVN[ WUKPI VJG RQUVGTKQT 2&( VQ UGTXG CU VJG RTKQT 2&( KP RTGFKEVKXG FGEKUKQP OCMKPI CPF ICXG C UKORNG GZCORNG KP YJKEJ C reproducing density GZKUVU 0Q GZRGTKOGPVCN TGUWNVU YGTG TGRQTVGF CPF VJG RCRGT ENQUGF D[ DTKGƀ[ FKUEWUUKPI VJG FKHſEWNV[ QH CRRN[KPI VJG VJGQT[ VQ *// DCUGF URGGEJ TGEQIPKVKQP 5VCTVKPI HTQO 0CFCUŏU HQTOWNCVKQP /GTJCX CPF 'RJTCKO =? UWIIGUVGF C UQECNNGF approximate Bayesian (AB) decision rule HQT URGGEJ TGEQIPKVKQP YJKEJ YCU DCUGF QP VJG IGPGTCNK\GF NKMGNKJQQF TCVKQU EQORWVGF HTQO VJG CXCKNCDNG VTCKPKPI CPF VGUVKPI FCVC 5WEJ CP #$ TWNG QRGTCVGU CU HQNNQYU
Ï
¼
+V KU ENGCT VJCV KH VJG VTCKPKPI UGSWGPEGU CTG EQPUKFGTCDN[ NQPIGT VJCP VJG VGUV UG SWGPEG YJKEJ KU VJG ECUG KP OQUV URGGEJ TGEQIPKVKQP CRRNKECVKQPU VJG RCTCOGVGT UGV VJCV OCZKOK\GU VJG FGPQOKPCVQT QH 'S KU XGT[ ENQUG VQ VJG RCTCOGVGT
E\&5&3UHVV//&
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ſECVKQP TWNG ECP DG XKGYGF CU C VYQUVGR RTQEGFWTG CPF KORNGOGPVGF KP 'S (KTUV GCEJ VGUVKPI WVVGTCPEG KU VTGCVGF CU RQUUKDN[ DGNQPIKPI VQ CP[ YQTF UGSWGPEG CPF C EQPUVTCKPGF /. GUVKOCVG QH VJG TGNCVGF RCTCOGVGTU KU QDVCKPGF 6JGP C RNWIKP /#2 TWNG KU WUGF HQT URGGEJ TGEQI PKVKQP D[ WUKPI VJG WRFCVGF RCTCOGVGTU 6JKU KPVWKVKXG KPVGTRTGVCVKQP QRGPU WR VJG RQUUKDKNKVKGU VQ WUG QVJGT GUVKOCVKQP CRRTQCEJGU GI VJG /#2 CRRTQCEJ KP VJG ſTUV UVGR 5WEJ C OQFKſGF OKPKOCZ FGEKUKQP TWNG YQTMU CU HQNNQYU
Ï
¼
YJGTG KU VJG /#2 GUVKOCVG QH (QT VJG EQPXGPKGPEG QH TGHGTGPEG YG ECNN VJKU OQFKſGF OKPKOCZ FGEKUKQP TWNG CU C Bayesian minimax rule VQ GORJCUK\G KVU FKHHGTGPEG HTQO VJG OKPKOCZ CRRTQCEJ KP =? 9G JCXG RTGXKQWUN[ FKUEWUUGF VJG $2% CRRTQCEJ CU C PGY FGEKUKQP TWNG YJKEJ CXGT CIGU QWV VJG UCORNKPI GTTQT KP RCTCOGVGT GUVKOCVKQP # TGNCVGF DWV UKORNGT CRRTQCEJ ECP CNUQ DG WUGF (QT GZCORNG HQT C %&*//DCUGF #54 U[UVGO KPUVGCF QH FKTGEVN[ OQFKH[KPI VJG DCUKE FGEKUKQP TWNG QPG ECP CNUQ CUUWOG VJCV VJG %&*// RCTCOGVGTU CTG WPEGTVCKP 6JGP QPG WUGU VJG Bayesian predictive density QH GCEJ )CWUUKCP OKZ VWTG EQORQPGPV VQ UGTXG CU VJG EQORGPUCVGF FKUVTKDWVKQP QH VJCV EQORQPGPV CPF RNWI VJGUG EQORGPUCVGF FKUVTKDWVKQPU KPVQ VJG RNWIKP /#2 FGEKUKQP TWNG KP 'S 6JG CRRTQCEJ KU VJWU ECNNGF Bayesian predictive density based model compensation OGVJQF QT UJQTVN[ $2/% OGVJQF VQ FKHHGTGPVKCVG KV HTQO VJG $2% TWNG FGſPGF KP 'S +P =? UWEJ CP KFGC KU GZRNQTGF KP VJG EQPVGZV QH $C[GUKCP URGCMGT CFCRVCVKQP YJGTG C )CWUUKCP RTKQT 2&( HQT OGCP XGEVQT KU CFQRVGF +P =? C UKO KNCT KFGC KU CRRNKGF VQ PQKU[ URGGEJ TGEQIPKVKQP YJGTG C WPKHQTO RTKQT 2&( QP C RTGURGEKſGF WPEGTVCKPV[ PGKIJDQTJQQF HQT OGCP XGEVQT KU CFQRVGF /QTG TGEGPVN[ UKOKNCT KFGCU CTG CRRNKGF VQ VJG VTCPUHQTOCVKQPDCUGF OQFGN EQORGPUCVKQP D[ WUKPI VJG RTGFKEVKXG 2&( QH VJG VTCPUHQTOCVKQP RCTCOGVGTU = ?
3.7 Summary +P VJKU EJCRVGT YG JCXG TGXKUKVGF VJG FGEKUKQP VJGQTGVKE HQWPFCVKQP QH VJG OQFGTP #54 VGEJPQNQI[ 9G JCXG GZRNCKPGF UGXGTCN MG[ EQPEGRVU CDQWV VJG QRVKOCN FGEK UKQP TWNG CFCRVKXG FGEKUKQP TWNG CPF TQDWUV FGEKUKQP TWNG 9G JCXG UJQYP JQY VJGUG FGEKUKQP TWNGU ECP DG FGTKXGF WPFGT FKHHGTGPV CUUWORVKQPU CPF QRVKOCNKV[ ETKVGTKC #
E\&5&3UHVV//&
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class YJKEJ EQWNF OGCP FKHHGTGPV VJKPIU CU GZRNCKPGF CV 9G CTG VTGCVKPI GCEJ VJG DGIKPPKPI QH 5GEVKQP +P VJG ECUG QH EQPVKPWQWU URGGEJ TGEQIPKVKQP VCMGU VJG HQTO QH C UGSWGPEG QH QVJGT UOCNNGT NKPIWKUVKE WPKVU UWEJ CU YQTFU KP PQTOCN UGPUG #NN QH VJG FGEKUKQP TWNGU FGUETKDGF KP VJKU EJCRVGT CKO CV CEJKGXKPI VJG OKPKOWO KPUVGCF QH VJG YQTF TGEQIPKVKQP GTTQT TCVG YJKEJ KU WUWCNN[ ENCUUKſECVKQP GTTQT QH WUGF CU C OGCUWTG QH RTCEVKECN #54 RGTHQTOCPEG #RRCTGPVN[ VJGTG KU C OKUOCVEJ JGTG VQQ 7PHQTVWPCVGN[ VJG OGVJQF VQ FGTKXG C FGEKUKQP TWNG YJKEJ CEJKGXGU VJG OKPKOWO YQTF TGEQIPKVKQP GTTQT TCVG RQUUKDN[ D[ WUKPI C NQUUHWPEVKQP DG[QPF ő Œ TGOCKPU CP KPVGTGUVKPI QRGP RTQDNGO 4GCFGTU CTG TGHGTTGF VQ UQOG KPVGTGUVKPI TGEGPV YQTMU VJCV VT[ VQ CVVCEM VJKU RTQDNGO = ? +V KU QWT JQRG VJCV VJG KPFGRVJ FKUEWUUKQPU KP VJKU EJCRVGT OC[ KPURKTG HWTVJGT KPPQ XCVKQPU VJCV YKNN NGCF VQ DGVVGT UQNWVKQPU HQT #54 CPF OCP[ QVJGT RCVVGTP TGEQIPKVKQP CRRNKECVKQPU
Ï
Ï
Ï
Acknowledgement 6JG CWVJQT ITCVGHWNN[ CEMPQYNGFIGU VJG EQPVTKDWVKQPU QH JKU RCUV CPF RTGUGPV EQN NCDQTCVQTU KPENWFKPI % %JCP %* .GG * ,KCPI $ /C CPF , 9W 6JKU YQTM YCU HWPFGF D[ ITCPVU HTQO VJG 4)% QH VJG *QPI -QPI 5#4 2TQLGEV 0WODGTU *-7' CPF *-7'
References =? #EGTQ # Acoustical and Environmental Robustness in Automatic Speech Recognition -NWYGT #ECFGOKE 2WDNKUJGTU =? #ſH[ / CPF 1 5KQJCP ő5GSWGPVKCN PQKUG GUVKOCVKQP YKVJ QRVKOCN HQTIGVVKPI HQT TQDWUV URGGEJ TGEQIPKVKQPŒ Proc. of ICASSP-2001 =? #ſH[ / 1 5KQJCP CPF %* .GG ő7RRGT CPF NQYGT DQWPFU QP VJG OGCP QH PQKU[ URGGEJ CRRNKECVKQP VQ OKPKOCZ ENCUUKſECVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR
E\&5&3UHVV//&
=? #KVEJKUQP , CPF + 4 &WPUOQTG Statistical Prediction Analysis %CODTKFIG 7- %CODTKFIG 7PKXGTUKV[ 2TGUU =? #OCTK 5 ő# VJGQT[ QH CFCRVKXG RCVVGTP ENCUUKſGTUŒ IEEE Trans. on Electronic Computers 8QN '% 0Q RR =? $CJN . 4 ( ,GNKPGM CPF 4 . /GTEGT ő# OCZKOWO NKMGNKJQQF CRRTQCEJ VQ EQPVKPWQWU URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Pattern Analysis and Machine Intelligence 8QN 2#/+ 0Q RR /CTEJ =? $CJN . 4 2 ( $TQYP 2 8 FG 5QW\C CPF 4 . /GTEGT ő/CZKOWO OW VWCN KPHQTOCVKQP GUVKOCVKQP QH JKFFGP /CTMQX OQFGN RCTCOGVGTU HQT URGGEJ TGEQIPKVKQPŒ KP Proc. of ICASSP-86 RR =? $CJN . 4 2 ( $TQYP 2 8 FG 5QW\C CPF 4 . /GTEGT ő'UVKOCVKPI JKFFGP /CTMQX OQFGN RCTCOGVGTU UQ CU VQ OCZKOK\G URGGEJ TGEQIPKVKQP CEEWTCE[Œ IEEE Trans. Speech and Audio Processing 8QN 0Q RR =? $CMGT , - ő5VQEJCUVKE OQFGNKPI HQT CWVQOCVKE URGGEJ WPFGTUVCPFKPIŒ KP Speech Recognition & 4 4GFF[ GF 0GY ;QTM #ECFGOKE RR =? $CMGT , - ő6JG &4#)10 U[UVGO Ō CP QXGTXKGYŒ IEEE Trans. on Acoustics, Speech, and Signal Processing 8QN #552 RR =? $CWO . ' ő#P KPGSWCNKV[ CPF CUUQEKCVGF OCZKOK\CVKQP VGEJPKSWGU KP UVCVKU VKECN GUVKOCVKQP HQT RTQDCDKNKUVKE HWPEVKQPU QH /CTMQX RTQEGUUGUŒ Inequalities 8QN RR =? %JGPICNXCTC[CP 4 ő5RGCMGT CFCRVCVKQP WUKPI FKUETKOKPCVKXG NKPGCT TGITGU UKQP QP VKOGXCT[KPI OGCP RCTCOGVGTU KP VTGPFGF *//Œ IEEE Signal Processing Letters 8QN 0Q RR =? %JKGP ,6 ő%QODKPGF NKPGCT TGITGUUKQP CFCRVCVKQP CPF $C[GUKCP RTGFKEVKXG ENCUUKſECVKQP HQT TQDWUV URGGEJ TGEQIPKVKQPŒ Proc. of Eurospeech-2001 #CN DQTI &GPOCTM 5GRV =? %JKGP ,6 CPF )* .KCQ ő6TCPUHQTOCVKQPDCUGF $C[GUKCP RTGFKEVKXG ENCU UKſECVKQP WUKPI QPNKPG RTKQT GXQNWVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? %JQW 9 ő&KUETKOKPCPVHWPEVKQPDCUGF OKPKOWO TGEQIPKVKQP GTTQT TCVG RCVVGTPTGEQIPKVKQP CRRTQCEJ VQ URGGEJ TGEQIPKVKQPŒ Proceedings of the IEEE 8QN 0Q RR =? &GNRJKP2QWNCV . % /QMDGN CPF , +FKGT ő(TCOGU[PEJTQPQWU UVQEJCUVKE OCVEJKPI DCUGF QP VJG -WNNDCEM.GKDNGT KPHQTOCVKQPŒ Proc. of ICASSP-1998 RR =? &GPI . ő# F[PCOKE HGCVWTG DCUGF CRRTQCEJ VQ VJG KPVGTHCEG DGVYGGP RJQPQN QI[ CPF RJQPGVKEU HQT URGGEJ OQFGNKPI CPF TGEQIPKVKQPŒ Speech Communication 8QN 0Q RR
E\&5&3UHVV//&
=? &WFC 4 1 CPF *CTV 2 ' Pattern Classification and Scene Analysis 0GY ;QTM 9KNG[ =? &WFC 4 1 *CTV 2 ' CPF 5VQTM & ) Pattern Classification PF GF 0GY ;QTM 9KNG[ =? &G /QTK 4 GF Spoken Dialogues with Computers #ECFGOKE 2TGUU =? 'RJTCKO ; # &GODQ CPF . 4 4CDKPGT ő# OKPKOWO FKUETKOKPCVKQP KPHQT OCVKQP CRRTQCEJ HQT JKFFGP /CTMQX OQFGNKPIŒ IEEE Trans. on Information Theory 8QN 0Q RR =? 'RJTCKO ; CPF . 4 4CDKPGT ő1P VJG TGNCVKQPU DGVYGGP OQFGNKPI CRRTQCEJGU HQT URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Information Theory 8QN 0Q RR =? 'RJTCKO ; ő5VCVKUVKECN OQFGN DCUGF URGGEJ GPJCPEGOGPV U[UVGOUŒ Proc. IEEE 8QN 0Q RR =? (GTIWUQP 6 5 Mathematical Statistics: a Decision Theoretic Approach 0GY ;QTM #ECFGOKE 2TGUU =? (WTWK 5 ő4GEGPV CFXCPEGU KP TQDWUV URGGEJ TGEQIPKVKQPŒ Proc. ETRW on Robust Speech Recognition for Unknown Communication Channels 2QPVC /QWUUQP (TCPEG #RTKN RR =? )CNGU / , ( ő2TGFKEVKXG OQFGNDCUGF EQORGPUCVKQP UEJGOGU HQT TQDWUV URGGEJ TGEQIPKVKQPŒ Speech Communication 8QN RR =? )CNGU / , ( ő%NWUVGT CFCRVKXG VTCKPKPI QH JKFFGP /CTMQX OQFGNUŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? )CQ ;3 / 2CFOCPCDJCP CPF / # 2KEJGP[ ő5RGCMGT CFCRVCVKQP DCUGF QP RTGENWUVGTKPI VTCKPKPI URGCMGTUŒ Proc. of Eurospeech-97 4JQFGU )TGGEG RR =? )CQ ;3 ;: .K CPF / 2KEJGP[ ő/CZKOCN TCPM NKMGNKJQQF CU CP QRVKOK\C VKQP HWPEVKQP HQT URGGEJ TGEQIPKVKQPŒ Proc. ICSLP-00 $GKLKPI 1EV =? )GKUUGT 5 ő$C[GUKCP &KUETKOKPCVKQPŒ KP Handbook of Statistics 2 4 -TKUJ PCKCJ CPF . 0 -CPCN GFU 8QN RR =? )GKUUGT 5 Predictive Inference: An Introduction 0GY ;QTM %JCROCP *CNN =? )NKEM 0 ő5CORNGDCUGF ENCUUKſECVKQP RTQEGFWTGU FGTKXGF HTQO FGPUKV[ GUVK OCVQTUŒ Journal of the American Statistical Association 8QN RR =? )NKEM 0 ő5CORNGDCUGF ENCUUKſECVKQP RTQEGFWTGU TGNCVGF VQ GORKTKE FKUVTKDW VKQPUŒ IEEE Trans. on Information Theory 8QN RR
E\&5&3UHVV//&
=? )QGN 8 CPF 9 $[TPG ő/KPKOWO $C[GUTKUM CWVQOCVKE URGGEJ TGEQIPKVKQPŒ Computer Speech and Language 8QN RR =? )QPI ; ő5RGGEJ TGEQIPKVKQP KP PQKU[ GPXKTQPOGPVU C UWTXG[Œ Speech Communication 8QN RR =? )QPI ; ő5VQEJCUVKE VTCLGEVQT[ OQFGNKPI CPF UGPVGPEG UGCTEJKPI HQT EQPVKPW QWU URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? )QQF + , ő6JG RQRWNCVKQP HTGSWGPEKGU QH URGEKGU CPF VJG GUVKOCVKQP QH RQR WNCVKQP RCTCOGVGTUŒ Biometrika 8QN RR =? )QRCNCMTKUJPCP 2 5 & -CPGXUM[ # 0CFCU & 0CJCOQQ CPF / # 2KEJGP[ ő&GEQFGT UGNGEVKQP DCUGF QP ETQUUGPVTQRKGUŒ KP Proc. ICASSP-88 RR =? )QVQJ ; *QEJDGTI / / CPF 5KNXGTOCP * ( ő'HſEKGPV VTCKPKPI CNIQ TKVJOU HQT *//U WUKPI KPETGOGPVCN GUVKOCVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? )WPCYCTFCPC # CPF 9 $[TPG ő&KUETKOKPCVKXG URGCMGT CFCRVCVKQP YKVJ EQP FKVKQPCN OCZKOWO NKMGNKJQQF NKPGCT TGITGUUKQPŒ Proc. Eurospeech-01 #CN DQTI &GPOCTM =? *CORNG ( 4 ' / 4QPEJGVVK 2 , 4QWUUGGWY CPF 9 # 5VCJGN Robust Statistics: The approach Based on Influence Functions 0GY ;QTM ,QJP 9KNG[ 5QPU =? *C\GP 6 , ő# EQORCTKUQP QH PQXGN VGEJPKSWGU HQT TCRKF URGCMGT CFCRVCVKQPŒ Speech Communication 8QN RR =? *GTOCPUM[ * ő5JQWNF TGEQIPK\GTU JCXG GCTU!Œ Speech Communication 8QN RR =? *Q ; % CPF #ITCYCNC # - ő1P RCVVGTP ENCUUKſECVKQP CNIQTKVJOU Ō KPVTQ FWEVKQP CPF UWTXG[Œ IEEE Trans. on Automatic Control 8QN #% RR =? *WCPI , CPF / 2CFOCPCDJCP ő# UVWF[ QH CFCRVCVKQP VGEJPKSWGU QP C XQKEG OCKN VTCPUETKRVKQP VCUMŒ Proc. of Eurospeech-99 =? *WCPI :& #EGTQ # CPF *QP * 9 Spoken language processing: a guide to theory, algorithm, and system development 7RRGT 5CFFNG 4KXGT 0, 2TGPVKEG *CNN =? *WDGT 2 , Robust Statistics 0GY ;QTM ,QJP 9KNG[ 5QPU =? *WQ 3 CPF %* .GG ő1PNKPG CFCRVKXG NGCTPKPI QH VJG EQPVKPWQWU FGPUKV[ JKFFGP /CTMQX OQFGN DCUGF QP CRRTQZKOCVG TGEWTUKXG $C[GU GUVKOCVGŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR
E\&5&3UHVV//&
=? *WQ 3 CPF %* .GG ő1PNKPG CFCRVKXG NGCTPKPI QH VJG EQTTGNCVGF EQPVKPWQWU FGPUKV[ JKFFGP /CTMQX OQFGNU HQT URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? *WQ 3 CPF %* .GG ő# $C[GUKCP RTGFKEVKXG ENCUUKſECVKQP CRRTQCEJ VQ TQDWUV URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? *WQ 3 CPF .GG %* ő4QDWUV URGGEJ TGEQIPKVKQP DCUGF QP CFCRVKXG ENCUUK ſECVKQP CPF FGEKUKQP UVTCVGIKGUŒ Speech Communication 8QN RR =? *WQ 3 CPF $ /C ő4QDWUV URGGEJ TGEQIPKVKQP DCUGF QP QHHNKPG GNKEKVCVKQP QH OWNVKRNG RTKQTU CPF QPNKPG CFCRVKXG RTKQT HWUKQPŒ Proc. ICSLP-2000 $GKLKPI %JKPC 1EVQDGT RR+8 =? *WQ 3 CPF $ /C ő1PNKPG CFCRVKXG NGCTPKPI QH EQPVKPWQWU FGPUKV[ JKFFGP /CTMQX OQFGNU DCUGF QP OWNVKRNGUVTGCO RTKQT GXQNWVKQP CPF RQUVGTKQT RQQN KPIŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR /C[ =? *WQ 3 5OKVJ 0 CPF /C $ ő'HſEKGPV /. VTCKPKPI QH %&*// RCTCOGVGTU DCUGF QP RTKQT GXQNWVKQP RQUVGTKQT KPVGTXGPVKQP CPF HGGFDCEMŒ Proc. ICASSP2000 6WTMG[ RR++ =? ,CKP # - &WKP 4 2 9 CPF /CQ , ő5VCVKUVKECN RCVVGTP TGEQIPKVKQP C TGXKGYŒ IEEE Trans. on Pattern Analysis and Machine Intelligence 8QN RR =? ,GNKPGM ( ő%QPVKPWQWU URGGEJ TGEQIPKVKQP D[ UVCVKUVKECN OGVJQFUŒ Proceedings of the IEEE 8QN 0Q RR #RTKN =? ,GNKPGM ( Statistical Method for Speech Recognition 6JG /+6 2TGUU %CO DTKFIG =? ,GNKPGM ( 4 . /GTEGT CPF 5 4QWMQU ő2TKPEKRNGU QH NGZKECN NCPIWCIG OQF GNKPI HQT URGGEJ TGEQIPKVKQPŒ KP Advances in Speech Signal Processing 5 (WTWK CPF / / 5QPFJK GFU 0GY ;QTM /CTEGN &GMMGT RR =? ,KCPI * CPF . &GPI ő# TQDWUV EQORGPUCVKQP UVTCVGI[ HQT GZVTCPGQWU CEQWU VKE XCTKCVKQPU KP URQPVCPGQWU URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? ,KCPI * - *KTQUG CPF 3 *WQ ő4QDWUV URGGEJ TGEQIPKVKQP DCUGF QP C $C[GUKCP RTGFKEVKQP CRRTQCEJŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? ,KCPI * - *KTQUG 3 *WQ ő+ORTQXKPI 8KVGTDK $C[GUKCP RTGFKEVKXG ENCUUKſ ECVKQP XKC UGSWGPVKCN $C[GUKCP NGCTPKPI KP TQDWUV URGGEJ 4GEQIPKVKQPŒ Speech Communication 8QN 0Q RR
E\&5&3UHVV//&
=? ,KCPI * - *KTQUG 3 *WQ ő# OKPKOCZ UGCTEJ CNIQTKVJO HQT TQDWUV EQPVKP WQWU URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? ,WCPI $* 5 ' .GXKPUQP CPF / / 5QPFJK ő/CZKOWO NKMGNKJQQF GUVK OCVKQP HQT OWNVKXCTKCVG OKZVWTG QDUGTXCVKQPU QH /CTMQX EJCKPUŒ IEEE Trans. on Information Theory 8QN +6 0Q RR =? ,WCPI $* CPF . 4 4CDKPGT ő6JG UGIOGPVCN MOGCPU CNIQTKVJO HQT GUVK OCVKPI RCTCOGVGTU QH JKFFGP /CTMQX OQFGNUŒ IEEE Transactions on Acoustics, Speech, and Signal Processing 8QN 0Q RR =? ,WCPI $* CPF . 4 4CDKPGT ő*KFFGP /CTMQX OQFGNU HQT URGGEJ TGEQIPK VKQPŒ Technometrics 8QN 0Q RR =? ,WCPI $* ő5RGGEJ TGEQIPKVKQP KP CFXGTUG GPXKTQPOGPVUŒ Computer Speech and Language 8QN RR =? ,WCPI $* CPF 5 -CVCIKTK ő&KUETKOKPCVKXG NGCTPKPI HQT OKPKOWO GTTQT ENCU UKſECVKQPŒ IEEE Trans. on Signal Processing 8QN 0Q RR =? ,WCPI $* ő#WVQOCVKE 5RGGEJ 4GEQIPKVKQP 2TQDNGOU 2TQITGUU 2TQURGEVUŒ *CPFQWV QH -G[PQVG 5RGGEJ 1996 IEEE Workshop on Neural Networks For Signal Processing -[QVQ =? ,WCPI $* 9 %JQW CPF %* .GG ő/KPKOWO ENCUUKſECVKQP GTTQT TCVG OGVJQFU HQT URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? ,WPSWC ,% CPF *CVQP ,2 Robustness in Automatic Speech Recognition: Fundamentals and Applications -NWYGT #ECFGOKE 2WDNKUJGTU $QUVQP =? -CPCN . ő2CVVGTPU KP RCVVGTP TGEQIPKVKQP Œ IEEE Trans. on Information Theory 8QN +6 RR =? -CVCIKTK 5 $* ,WCPI CPF %* .GG ő2CVVGTP TGEQIPKVKQP WUKPI C HCOKN[ QH FGUKIP CNIQTKVJOU DCUGF WRQP VJG IGPGTCNK\GF RTQDCDKNKUVKE FGUEGPV OGVJQFŒ Proc. of IEEE 8QN 0Q RR =? -CV\ 5 / ő'UVKOCVKQP QH RTQDCDKNKVKGU HTQO URCTUG FCVC HQT VJG NCPIWCIG OQFGN EQORQPGPV QH C URGGEJ TGEQIPK\GTŒ IEEE Trans. Acoust., Speech, Signal Processing 8QN 0Q RR =? -QUCMC 6 5 /CVUWPCIC CPF 5 5CIC[COC ő5RGCMGTKPFGRGPFGPV URGGEJ TGEQIPKVKQP DCUGF QP VTGGUVTWEVWTGF URGCMGT ENWUVGTKPIŒ Computer Speech and Language 8QN RR =? -WJP 4 ,% ,WPSWC 2 0IW[GP CPF 0 0KGF\KGNUMK ő4CRKF URGCMGT CFCRVC VKQP KP GKIGPXQKEG URCEGŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR
E\&5&3UHVV//&
=? /CPIW . ' $TKNN CPF # 5VQNEMG ő(KPFKPI EQPUGPUWU KP URGGEJ TGEQIPK VKQP YQTF GTTQT OKPKOK\CVKQP CPF QVJGT CRRNKECVKQPU QH EQPHWUKQP PGVYQTMUŒ Computer Speech and Language 8QN RR =? /CPIW . CPF / 2CFOCPCDJCP ő'TTQT EQTTGEVKXG OGEJCPKUOU HQT URGGEJ TGEQIPKVKQPŒ Proc. of ICASSP-2001 =? -GTOQTXCPV % CPF % /QMDGN ő6QYCTFU KPVTQFWEKPI NQPIVGTO UVCVKUVKEU KP /75' HQT TQDWUV URGGEJ TGEQIPKVKQPŒ Proc. of ASRU-1999 =? -JCTKP ; Robustness in Statistical Pattern Recognition $QUVQP -NWYGT #ECFGOKE =? -KO 05 ő0QPUVCVKQPCT[ GPXKTQPOGPV EQORGPUCVKQP DCUGF QP UGSWGPVKCN GUVKOCVKQPŒ IEEE Signal Processing Letters 8QN 0Q RR =? -KO 05 ő+//DCUGF GUVKOCVKQP HQT UNQYN[ GXQNXKPI GPXKTQPOGPVUŒ IEEE Signal Processing Letters 8QN 0Q RR =? -QTMOC\UMK[ ( CPF $* ,WCPI ő&KUETKOKPCVKXG CFCRVCVKQP HQT URGCMGT XGTK ſECVKQPŒ Proc. ICSLP-96 =? .CWTKNC - / 8CUKNCEJG CPF 1 8KKMMK ő# EQODKPCVKQP QH FKUETKOKPCVKXG CPF OCZKOWO NKMGNKJQQF VGEJPKSWGU HQT PQKUG TQDWUV URGGEJ TGEQIPKVKQPŒ Proc. ICASSP-98 RR =? .GG %* (- 5QQPI CPF -- 2CNKYCN GFU Automatic Speech and Speaker Recognition: Advanced Topics $QUVQP -NWYGT #ECFGOKE 2WDNKUJ GTU =? .GG %* ő1P UVQEJCUVKE HGCVWTG CPF OQFGN EQORGPUCVKQP CRRTQCEJGU VQ TQ DWUV URGGEJ TGEQIPKVKQPŒ Speech Communication 8QN RR =? .GG %* CPF *WQ 3 ő1P CFCRVKXG FGEKUKQP TWNGU CPF FGEKUKQP RCTCOGVGT CFCRVCVKQP HQT CWVQOCVKE URGGEJ TGEQIPKVKQPŒ Proceedings of the IEEE 8QN 0Q RR =? .GXKPUQP 5 ' 4CDKPGT . 4 CPF 5QPFJK / / ő#P KPVTQFWEVKQP VQ VJG CRRNKECVKQP QH VJG VJGQT[ QH RTQDCDKNKUVKE HWPEVKQPU QH C /CTMQX RTQEGUU VQ CWVQOCVKE URGGEJ TGEQIPKVKQPŒ The Bell System Technical Journal 8QN 0Q RR =? .GXKPUQP 5 ' ő5VTWEVWTCN /GVJQFU KP #WVQOCVKE 5RGGEJ 4GEQIPKVKQPŒ Proc. IEEE 8QN RR =? .KP %* %* 9W CPF 2% %JCPI ő# UVWF[ QP URGCMGT CFCRVCVKQP HQT /CPFCTKP U[NNCDNG TGEQIPKVKQP YKVJ OKPKOWO GTTQT FKUETKOKPCVKXG VTCKPKPIŒ IEICE Trans. Inf. & Syst. 8QN '& 0Q RR =? .GG -( Automatic Speech Recognition – The Development of the SPHINXSystem -NWYGT #ECFGOKE 2WDNKUJGTU $QUVQP
E\&5&3UHVV//&
=? .KRQTCEG . 4 ő/CZKOWO NKMGNKJQQF GUVKOCVKQP HQT OWNVKXCTKCVG QDUGTXCVKQPU QH /CTMQX UQWTEGUŒ IEEE Trans. on Information Theory 8QN +6 RR =? .LQNLG # ; 'RJTCKO CPF . 4 4CDKPGT ő'UVKOCVKQP QH JKFFGP /CTMQX OQFGN RCTCOGVGTU D[ OKPKOK\KPI GORKTKECN GTTQT TCVGŒ KP Proc. ICASSP-90 RR =? /CVUWK 6 CPF 5 (WTWK ő# UVWF[ QH URGCMGT CFCRVCVKQP DCUGF QP OKPKOWO ENCUUKſECVKQP GTTQT VTCKPKPIŒ Proc. Eurospeech-95 /CFTKF 5GRVGODGT RR =? /GTJCX 0 CPF ; 'RJTCKO ő# $C[GUKCP ENCUUKſECVKQP CRRTQCEJ YKVJ CRRNK ECVKQP VQ URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Signal Processing 8QN 0Q RR =? /GTJCX 0 CPF %* .GG ő# OKPKOCZ ENCUUKſECVKQP CRRTQCEJ YKVJ CRRNKEC VKQP VQ TQDWUV URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? /QQP 5 CPF ,0 *YCPI ő4QDWUV URGGEJ TGEQIPKVKQP DCUGF QP LQKPV OQFGN CPF HGCVWTG URCEG QRVKOK\CVKQP QH JKFFGP /CTMQX OQFGNUŒ IEEE Trans. on Neural Networks 8QN 0Q RR =? 0CFCU # ő# FGEKUKQP VJGQTGVKE HQTOWNCVKQP QH C VTCKPKPI RTQDNGO KP URGGEJ TGEQIPKVKQP CPF C EQORCTKUQP QH VTCKPKPI D[ WPEQPFKVKQPCN XGTUWU EQPFKVKQPCN OCZKOWO NKMGNKJQQFŒ IEEE Trans. on Acoustics, Speech, and Signal Processing 8QN #552 0Q RR =? 0CFCU # ő1RVKOCN UQNWVKQP QH C VTCKPKPI RTQDNGO KP URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Acoustics, Speech, and Signal Processing 8QN #552 0Q RR =? 0CFCU # & 0CJCOQQ CPF / # 2KEJGP[ ő1P C OQFGNTQDWUV VTCKPKPI OGVJQF HQT URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Acoustics, Speech, and Signal Processing 8QN #552 0Q RR =? 0CI[ ) ő5VCVG QH VJG CTV KP RCVVGTP TGEQIPKVKQPŒ Proceedings of the IEEE 8QN RR =? 0G[ * CPF 1TVOCPPU 5 ő2TQITGUU KP F[PCOKE RTQITCOOKPI UGCTEJ HQT .8%54Œ Proceedings of the IEEE 8QN 0Q RR =? 1UVGPFQTH / 8 8 &KICNCMKU CPF 1 # -KODCNN ő(TQO *//ŏU VQ UGIOGPV OQFGNU C WPKſGF XKGY QH UVQEJCUVKE OQFGNKPI HQT URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR
=? 4CDKPGT . 4 , ) 9KNRQP CPF $* ,WCPI ő# UGIOGPVCN OGCPU VTCKPKPI RTQEGFWTG HQT EQPPGEVGF YQTF TGEQIPKVKQPŒ AT&T Tech. Journal 8QN RR
E\&5&3UHVV//&
=? 4CDKPGT . 4 ő# VWVQTKCN QP JKFFGP /CTMQX OQFGNU CPF UGNGEVGF CRRNKECVKQPU KP URGGEJ TGEQIPKVKQPŒ Proceedings of the IEEE 8QN 0Q RR =? 4CDKPGT . 4 CPF ,WCPI $* Fundamentals of Speech Recognition 2TGP VKEG *CNN =? 4KRNG[ $ & Pattern Recognition and Neural Networks %CODTKFIG 7- %CODTKFIG 7PKXGTUKV[ 2TGUU =? 4QDDKPU * CPF * /QPTQG ő# UVQEJCUVKE CRRTQZKOCVKQP OGVJQFŒ Annals of Mathematical Statistics 8QN RR =? 4WUUGNN / , CPF *QNOGU 9 , ő.KPGCT VTCLGEVQT[ UGIOGPVCN *//ŏUŒ IEEE Signal Processing Letters 8QN 0Q RR =? 5CPMCT # CPF %* .GG ő# OCZKOWO NKMGNKJQQF CRRTQCEJ VQ UVQEJCUVKE OCVEJKPI HQT TQDWUV URGGEJ TGEQIPKVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? 5JCJUJCJCPK $ / ő# /CTMQX TCPFQO ſGNF CRRTQCEJ VQ $C[GUKCP URGCMGT CFCRVCVKQPŒ IEEE Trans. on Speech and Audio Processing 8QN 0Q RR =? 5KQJCP 1 CPF # % 5WTGPFTCP ő5VTWEVWTCN $C[GUKCP RTGFKEVKXG CFCRVCVKQP QH JKFFGP /CTMQX OQFGNUŒ Proc. Workshop on Adaptation Methods for Speech Recognition 5QRJKC#PVKRQNKU (TCPEG #WIWUV =? 5VGTP 4 / 4CL $ CPF /QTGPQ 2 , ő%QORGPUCVKQP HQT GPXKTQP OGPVCN FGITCFCVKQP KP CWVQOCVKE URGGEJ TGEQIPKVKQPŒ Proc. ETRW on Robust Speech Recognition For Unknown Communication Channels 2QPVC /QWUUQP (TCPEG #RTKN RR =? 5WTGPFTCP # % %* .GG CPF / 4CJKO ő0QPNKPGCT EQORGPUCVKQP HQT UVQEJCUVKE OCVEJKPIŒ IEEE Trans. on Audio and Speech Processing 8QN 0Q RR =? 5WTGPFTCP # % CPF %* .GG ő6TCPUHQTOCVKQPDCUGF $C[GUKCP RTGFKEVKQP HQT CFCRVCVKQP QH *//UŒ Speech Communication 8QN RR =? 6CMCJCUJK , CPF 5 5CIC[COC ő&KUETKOKPCVKXG VTCKPKPI DCUGF QP OKPKOWO ENCUUKſECVKQP GTTQT HQT C UOCNN COQWPV QH FCVC GPJCPEGF D[ XGEVQTſGNF UOQQVJGF $C[GUKCP NGCTPKPIŒ IEICE Trans. Inf. & Syst. 8QN '& 0Q RR =? 6U[RMKP ; < Adaptation and learning in automatic systems 0GY ;QTM #EC FGOKE 2TGUU =? 6U[RMKP ; < Foundations of the theory of learning systems 0GY ;QTM #EC FGOKE 2TGUU
E\&5&3UHVV//&
=? 7GDGN . CPF 2 % 9QQFNCPF ő+ORTQXGOGPVU KP NKPGCT VTCPUHQTO DCUGF URGCMGT CFCRVCVKQPŒ Proc. ICASSP-01 =? 9GUUGN ( 4 5EJNWVGT CPF * 0G[ ő'ZRNKEKV YQTF GTTQT OKPKOK\CVKQP WUKPI YQTF J[RQVJGUKU RQUVGTKQT RTQDCDKNKVKGUŒ Proc. of ICASSP-2001 =? 9CNNJQHH 4 & 9KNNGVV CPF ) 4KIQNN ő(TCOGFKUETKOKPCVKXG CPF EQPſFGPEG FTKXGP CFCRVCVKQP HQT .8%54Œ Proc. ICASSP-00 =? 9CNF # Statistical Decision Functions 0GY ;QTM 9KNG[ =? 9CPI < CPF ( .KW ő5RGCMGT CFCRVCVKQP WUKPI OCZKOWO NKMGNKJQQF OQFGN KPVGTRQNCVKQPŒ Proc. of ICASSP-99 =? 9W , CPF 3 *WQ ő5WRGTXKUGF CFCRVCVKQP QH /%'VTCKPGF %&*//U WUKPI OKPKOWO ENCUUKſECVKQP GTTQT NKPGCT TGITGUUKQPŒ Proc. ICASSP-2002 1TNCPFQ (NQTKFC /C[ =? ;CQ - - - 2CNKYCN CPF 5 0CMCOWTC ő5GSWGPVKCN PQKUG EQORGPUCVKQP D[ C UGSWGPVKCN -WNNDCEM RTQZKOCN CNIQTKVJOŒ Proc. of Eurospeech-2001 =? ;QWPI 5 -GTUJCY & , 1FGNN , , 1NNCUQP & 8CNVEJGX 8 CPF 9QQFNCPF 2 % The HTK Book Version 3.0 %CODTKFIG 7PKXGTUKV[ %CODTKFIG 'PINCPF
E\&5&3UHVV//&
4 Speech Pattern Recognition using Neural Networks Shigeru Katagiri NTT Communication Science Laboratories
CONTENTS
+PVTQFWEVKQP $C[GU &GEKUKQP 6JGQT[ 5RGGEJ 4GEQIPK\GTU $CUGF QP 0GWTCN 0GVYQTMU (WUKQP QH /WNVKRNG %NCUUKſECVKQP &GEKUKQPU %QPENWFKPI 4GOCTMU 4GHGTGPEGU #RRGPFKZ /CZKOK\KPI /WVWCN +PHQTOCVKQP
4.1 Introduction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ŏU VJGTG YGTG VYQ GRQEJU KP URGGEJ RCVVGTP TGEQIPKVKQP TGUGCTEJ 1PG YCU C RCTCFKIO UJKHV HTQO RCVVGTP OCVEJKPI VQ C PGY RTQDCDKNKUVKE ENCUUKſECVKQP FG EKUKQP RCTCFKIO YJKEJ OCKPN[ TGNKGF QP VJG WUG QH JKFFGP /CTMQX OQFGNU *//U 6JG QVJGT YCU VJG EJCNNGPIG QH GORNQ[KPI C TCRKFN[ ITQYKPI VGEJPQNQIKECN RCTCFKIO ECNNGF PGWTCN PGVYQTMU 00U GI UGG = ? +P VJGUG PGY RCTCFKIOCVKE UVCIGU URGGEJ RCVVGTPU YGTG DCUKECNN[ TGRTGUGPVGF KP VJG UCOG HCUJKQP CU KP VJG ENCUUKE RCV VGTP OCVEJKPI *QYGXGT KP VJGUG PGY UVCIGU ENCUUGU VQ YJKEJ CP KPRWV UJQWNF DG FGEQFGF YGTG OQFGNGF D[ OQTG GNCDQTCVGF UVTWEVWTGU UWEJ CU *//U CPF 00U 6JG RTQDCDKNKUVKE FGEKUKQP HTCOGYQTM KU UWKVCDNG HQT GHſEKGPVN[ OQFGNKPI VJG UVCVKU
E\&5&3UHVV//&
VKECN XCTKCVKQP QH URGGEJ UCORNGU +V KU CNUQ WUGHWN HQT WPKH[KPI FKHHGTGPV V[RGU QH KP HQTOCVKQP UQWTEGU UWEJ CU CEQWUVKE URGGEJ OQFGNU CPF NKPIWKUVKE URGGEJ OQFGNU +P FGGF VJG WUG QH *// JCU ITGCVN[ EQPVTKDWVGF VQ CFXCPEKPI URGGEJ TGEQIPKVKQP VGEJ PQNQIKGU CPF *// KU UVKNN C OCKPUVTGCO EJQKEG CU C TGEQIPK\GT UVTWEVWTG 6JG WUG QH *// KU RTQDCDN[ LWUV QPG QH OCP[ HCEVQTU EQPVTKDWVKPI VQ TGEGPV CFXCPEGU *QY GXGT KV UJQWNF DG CNUQ PQVGF VJCV OQUV QH VJG EWTTGPV TGEQIPK\GTU UWEEGUUHWNN[ GO RNQ[ *//U KP EQOOGTEKCN UGTXKEGU VJCV GXGP GPEQORCUU NCTIGXQECDWNCT[ URGCMGT KPFGRGPFGPV EQPPGEVGFYQTF CPF CEQWUVKECNN[ EQTTWRVGF VGNGRJQPGDCUGF TGEQIPK VKQP VCUMU 00 KU EJCTCEVGTK\GF D[ KVU JKIJ FKUETKOKPCVKXG ECRCDKNKV[ #EVWCNN[ UKPEG VJG CFXGPV QH VJG JKUVQTKE DTGCMVJTQWIJ QH VJG OWNVKNC[GT RGTEGRVTQP /.2 PGVYQTM 00 JCU DGGP GZVGPUKXGN[ CRRNKGF VQ URGGEJ TGEQIPKVKQP KP XCTKQWU HQTOU UWEJ CU VJQTQWIJ DTGF 00DCUGF TGEQIPK\GTU CPF J[DTKFU QH 00 CPF *// 6JTQWIJ XKIQTQWU UVWF KGU KV YCU UJQYP VJCV VJG 00ŏU JKIJ ENCUUKſECVKQP ECRCDKNKV[ OCKPN[ QTKIKPCVGF HTQO KVU FKUETKOKPCVKXG VTCKPKPI OGVJQFU UWEJ CU WUKPI VJG OKPKOK\CVKQP QH VJG USWCTGF GTTQT NQUU QT WUKPI VJG OKPKOK\CVKQP QH VJG ENCUUKſECVKQP GTTQT EQWPV NQUU &WG VQ VJKU URGEKCN HGCVWTG HQT VTCKPKPI 00 KU PQY GUVCDNKUJGF CU C OQFGTP ENCUU QH FGUKIP OGVJQFQNQI[ HQT CEJKGXKPI JKIJRGTHQTOCPEG URGGEJ RCVVGTP TGEQIPK\GTU #U KVU PCOG KORNKGU CP 00DCUGF U[UVGO WUWCNN[ EQPUKUVU QH C PGVYQTM UVTWEVWTG CPF OCP[ UKORNG QRGTCVKQP WPKVU 00 VJWU GUUGPVKCNN[ RQUUGUUGU C FKUVTKDWVGF CPF RCTCNNGN EQORWVCVKQP OGEJCPKUO VJCV KU CFXCPVCIGQWU KP URGGF CPF TQDWUVPGUU QH VJG QRGTCVKQP 5RGEKſECNN[ VJGTG CTG VYQ V[RGU QH TQDWUVPGUU QPG VJCV KPETGCUGU VJG HCWNV VQNGTCPEG QH QRGTCVKQP QH YJKEJ FGITCFCVKQP KU FWG VQ JCTFYCTG VTQWDNGU CPF QPG VJCV KPETGCUGU VJG UVCDKNKV[TGNKCDKNKV[ QH C ENCUUKſECVKQP FGEKUKQP YJKEJ KU C NQPIUVCPFKPI OCVJGOCVKECN TGUGCTEJ KUUWG KP FGEKUKQP VJGQTKGU UWEJ CU VJG $C[GU FG EKUKQP VJGQT[ 6JG NCVVGT KU ENGCTN[ CP KORQTVCPV CNIQTKVJOTGNCVGF KUUWG VJCV UJQWNF DG XKIQTQWUN[ UVWFKGF KP VJG TGUGCTEJ ſGNF QH 00DCUGF RCVVGTP TGEQIPKVKQP #OQPI VJG OCP[ ECPFKFCVG UQNWVKQPU HQT KORTQXKPI FGEKUKQP TGNKCDKNKV[ VJGTG JCU DGGP TKI QTQWU KPXGUVKICVKQP QH EQODKPKPI OWNVKRNG FGEKUKQPU D[ FGEKUKQP HWUKQP VQ QDVCKP C ſPCN TGNKCDNG ENCUUKſ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
E\&5&3UHVV//&
YKVJ CP GORJCUKU QP VJGKT VTCKPKPI RTQEGFWTGU +P 5GEVKQP YG FKUEWUU VJG TQ DWUVPGUU QH 00DCUGF TGEQIPKVKQP RC[KPI CVVGPVKQP VQ VJG KUUWG QH FGEKUKQP HWUKQP (KPCNN[ YG RTQXKFG EQPENWFKPI TGOCTMU KP 5GEVKQP
4.2 Bayes Decision Theory 4.2.1 Preparations (QT FKUEWUUKQP RWTRQUGU YG WUG CP GZGORNCT VCUM UGVVKPI QH TGEQIPK\KPI C URGGEJ RCV VGTP YKVJ C OQFWNCT TGEQIPK\GT EQORQUGF QH C HTQPVGPF HGCVWTG GZVTCEVKQP OQFWNG
HGCVWTG GZVTCEVQT CPF C DCEMGPF ENCUUKſECVKQP OQFWNG ENCUUKſGT 6JG PWODGT QH ENCUUGU VQ YJKEJ OC[ DGNQPI KU 6JG TGEQIPK\GT EQPVCKPU C UGV QH VTCKPCDNG RCTCOGVGTU YJGTG KU VJG UGV QH VTCKPCDNG RCTCOGVGTU QH VJG HGCVWTG GZVTCEVQT CPF KU HQT VJG UCOG ENCUUKſGT 5RGGEJ RCVVGTPU CTG GUUGPVKCNN[ F[PCOKE KG QH C XCTKCDNG NGPIVJ CPF C PQPNKPGCTN[ YCTRKPI VGORQTCN UVTWEVWTG 6Q OCKPVCKP KVU F[PCOKE PCVWTG C URGGEJ RCVVGTP KU WUW CNN[ TGRTGUGPVGF CU C UGSWGPEG QH CEQWUVKE HGCVWTG XGEVQTU GCEJ ECNEWNCVGF D[ UJKHVKPI C UJQTVVKOG YKPFQY QXGT VJG QDUGTXGF RCVVGTP (QT QWT VCUM UGVVKPI YG CUUWOG VJCV CV VJG HGCVWTG GZVTCEVQT KU EQPXGTVGF VQ XGEVQT UGSWGPEG Ü ½ ܾ Ü YJGTG Ü KU VJG FKOGPUKQPCN CEQWUVKE HGCVWTG XGEVQT CV VKOG KPFGZ QXGT CPF KU C ſZGF PWODGT 6JG VCUM QH VJG ENCUUKſGT KU VJGP VQ FGEQFG VQ KVU EQTTGURQPFKPI EQTTGEV ENCUU YJKEJ KU QPG QH VJG RQUUKDNG ENCUUGU QH ½ VJTQWIJ
4.2.2 Decision Rule +PVWKVKQP UWIIGUVU VJCV C PCVWTCN TWNG QH C ENCUUKſECVKQP FGEKUKQP KU VQ ENCUUKH[ CP KPRWV KPVQ KVU OQUV NKMGN[ ENCUU # FKUETKOKPCPV HWPEVKQP KU KPVTQFWEGF VQ OGCUWTG VJKU NKMGNKJQQF (QT UKORNKEKV[ NGV WU CUUWOG VJCV VJG HGCVWTG GZVTCEVQT KU FGVGTOKPGF KPFGRGPFGPVN[ QH VJG ENCUUKſGTŏU FGUKIP )KXGP VJG HGCVWTG GZVTCEVKQP QRGTCVKQP YG HQTOCNN[ FGſPG VJKU ENCUUKſECVKQP QRGTCVKQP CU HQNNQYU
Ü
KHH
YJGTG Ü KU VJG ENCUUKſECVKQP QRGTCVKQP CPF KU VJG FKUETKOKPCPV HWPEVKQP QH VJCV OGCUWTGU VJG FGITGG VQ YJKEJ DGNQPIU VQ +V VWTPU QWV JGTG VJCV CP KOOGFKCVG IQCN QH TGEQIPK\GT FGUKIP KU VQ VTCKP UQ VJCV VJG ENCUUKſGT ECP FGEQFG EQTTGEVN[ *QYGXGT VJG TGEQIPK\GT KU PCVWTCNN[ GZRGEVGF VQ JCPFNG OCP[ KPRWV RCV VGTPU 6JG WNVKOCVG FGUKIP IQCN VJWU DGEQOGU CEJKGXKPI VJG UVCVWU QH VJCV NGCFU VQ VJG OQUV CEEWTCVG ENCUUKſECVKQP QXGT VJG GPVKTG UGV QH CXCKNCDNG KPRWV URGGEJ RCVVGTPU
E\&5&3UHVV//&
4.2.3 Minimum Error-rate Classification 4WNG CPF VJG VGTO QH NKMGNKJQQF KORN[ VJCV YG UJQWNF WUG C RTQDCDKNKV[ HWPEVKQP CU C FKUETKOKPCPV HWPEVKQP 6JGP CUUWOKPI VJCV VJG RTQDCDKNKV[ HQT F[PCOKE RCVVGTPU KU FGſPGF RTQRGTN[ DGEQOGU VJG HQNNQYKPI SWKVG PCVWTCN TWNG
KHH
YJKEJ TGSWKTGU VQ DG ENCUUKſGF CU VJG ENCUU JCXKPI VJG NCTIGUV C RQUVGTKQTK RTQDC DKNKV[ 6JG FGEKUKQP WUKPI NGCFU VQ VJG QRVKOCN OKPKOWO GTTQTTCVG ENCUUKſECVKQP QT KP QVJGT YQTFU VJG OKPKOWO ENCUUKſECVKQP GTTQT RTQDCDKNKV[ EQPFKVKQP YJKEJ KU CP QRVKOCN FGUKIP IQCN KP VJG UVCVKUVKEUDCUGF VJGQT[ GI UGG =?
4.2.4 Probability Function Estimation +H QPG ECP CEEWTCVGN[ GUVKOCVG YKVJ QPG ECP KP RTKPEKRNG CEJKGXG VJG QRVKOCN OKPKOWO GTTQTTCVG ENCUUKſECVKQP +V UJQWNF DG PQVGF JQYGXGT VJCV VJG C RQUVGTKQTK RTQDCDKNKV[ KU QPN[ EQORWVGF QXGT CP KPſPKVG UGV QH UCORNGU CPF VJGTGHQTG UWEJ CP CEJKGXGOGPV WUKPI KU PGCTN[ KORQUUKDNG KP RTCEVKECN FGUKIP RTQDNGOU 0GXGTVJGNGUU OCP[ CVVGORVU JCXG DGGP OCFG VQ CEEWTCVGN[ GUVKOCVG VJG RTQDCDKNKVKGU DCUGF QP VJG UQWPF CPF VTCEVCDNG OCVJGOCVKECN DCUGU QH VJG OCZKOWO NKMGNKJQQF GUVKOCVKQP /.' OGVJQF CPF VJG $C[GUKCP GUVKOCVKQP OGVJQF +P VJGUG OGVJQFU KU WUWCNN[ TGYTKVVGP CU
KHH
YJGTG KU TGRNCEGF D[ VJG ENCUUEQPFKVKQPCN RTQDCDKNKV[ CPF VJG C RTKQTK RTQDCDKNKV[ 6JGUG RTQDCDKNKV[ HWPEVKQPU CTG OQTG UWKVCDNG VQ VJG WUG QH VJG /.' CPF $C[GUKCP OGVJQFU $CUGF QP VJG CEEWOWNCVGF TGUGCTEJ TGUWNVU VJG CRRTQCEJ WUKPI VJG RTQDCDKNKV[ HWPE VKQP GUVKOCVKQP JCU OCFG KORQTVCPV CFXCPEGU *QYGXGT KP TGCNKV[ KV UVKNN UWHHGTU HTQO FKHſEWNV RTQDNGOU CPF VJGTG CTG OCP[ QPIQKPI TGUGCTEJ GHHQTVU (WPFCOGP VCN QRGP SWGUVKQPU CTG UWOOCTK\GF CU HQNNQYU (KTUV HWPEVKQPCN HQTOU UWEJ CU C )CWUUKCP HWPEVKQP QH VJG RTQDCDKNKV[ HWPEVKQPU CTG TCTGN[ MPQYP CPF VJGTGHQTG VJG OKUOCVEJ DGVYGGP C VTWG HWPEVKQPCN HQTO CPF KVU EQWPVGTRCTV HQTO UGNGEVGF HQT GUVK OCVKQP ECWUGU WPCXQKFCDNG GUVKOCVKQP GTTQTU 5GEQPF VJG CRRTQCEJ CUUWOGU VJCV C NCTIG PWODGT QH FGUKIP UCORNGU CTG CXCKNCDNG HQT GUVKOCVKQP YJKNG KV KU QHVGP FKHſEWNV VQ EQNNGEV C UWHſEKGPV PWODGT QH FGUKIP UCORNGU 6JG GUVKOCVKQP RTQEGUU KU GUUGPVKCNN[ DCUGF QP VJG OCVEJKPI DGVYGGP C UGNGEVGFHQTO RTQDCDKNKV[ HWPEVKQP CPF VJG UCORNG FKUVTKDWVKQP CPF VJWU VJG CRRTQCEJ KPGXKVCDN[ TGNKGU QP VJG CDQXG CUUWORVKQP +V VJGP DGEQOGU CP KPVTCEVCDNG O[UVGT[ JQY EQTTWRVGF VJG RTQDCDKNKV[ HWPEVKQP GUVKOCVGU CTG QXGT C NKOKVGF PWODGT QH FGUKIP UCORNGU 6JG NCUV SWGUVKQP KU JQY VJG EQTTWRVGF GUVKOCVGU CHHGEV VJG RGTHQTOCPEG QH 6JG QRVKOCNKV[ QH JQNFU QPN[ KP VJG ECUG QH GTTQTHTGG GUVKOCVKQP QH VJG RTQDCDKNKVKGU $CUKECNN[ VJG OQTG CEEWTCVG VJG
E\&5&3UHVV//&
RTQDCDKNKV[ GUVKOCVGU CTG VJG OQTG CEEWTCVG VJG ENCUUKſECVKQP FGEKUKQP WUKPI KU *QYGXGT VJG GUVKOCVKQP KORTQXGOGPV ECP QEEWT QPN[ KPFKXKFWCNN[ QP VJG RTQDCDKNKV[ HWPEVKQPU GCEJ QH C FKHHGTGPV ENCUU CPF PQ ENGCT RGTURGEVKXG JCU DGGP CXCKNCDNG QP VJG KPVGTCEVKQP KP ENCUUKſECVKQP CEEWTCE[ DGVYGGP VJG RTQDCDKNKV[ GUVKOCVKQP CPF VJG ENCUUKſECVKQP FGEKUKQP +V UJQWNF DG PQVGF JGTG VJCV VJG OGVJQF WUKPI VJG RTQDCDKNKV[ HWPEVKQP GUVKOCVGU KU CP KPFKTGEV CRRTQCEJ VQ KORTQXKPI VJG RGTHQTOCPEG QH VJG GPVKTG FGEKUKQP TWNG CPF VJG KPFKTGEVPGUU KU C HWPFCOGPVCN ECWUG QH VJG NCUV SWGUVKQP
4.2.5 Discriminative Training # OQTG FKTGEV CRRTQCEJ VJCP VJG /.' CPF $C[GUKCP OGVJQFU VQ C UWEEGUUHWN GZG EWVKQP QH VJG ENCUUKſECVKQP FGEKUKQP TGVWTPU VQ 6JG CRRTQCEJ QTKIKPCVGF KP VJG ENCUUKE VTCKPKPI OGVJQF ECNNGF NKPGCT FKUETKOKPCPV CPCN[UKU CPF KV KU WUWCNN[ TGHGTTGF VQ CU FKUETKOKPCVKXG VTCKPKPI +V GOWNCVGU VJG GPVKTG ENCUUKſECVKQP FKUETKOKPCVKQP QR GTCVKQP QH CPF CVVGORVU VQ TGCNK\G C UGV QH FKUETKOKPCPV HWPEVKQPU EQPUGSWGPVN[ VJCV CEJKGXG VJG FGUKTCDNG KP VGTOU QH C RTGUGV VTCKPKPI QDLGEVKXG ETKVGTKQP ENCU UKſECVKQP RGTHQTOCPEG HQT VTCKPKPI UCORNGU 6JG MG[ FGUKIP KUUWGU VJCV FGVGTOKPG VJG RGTHQTOCPEG QH FKUETKOKPCVKXG VTCKPKPI KP ENWFG VJG HQNNQYKPI JQY VQ FGſPG VJG FKUETKOKPCPV HWPEVKQPU JQY VQ GXCNWCVG VJG RGTHQTOCPEG QH FGEKUKQPU KP VJG VTCKPKPI UVCIG JQY VQ CFLWUV VTCKPCDNG U[UVGO RCTCOGVGTU GI CPF JQY VQ EQRG YKVJ WPMPQYP UCORNGU VJCV FQ PQV CRRGCT KP VJG VTCKPKPI UVCIG +V VWTPU QWV VJCV VJGUG RQKPVU EQORTGJGPUKXGN[ EQXGT VJG ENCUUKſGT FGUKIP RTQEGFWTG 6JGUG KUUWGU YKNN DG FGUETKDGF KP VWTP KP NCVGT RCIGU 4.2.5.1 Functional Form Embodiment of the Entire Process 'OWNCVKQP QH VJG ENCUUKſECVKQP RTQEGUU KPGXKVCDN[ TGSWKTGU VJCV VJG RTQEGUU DG TGRTG UGPVGF KP C VTCEVCDNG HWPEVKQPCN HQTO YJKEJ GPCDNGU QPG VQ FGCN YKVJ VJG FGUKIP QH KVU EQTTGURQPFKPI ENCUUKſECVKQP U[UVGO OCVJGOCVKECNN[ $CUGF QP CP QDUGTXCVKQP QH VJG FGEKUKQP RTQEGUU KP QPG ECP UGG VJCV VJG FGEKUKQP EQPUKUVU QH VJG EQORCTKUQP QH VJG FKUETKOKPCPV HWPEVKQPU QXGT CNN RQUUKDNG ENCUUGU # IGPGTCN YC[ QH GOWNCVKPI VJG RTQEGUU KU IKXGP KP VJG HQNNQYKPI HWPEVKQPCN HQTO YJKEJ KU QHVGP TGHGTTGF VQ CU C OKUENCUUKſECVKQP OGCUWTG
YJGTG KU C RQUKVKXG EQPUVCPV 1PG ECP PQVKEG JGTG VJCV IKXGP QH KPFKECVGU C OKUENCUUKſECVKQP CPF KPFKECVGU C EQTTGEV ENCUUKſECVKQP +P CFFKVKQP EQPVTQNNKPI GPCDNGU VJG UKOWNCVKQP QH XCTKQWU FGEKUKQP TWNGU +P RCTVKEWNCT YJGP CRRTQCEJGU ENQUGN[ GOWNCVGU TWNG 4.2.5.2 Discriminant Functions 6JG UGNGEVKQP QH C HWPEVKQPCN HQTO QH VJG FKUETKOKPCPV HWPEVKQP KU DCUKECNN[ GSWKXCNGPV VQ VJG UGNGEVKQP QH C OGCUWTG QT OGCUWTGOGPV WUGF VQ TGRTGUGPV VJG FGITGG VQ YJKEJ
E\&5&3UHVV//&
CP KPRWV UCORNG DGNQPIU VQ UQOG ENCUU +P CFFKVKQP VJG HQTO TGNKGU QP VJG V[RG QH VTCKPCDNG U[UVGO RCTCOGVGTU KG CPF KV KU IGPGTCNN[ UGNGEVGF DCUGF QP VJG PCVWTG QH VJG RCVVGTPU 6[RKECN GZCORNGU QH ENCUUKECN HQTOU HQT ſZGFFKOGPUKQPCN RCVVGTPU KPENWFG VJG NKP GCT FKUETKOKPCPV HWPEVKQP CPF VJG FKUVCPEG 6JG OGCUWTG WUGF KP VJG NKPGCT FKUETKO KPCPV HWPEVKQP KU C NKPGCTN[YGKIJVGF UWO QH KPRWV XGEVQT EQORQPGPVU YJGTG C UGV QH YGKIJVU EQTTGURQPFU VQ VTCKPCDNG U[UVGO RCTCOGVGTU +P VJG FKUVCPEG ECUG VJG OGCUWTG KU C TGCUQPCDN[ UGNGEVGF FKUVCPEG DGVYGGP CP KPRWV CPF C TGHGTGPEG XGEVQT UWEJ CU 'WENKFGCP FKUVCPEG YJGTG VJG TGHGTGPEG XGEVQTU YQTM CU (QT VJG ECUG QH ENCUUKH[KPI F[PCOKE RCVVGTPU UWEJ CU URGGEJ UCORNGU U[UVGO RCTCO GVGTU UJQWNF EQPVCKP UQOG OGCPU QH TGRTGUGPVKPI VJG VGORQTCN UVTWEVWTG # VTCFKVKQPCN UGNGEVKQP GORNQ[U TGHGTGPEG XGEVQTU CPF WUGU C &2DCUGF PQTOCNK\GF FKUVCPEG OGC UWTG DGVYGGP CP KPRWV RCVVGTP CPF C TGHGTGPEG XGEVQT UGSWGPEG 1P VJG QVJGT JCPF KP OQUV TGEGPV ECUGU VTCKPCDNG U[UVGO RCTCOGVGTU CTG VJG EQORQPGPV RTQDCDKNKVKGU QH *// UWEJ CU UVCVG VTCPUKVKQP RTQDCDKNKV[ CPF QDUGTXCVKQP GOKUUKQP RTQDCDKNKV[ 6JG VGORQTCN UVTWEVWTG KU TGRTGUGPVGF VJGTGKP CU VJG UVCVG VTCPUKVKQP UVTWEVWTG +P VJG ECUGU WUKPI 00U C FKUETKOKPCPV HWPEVKQP KU DCUKECNN[ CP QWVRWV QH 00ŏU QWVRWV PQFG 6JG V[RG QH OGCUWTG VJGTGKP TGNKGU QP VJG UGNGEVKQP QH PQFG HWPEVKQPU UWEJ CU RGTEGRVTQPNKMG CEVKXCVKQP HWPEVKQP TCFKCNDCUKU CEVKXCVKQP HWPEVKQP CPF UKIOQKFCN QWVRWV HWPEVKQP 4.2.5.3 Loss over an Individual Pattern 6Q GXCNWCVG FGEKUKQP RGTHQTOCPEG KP VJG FGUKIP UVCIG KPFKXKFWCN NQUU KU KPVTQFWEGF HQT GXGT[ VTCKPKPI RCVVGTP YJKEJ KU C HWPEVKQP QH VJG OKUENCUUKſECVKQP OGCUWTG CPF TGƀGEVU ENCUUKſECVKQP FGITCFCVKQP +V UJQWNF DG PQVGF VJCV KPFKXKFWCN NQUU KU UQOG VKOGU TGHGTTGF VQ CU KPFKXKFWCN TKUM QT KPFKXKFWCN FGUKIP QDLGEVKXG 1DXKQWUN[ VJG UOCNNGT VJG NQUU KU VJG OQTG FGUKTCDNG KVU EQTTGURQPFKPI FGEKUKQP KU # PCVWTCN CPF IGPGTCN HWPEVKQPCN HQTO QH NQUU KU IKXGP HQT URGGEJ RCVVGTP QH CU
YJGTG KU C UECNCT HWPEVKQP VJCV FGVGTOKPGU VJG EJCTCEVGTKUVKEU QH VJG NQUU HWPEVKQP CPF CPF CTG EQPUVCPVU (WTVJGTOQTG KU QH VJG UKIOQKFCN HQTO YJKEJ KU MPQYP CU C UOQQVJGF NQIKUVKE HWPEVKQP CPF KV KU CNUQ C UOQQVJGF XGTUKQP QH VJG ENCUUKſECVKQP GTTQT EQWPV
QVJGTU
4.2.5.4 Loss over Multiple Patterns 6Q FGVGTOKPG VJG RGTHQTOCPEG QH C FGUKIPGF ENCUUKſGT QXGT OWNVKRNG RCVVGTP UCORNGU VJG KPFKXKFWCN NQUU KU PCVWTCNN[ CRRNKGF KP VJG HQNNQYKPI GORKTKECN CXGTCIG NQUU HQTO
E\&5&3UHVV//&
VQ C ſPKVG DWV WUWCNN[ NCTIG UGV QH VTCKPKPI RCVVGTP UCORNGU
KU VJG PWODGT QH VTCKPKPI UCORNGU KP CPF QH GZRNKEKVN[ OGCPU YJGTG VJCV VJG UCORNG KU VJG VJ UCORNG QH #UUWOKPI VJG KPFKXKFWCN NQUU VQ DG
VJKU GORKTKECN CXGTCIG NQUU DGEQOGU C UOQQVJGF XGTUKQP QH VJG VQVCN EQWPV QH ENCUUKſECVKQP GTTQTU OGCUWTGF QXGT
4.2.5.5 Adjustment of Trainable System Parameters 7UKPI VJG GORKTKECN CXGTCIG NQUU VJG FGUKIP RTQEGGFU VQ CP CEVWCN VTCKPKPI RTQEGFWTG QT KP QVJGT YQTFU VJG CFLWUVOGPV QH VTCKPCDNG U[UVGO RCTCOGVGTU # IQCN QH VJG CFLWUVOGPV KU QDXKQWUN[ VQ CEJKGXG VJG UVCVWU QH ENCUUKſGT RCTCOGVGT KP QWT ECUG VJCV TGUWNVU KP VJG OKPKOK\CVKQP QH VJG GORKTKECN CXGTCIG NQUU 7UWCNN[ VJG HWPEVKQPCN HQTO QH VJG GORKTKECN CXGTCIG NQUU KU WPMPQYP CPF KV KU VJWU TCTGN[ RQUUKDNG VQ CEJKGXG VJG VTWG INQDCN OKPKOWO UVCVWU QH VJG NQUU CPCN[VKECNN[ #EEQTFKPIN[ KP OQUV ECUGU VJG CFLWUVOGPV KU HQTOWNCVGF CU CP CU[ORVQVKECN VTCKP KPI OGVJQF VJCV DCUKECNN[ IWCTCPVGGU ſ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robabilistic Descent Theorem]
KU IKXGP CV VTCKPKPI VKOG KPFGZ +H VJG Æ KU URGEKſGF D[ Æ Í
CPF C UGSWGPEG QH RQUKVKXG TGCN PWODGTU KP TGHGTTGF VQ CU NGCTPKPI YGKIJVU UCVKUſGU ½ ½ CPF
#UUWOG VJCV C VTCKPKPI UCORNG ENCUUKſGT RCTCOGVGT CFLWUVOGPV
VJGP VJG RCTCOGVGT CFLWUVOGPV CEEQTFKPI VQ
Æ E\&5&3UHVV//&
EQPXGTIGU YKVJ RTQDCDKNKV[ QPG CV NGCUV VQ YJKEJ TGUWNVU KP C NQECN OKPKOWO QH YJKEJ KU VJG GZRGEVGF NQUU FGſPGF CU HQNNQYU
Í
YJGTG TGRTGUGPVU VJG UVCVG QH CV KU C RQUKVKXGFGſPKVG OCVTKZ CPF KU VJG GPVKTG UCORNG URCEG QH VJG RCVVGTPU +V KU CUUWOGF VJCV CPF VJCV KU CP KPFKECVQT HWPEVKQP 6JG KORQTVCPV RQKPVU JGTG CTG VJCV VJG CFLWUVOGPV CNYC[U CVVGORVU VQ TGOQXG VJG ENCUUKſECVKQP GTTQT ECWUGF D[ C PGYN[ RTGUGPVGF VTCKPKPI UCORNG CPF CUUWOKPI VJG KPFKXKFWCN NQUU VQ DG VJG CFLWUVOGPV ECP TGUWNV KP VJG NQECN OKPKOK\CVKQP QH VJG GZRGEVGF ENCUUKſECVKQP GTTQT EQWPV NQUU KP VJG RTQDCDKNKUVKE UGPUG 6JG VJGQTGO FQGU PQV IWCTCPVGG VJG WPEQPFKVKQPCN OKPKOK\CVKQP QH VJG GZRGEVGF NQUU KG VJG CEJKGXGOGPV QH VJG OKPKOWO ENCUUKſECVKQP GTTQT RTQDCDKNKV[ EQPFKVKQP *QYGXGT KV ENGCTN[ VWTPU QWV VJCV VJG FKUETKOKPCVKXG VTCKPKPI FGUETKDGF KP VJG CDQXG RCTCITCRJU OQTG ENGCTN[ TGUGODNGU VJG $C[GU FGEKUKQP TWNG VJCP VJG RTQDCDKNKV[ HWPEVKQP GUVKOCVKQP CRRTQCEJGU UWEJ CU VJG OCZKOWO NKMGNKJQQF OGVJQF 4.2.5.6 Training Optimality 6JG WNVKOCVG IQCN QH ENCUUKſGT FGUKIP KU VQ ſPF VJG ENCUUKſGT RCTCOGVGT UGV VJCV CEJKGXGU VJG OKPKOWO ENCUUKſECVKQP GTTQT RTQDCDKNKV[ EQPFKVKQP *QYGXGT CU EKVGF CDQXG RCVVGTPU CXCKNCDNG HQT VTCKPKPI CTG WUWCNN[ ſPKVG CPF VJWU KV KU FKHſEWNV VQ FK TGEVN[ CKO CV VJG WNVKOCVG IQCN YJKEJ YQWNF KPGXKVCDN[ TGSWKTG VJG EQORNGVG EQO RWVCVKQP QH VJG RTQDCDKNKV[ TGNCVGF VQ VJG UCORNG FKUVTKDWVKQP QT KVU EQTTGURQPFKPI QDUGTXCVKQP QH CNN RQUUKDNG RCVVGTP UCORNGU #EVWCNN[ VTCKPKPI WUKPI C NKOKVGF PWO DGT QH UCORNGU DCUKECNN[ NGCFU VQ CV OQUV C OKPKOWO ENCUUKſECVKQP GTTQT EQPFKVKQP QXGT VJG ſPKVG RCVVGTP UCORNGU CPF VJKU FGUETKDGU NKVVNG KP C OCVJGOCVKECNN[ TKIQTQWU UGPUG CDQWV VJG ENCUUKſECVKQP RGTHQTOCPEG QXGT WPMPQYP HWVWTG RCVVGTPU 6JGTGHQTG KP QTFGT VQ DTKFIG VJG WNVKOCVG IQCN KG VJG OKPKOWO ENCUUKſECVKQP GTTQT RTQDCDKN KV[ EQPFKVKQP CPF VJG RTCEVKECN FGUKIP CVVGORVU CPCN[UGU QH VJG VTCKPKPI QRVKOCNKV[ QXGT KPſPKVG VTCKPKPI UCORNGU CTG PGGFGF VQ UQOG GZVGPV GXGP VJQWIJ VJG[ CTG QPN[ VJGQTGVKECN (QT GZRNCPCVKQP RWTRQUGU NGV WU CUUWOG VJCV C RTQDCDKNKV[ OGCUWTG KU RTQ XKFGF KP C MPQYP HWPEVKQPCN HQTO HQT RCVVGTP UCORNG CPF C RCTCOGVGT UGV FG VGTOKPKPI VJG HWPEVKQPCN HQTO KU 6JGP EQPUKFGTKPI VJG FKUETKOKPCPV HWPEVKQP
CPF VJG OKUENCUUKſECVKQP OGCUWTG QH YG ECP TGYTKVG VJG GZRGEVGF NQUU VJCV KU
E\&5&3UHVV//&
FGſPGF D[ WUKPI VJG UOQQVJ ENCUUKſECVKQP GTTQT EQWPV NQUU CU
1DXKQWUN[ VJG NCUV GZRTGUUKQP KU GSWKXCNGPV VQ VJG GZRGEVGF GTTQT ECWUGF D[ VJG ENCU UKſECVKQP FGEKUKQP WUKPI VJG GUVKOCVGU QH VJG C RQUVGTKQTK RTQDCDKNKVKGU 0QVG JGTG VJCV VJG RCTCOGVTKE HQTO KU MPQYP DWV KV KU PQV MPQYP YJGVJGT VJG RTGUGPV UVC VWU QH TGUWNVU KP VJG OKPKOWO GTTQTTCVG ENCUUKſECVKQP *GTG VJG FKHHGTGPEG KP VJG PGCT GSWCNKV[ QH QTKIKPCVGU KP VJG UOQQVJPGUU EQPVCKPGF KP VJG GTTQT EQWPV NQUU %QPUGSWGPVN[ D[ EQPVTQNNKPI VJG UOQQVJPGUU QH HWPEVKQPU UWEJ CU VJG PQTO WUGF ENQUGT KP CPF VJG UKIOQKFCN HWPEVKQP WUGF KP YG ECP CTDKVTCTKN[ OCMG VQ VJG NCUV GSWCVKQP KP +V UJQWNF DG PQVGF JGTG VJCV VJGTG KU C NKPM DGVYGGP VJG RTCEVKECN FGUKIP QRGTCVKQP DCUGF QP VJG FKUETKOKPCVKXG VTCKPKPI CPF VJG EQORWVCVKQP QH VJG GZRGEVGF NQUU YJKEJ KU VJG WNVKOCVG VJGQTGVKECN IQCN QH ENCUUKſGT FGUKIP 0GZV NGV WU TGECNN VJCV YG WUG YJQUG HWPEVKQPCN HQTO KU CUUWOGF VQ DG MPQYP $CUGF QP VJKU HCEV CPF CNUQ VJG TGNCVKQP DGVYGGP VJG OKPKOWO GTTQTTCVG ENCUUKſECVKQP CPF KVU EQTTGURQPFKPI C RQUVGTKQTK RTQDCDKNKVKGU VJG UVCVWU QH VJCV EQTTGURQPFU VQ KP YJKEJ KU CEJKGXGF D[ CFLWUVKPI KU ENGCTN[ GSWCN VJG OKPKOWO QH VQ VJG VJCV CEJKGXGU VJG OCZKOWO C RQUVGTKQTK RTQDCDKNKV[ EQPFKVKQP +P UJQTV KV ECP DGEQOG CTDKVTCTKN[ ENQUG VQ VJG VWTPU QWV VJCV VJG OKPKOWO EQPFKVKQP QH KFGCN OKPKOWO GTTQTTCVG EQPFKVKQP VJCV KU CUUQEKCVGF D[ VJG OKPKOWO ENCUUKſECVKQP GTTQT RTQDCDKNKV[
YJGTG KU C RCTVKCN URCEG QH VJCV ECWUGU C ENCUUKſECVKQP GTTQT CEEQTFKPI VQ VJG OCZKOWO C RQUVGTKQTK RTQDCDKNKV[ FGEKUKQP TWNG KG
6JG CDQXG CPCN[UKU QH VJG OKPKOWO GTTQT EQPFKVKQP KU KORQTVCPV HTQO VJG VJGQTGVK ECN XKGYRQKPV QH UJQYKPI VJG TCVKQPCNKV[ QH VJG FKUETKOKPCVKXG VTCKPKPI HQTOCNKUO +V RTQXKFGU C OCVJGOCVKECNN[ UQWPF DCEMITQWPF VQ RTCEVKECN CVVGORVU CV ENCUUKſGT FG UKIP DCUGF QP FKUETKOKPCVKXG VTCKPKPI +P TGCNKV[ JQYGXGT VJG RCTCOGVGT UGV KU TCTGN[ MPQYP CPF VJWU KV KU WUWCNN[ KORQUUKDNG VQ CEJKGXG VJG OKPKOWO ENCUUKſEC VKQP GTTQTTCVG EQPFKVKQP VJTQWIJ FKUETKOKPCVKXG VTCKPKPI QT KP QVJGT YQTFU VJG NQUU OKPKOK\CVKQP QXGT ſPKVG VTCKPKPI UCORNGU
E\&5&3UHVV//&
4.2.5.7 Global Design Scope 9G JCXG CUUWOGF VJCV C HGCVWTG GZVTCEVKQP OQFWNG KU FGVGTOKPGF UGRCTCVGN[ HTQO VJG FGUKIP QH C ENCUUKſGT DWV VJKU CUUWORVKQP KU QPN[ CFQRVGF VQ UKORNKH[ FKUEWUUKQP #EVWCNN[ HGCVWTG GZVTCEVQTU CTG WUWCNN[ FGUKIPGF DCUGF QP UEKGPVKſE GZRGTVKUG CPF UWEJ UGRCTCVG CPF GORKTKECN FGUKIP KU C UVCPFCTF CRRTQCEJ VQ HGCVWTG GZVTCEVQT FGUKIP
CU C TGUWNV TGEQIPK\GT FGUKIP *QYGXGT HTQO VJG HQTOCNKUO KPVTQFWEGF CDQXG QPG ECP GCUKN[ EQPENWFG VJCV VJG FGUKIP UEQRG QH VJG FKUETKOKPCVKXG VTCKPKPI HQT C DCEMGPF ENCUUKſGT ECP DG GZVGPFGF VQ KVU EQTTGURQPFKPI HTQPVGPF HGCVWTG GZVTCEVQT # MG[ KP VJKU GZVGPUKQP KU ENGCTN[ VQ WUG VJG EJCKP TWNG QH ECNEWNWU 4GECNN VJCV VJKU TWNG RNC[U C EGPVTCN TQNG KP VJG FGſPKVKQPU QH VJG OKUENCUUKſECVKQP OGCUWTG CPF VJG KPFKXKFWCN NQUU (TQO VJG FGſPKVKQP QH QWT TGEQIPK\GT UVTWEVWTG C FKUETKOKPCPV HWPEVKQP HQT VJG GPVKTG TGEQIPK\GT ECP DG FGſPGF CU
YJGTG KU CP QWVRWV QH VJG HGCVWTG GZVTCEVKQP OQFWNG %NGCTN[ VJG CFLWUVOGPV VJCV KU RGTHQTOGF HQT VJG NQUU OKPKOK\CVKQP CV VJG NGXGN QH KU GCUKN[ RTQRCICVGF VQ VJG NGXGN QH NGCFKPI VQ C INQDCN QRVKOK\CVKQP QH VJG GPVKTG U[UVGO
4.3 Speech Recognizers Based on Neural Networks 4.3.1 Preparations +P VJG RTGXKQWU UGEVKQP YG UWOOCTK\GF VJG $C[GU FGEKUKQP VJGQT[ KP RCTVKEWNCT VJG FKUETKOKPCVKXG VTCKPKPI VJCV WPFGTNKGU VJG 00DCUGF URGGEJ RCVVGTP TGEQIPKVKQP (TQO VJQUG FGUETKRVKQPU QPG ECP UGG VJCV VJGTG ECP DG XCTKQWU MKPFU QH GODQFKOGPVU CPF KORNGOGPVCVKQPU QH VJG VTCKPKPI (QT GZCORNG D[ EQPVTQNNKPI KP QPG ECP CEJKGXG XCTKQWU KORNGOGPVCVKQPU QH VJG OKUENCUUKſECVKQP OGCUWTG CPF D[ EQPVTQNNKPI CPF KP QPG ECP CNUQ CEJKGXG XCTKQWU UJCRGU QH VJG UOQQVJGF ENCUUKſECVKQP GTTQT EQWPV NQUU +P CFFKVKQP VJGTG CTG QDXKQWUN[ OCP[ QVJGT RQUUKDKNKVKGU QH FGſPKPI KPFKXKFWCN NQUU KP RNCEG QH VJG UKIOQKFCN HQTO NQUU QH #EVWCNN[ KP VJG JKUVQT[ QH 00DCUGF URGGEJ RCVVGTP TGEQIPKVKQP NQUU HWPEVKQPU QVJGT VJCP VJG UOQQVJGF GTTQT EQWPV NQUU JCXG DGGP TCVJGT RQRWNCT 6JG OQUV V[RKECN UGNGEVKQP HTQO UWEJ RQRWNCT HWPEVKQPU KU VJG USWCTGF GTTQT NQUU YJKEJ KU FGſPGF DGVYGGP C ENCUUKſGT QWVRWV CPF KVU EQWPVGTRCTV VCTIGV VGCEJKPI UKIPCN 6JG UGEQPF V[RKECN UGNGEVKQP KU VJG ETQUU GPVTQR[ NQUU FGſPGF D[ WUKPI VJG GUVKOCVGU QH VJG ENCUUEQPFKVKQPCN RTQDCDKNKVKGU 6Q KPVTQ FWEG 00DCUGF URGGEJ RCVVGTP TGEQIPK\GTU YG VJWU PGGF VQ TGHGT VQ VJG FGſPKVKQPU QH VJGUG RQRWNCT NQUU HWPEVKQPU #U RTGXKQWUN[ UVCVGF C URGGEJ UKIPCN KU F[PCOKE 1P VJG QVJGT JCPF C DCUKE UVTWEVWTG QH 00 UWEJ CU C UVCPFCTF OWNVKNC[GT 2GTEGRVTQP PGVYQTM KU UGV HQT JCPFNKPI UVCVKE
ſZGFFKOGPUKQPCN XGEVQT RCVVGTPU +PFGGF QPG QH VJG KORQTVCPV CURGEVU QH 00
E\&5&3UHVV//&
DCUGF URGGEJ TGEQIPKVKQP TGUGCTEJ JCU DGGP VQ EQRG YKVJ VJG FKUETGRCPE[ DGVYGGP VJG F[PCOKE PCVWTG QH URGGEJ UKIPCNU CPF VJG 00ŏU VTCFKVKQPCN UVTWEVWTG YJKEJ KU QPN[ UWKVGF HQT UVCVKE RCVVGTPU +P VJG GCTN[ UVCIG QH TGUGCTEJ UJKHVVQNGTCPV UVTWEVWTGU YGTG GZCOKPGF HQT ENCUUKH[KPI UJQTV UGIOGPVU QH URGGEJ UKIPCNU 0GZV PGVYQTMU JCXKPI DQVJ C UVCPFCTF UVTWEVWTG UWKVGF HQT UVCVKE RCVVGTPU CPF C UJKHVVQNGTCPV UVTWE VWTG YGTG WUGF KP C J[DTKF HQTO YKVJ VJG UVCPFCTF URGGEJ RCVVGTP ENCUUKſGT UVTWEVWTG QH *// +P CFFKVKQP PGVYQTM UVTWEVWTG KVUGNH YCU HWTVJGT GZCOKPGF TGUWNVKPI KP VJG FGXGNQROGPV QH VJG TGEWTTGPV PGVYQTM YJKEJ RQUUGUUGU TGEWTTGPV UKIPCN ƀQYU KP QTFGT VQ TGRTGUGPV VJG VGORQTCN UVTWEVWTG QH URGGEJ 1DXKQWUN[ VJGTG CTG OCP[ RQUUKDKNKVKGU KP EQODKPKPI VJG NQUU HWPEVKQPU CPF VJG 00 UVTWEVWTGU #EVWCNN[ OCP[ V[RGU QH EQODKPCVKQPU JCXG DGGP TGRQTVGF CPF KV KU FKHſ EWNV VQ KPVTQFWEG VJGO EQORTGJGPUKXGN[ KP VJG NKOKVGF URCEG QH VJKU EJCRVGT 6JGTG HQTG YG UGNGEVKXGN[ HQEWU QP UGXGTCN JKUVQTKE ECUGU QH 00DCUGF URGGEJ TGEQIPKVKQP CVVGORVU +PVTQFWEVKQPU YKNN DG QTICPK\GF CNQPI VJG NKPG QH VJG NQUU CPF 00 UVTWEVWTG UGNGEVKQPU
4.3.2 Classification Error Minimization 4.3.2.1 Learning Vector Quantization .GCTPKPI XGEVQT SWCPVK\CVKQP .83 KU QPG QH VJG RKQPGGT 00DCUGF RCVVGTP ENCUUK ſGTU = ? 1TKIKPCNN[ KV YCU FGXGNQRGF KP VJG HTCOGYQTM QH C UGNHQTICPK\KPI HGCVWTG OCR VJCV UKOWNCVGF VJG RJ[UKQNQIKECN TGRTGUGPVCVKQP QH OGOQT[ *QYGXGT KVU DGJCXKQTCN RTKPEKRNG KU UKORNG CPF ECP DG EQPUKFGTGF CP CFCRVKXG VTCKPKPI QH TGHGTGPEG XGEVQTU GCEJ QH YJKEJ DCUKECNN[ TGRTGUGPVU C ENCUU OQFGN KP C RTGUGV FKUVCPEG URCEG 5GXGTCN XGTUKQPU QH .83 JCXG DGGP RTQRQUGF +P VJG URGGEJ TGEQIPKVKQP ſGNF VJG UGE QPF XGTUKQP QH .83 KG .83 JCU DGGP VJG OQUV GZVGPUKXGN[ CRRNKGF =? +P CFFK VKQP VJG FKHHGTGPEGU COQPI VJG XCTKQWU .83ŏU CTG KPUKIPKſECPV HTQO VJG XKGYRQKPV QH FKUETKOKPCVKXG VTCKPKPI HQTOCNKUO 6JGTGHQTG JGTG YG UWOOCTK\G VJG VTCKPKPI RTKP EKRNG QH .83 CPF FKUEWUU KV HTQO VJG NQUU UGNGEVKQP XKGYRQKPV QH VJG FKUETKOKPCVKXG VTCKPKPI # ENCUUKſGT VQ DG VTCKPGF YKVJ .83 CUUWOGU CP KPRWV VQ DG C UVCVKE XGEVQT CPF CNUQ CUUWOGU GCEJ ENCUU VQ DG OQFGNGF D[ OWNVKRNG TGHGTGPEG XGEVQTU GCEJ DGKPI KP VJG UCOG XGEVQT URCEG CU VJG KPRWV +P VJG ENCUUKſECVKQP UVCIG CP WPMPQYP KPRWV XGEVQT KU ENCUUKſGF CU VJG ENCUU QH VJG TGHGTGPEG XGEVQT VJCV JCU VJG UOCNNGUV FKUVCPEG VQ VJCV KPRWV XGEVQT 6JKU ENCUUKſECVKQP UEJGOG OGCPU RCTVKVKQPKPI VJG XGEVQT URCEG KPVQ TG IKQPU FGſPGF D[ KPFKXKFWCN TGHGTGPEG XGEVQTU QT KP QVJGT YQTFU XGEVQT SWCPVK\CVKQP QH VJG QTKIKPCN XGEVQT URCEG .83 VTCKPKPI CFLWUVU VJG TGHGTGPEG XGEVQTU UQ VJCV GCEJ KPRWV XGEVQT JCU C TGHGTGPEG XGEVQT QH VJG TKIJV ENCUU CU KVU ENQUGUV TGHGTGPEG XGEVQT /QTG RTGEKUGN[ .83 VTCKPKPI KU UWOOCTK\GF CU HQNNQYU (QT C IKXGP VTCKPKPI KPRWV XGEVQT Ü QH VJTGG EQPFKVKQPU OWUV DG OGV HQT VTCKPKPI VQ QEEWT VJG PGCTGUV ENCUU OWUV DG KPEQTTGEV VJG PGZVPGCTGUV ENCUU OWUV DG EQTTGEV CPF VJG VTCKP KPI XGEVQT OWUV HCNN KPUKFG C UOCNN U[OOGVTKE YKPFQY FGſPGF CTQWPF VJG OKFRNCPG QH VJG TGHGTGPEG XGEVQTU ¾ DGKPI CP KPEQTTGEV ENCUU CPF ¾ DGKPI VJG EQTTGEV ENCUU +H VJGUG EQPFKVKQPU CTG OGV VJG KPEQTTGEV TGHGTGPEG XGEVQT KU OQXGF
E\&5&3UHVV//&
HWTVJGT CYC[ HTQO VJG KPRWV YJKNG VJG EQTTGEV TGHGTGPEG XGEVQT KU OQXGF ENQUGT CE EQTFKPI VQ
Ü Ö Ö Ö Ü Ö
YJGTG OGCPU VJG UVCVWU QH KVU EQTTGURQPFKPI XGEVQT CV VKOG KPFGZ CPF KU C OQPQVQPKECNN[ FGETGCUKPI UOCNN XCNWG HWPEVKQP QH VJG VKOG KPFGZ 1PG OC[ PQVKEG JGTG VJCV VJG CFLWUVOGPV TWNG QH .83 KU UKOKNCT VQ VJG VTCKPKPI TWNG FGſPGF D[ VJG RTQDCDKNKUVKE FGUEGPV VJGQTGO #EVWCNN[ KV YCU FGOQPUVTCVGF VJCV .83 YCU C JGWTKUVKE CPF OQFKſGF XGTUKQP QH )2& VTCKPKPI WUKPI VJG UOQQVJGF ENCU UKſECVKQP GTTQT EQWPV NQUU KG YJKNG NGVVKPI IQ VQ ½ YKVJ UQOG OQFKſECVKQP KP VJG FGſPKVKQP QH VJG OKUENCUUKſECVKQP OGCUWTG HQT C OWNVKRNG TGHGTGPEG FKUVCPEG ENCUUKſGT =? 5GG =? CPF =? HQT FGVCKNU +P UJQTV .83 KU CP 00OQVKXCVGF FKUETKOKPCVKXG VTCKPKPI OGVJQF VJCV CKOU CV VJG OKPKOK\CVKQP QH VJG UOQQVJGF ENCU UKſECVKQP GTTQT EQWPV NQUU 4.3.2.2 Shift-tolerant LVQ Classifier +P VJG UKORNGUV ECUG CP .83VTCKPGF FKUVCPEG ENCUUKſGT KU CRRNKGF VQ C UVCVKE HGCVWTG XGEVQT YJKEJ KU ECNEWNCVGF HQT GXGT[ VKOG YKPFQY RQUKVKQP QXGT C F[PCOKE KPRWV URGGEJ RCVVGTP (QT GXGT[ HGCVWTG XGEVQT .83 VTCKPKPI KU GZGEWVGF CPF VJG ENCUUKſ ECVKQP WUKPI VJG VTCKPGF TGHGTGPEG XGEVQTU KU RGTHQTOGF *QYGXGT KPFKXKFWCN HGCVWTG XGEVQTU CTG QHVGP KPUWHſEKGPV HQT CEJKGXKPI EQTTGEV URGGEJ RCVVGTP ENCUUKſECVKQP CV C OGCPKPIHWN URGGEJ WPKV NGXGN UWEJ CU RJQPGOG QT YQTF 6Q EQRG YKVJ VJKU KPUWHſ EKGPE[ .83 YKVJ C UJKHVVQNGTCPV CTEJKVGEVWTG YCU ſTUV CRRNKGF VQ RJQPGOG RCVVGTP ENCUUKſECVKQP =? (KIWTG KNNWUVTCVGU VJG CTEJKVGEVWTG QH C UJKHVVQNGTCPV .83 56.83 U[UVGO HQT ENCUUKH[KPI VJTGG RJQPGOG ENCUUGU +V KU CUUWOGF JGTG VJCV CP KPRWV URGGEJ RCVVGTP KU C RTKQTK UGIOGPVGF CPF NCDGNGF YKVJ KVU EQTTGURQPFKPI EQTTGEV RJQPGOG ENCUU +P VJG ENCUUKſGT GCEJ ENCUU KU CUUKIPGF C PWODGT QH TGHGTGPEG XGEVQTU 6JG .83 VTCKPKPI RTQEGFWTG KU CRRNKGF VQ URGGEJ HGCVWTG XGEVQT RCVVGTPU VJCV CTG UVGRRGF VJTQWIJ KP VKOG (QT GXGT[ HGCVWTG XGEVQT RQUKVKQP VJG ENCUUKſECVKQP FGEKUKQP KU GXCNWCVGF D[ WUKPI VJG RJQPGOG NCDGN CPF TGHGTGPEG XGEVQTU CTG CFLWUVGF CEEQTFKPI VQ VJG .83 VTCKPKPI TWNG +P VJG ENCUUKſECVKQP UVCIG HQT CP WPMPQYP KPRWV C UNKIJVN[ FKHHGTGPV RTQEGFWTG VJCP UKORN[ ſPFKPI VJG ENQUGUV XGEVQT VQ VJG KPRWV KU GORNQ[GF 6JG UJKHV VQNGTCPV CTEJKVGEVWTG RTQFWEGU UGXGTCN ENQUGUV TGHGTGPEG XGEVQTU QPG HQT GCEJ YKPFQY RQUKVKQP 6JG RTQEGFWTG ECP DG UWOOCTK\GF CU HQNNQYU HQT GCEJ YKPFQY RQUKVKQP CPF HQT GCEJ ENCUU VJG ENCUUKſGT ECNEWNCVGU VJG FKUVCPEG DGVYGGP VJG KPRWV XGEVQT CPF VJG ENQUGUV TGHGTGPEG XGEVQT YKVJKP QPG ENCUU HTQO VJKU FKUVCPEG OGCUWTG GCEJ ENCUU KU CUUKIPGF CP CEVKXCVKQP XCNWG VJCV KU JKIJ HQT UOCNN FKUVCPEGU NQY HQT NCTIG FKUVCPEGU CHVGT VJG YKPFQY JCU DGGP UJKHVGF QXGT VJG GPVKTG KPRWV RCVVGTP VJG CEVKXCVKQPU ECNEWNCVGF CV GCEJ YKPFQY RQUKVKQP CTG UWOOGF HQT GCEJ ENCUU VJG ENCUU YKVJ VJG JKIJGUV QXGTCNN CEVKXCVKQP KU EJQUGP CU VJG GPEQFGF ENCUU 'ZRGTKOGPVCN GXCNWCVKQPU QH VJG 56.83 ENCUUKſGT CTG TGRQTVGF KP FGVCKN KP =?
E\&5&3UHVV//&
EGJ )LQDODFFXPXODWHG DFWLYDWLRQ
$FWLYDWLRQV RYHUWLPH
E G J
EUHIHUHQFHYHFWRUV
VKLIWLQJRYHU LQSXW
GUHIHUHQFHYHFWRUV JUHIHUHQFHYHFWRUV
/94 WUDLQLQJ LQSXWIHDWXUH YHFWRUVHTXHQFH
VKLIWLQJRYHU LQSXW
FIGURE 4.1 Architecture of shift-tolerant LVQ classifier [20].
4.3.2.3 LVQ/HMM Hybrid Classifier 6JG UJKHVVQNGTCPV CTEJKVGEVWTG KU WUGHWN HQT CNNGXKCVKPI VJG NKOKVCVKQP QH VJG QTKIKPCN .83VTCKPGF RCVVGTP ENCUUKſGT YJKEJ KU UWKVGF QPN[ HQT UVCVKE XGEVQT RCVVGTPU *QY GXGT VJG WUWCN IQCN QH URGGEJ RCVVGTP TGEQIPKVKQP KU VQ GPEQFG C PCVWTCNNGPIVJ URGGEJ KPRWV VQ KVU EQTTGURQPFKPI YQTF QT UGPVGPEG YQTF UGSWGPEG ENCUU 1DXKQWUN[ VJG UJKHVVQNGTCPV CTEJKVGEVWTG KU KPUWHſEKGPV HQT CEJKGXKPI VJKU V[RG QH IQCN # UVTCKIJVHQTYCTF UQNWVKQP KU VQ EQODKPG .83 CPF *// 6TCFKVKQPCNN[ *// JCU DGGP VTCKPGF KP VJG RTQDCDKNKV[ HWPEVKQP GUVKOCVKQP CRRTQCEJ YJGTG VJG ENCUUKſECVKQP RQYGT KU WUWCNN[ RQQTGT VJCP VJG FKUETKOKPCVKXG VTCKPKPI 6JWU C J[DTKF QH .83 CPF *// KU C PCVWTCN CPF RTQOKUKPI EJQKEG .83 KU CP GODQFKOGPV QH FKUETKOKPCVKXG VTCKPKPI 6JGTG CTG VYQ V[RGU QH *// FKUETGVG *// JCXKPI C EQFGDQQM HQT FGCNKPI YKVJ CP KPRWV CU CP QDUGTXCVKQP QH VJG OWNVKPQOKCN FKUVTKDWVKQP CPF EQPVKPWQWU *// FGCNKPI YKVJ CP KPRWV CU CP QDUGTXCVKQP QH C EQPVKPWQWU RTQDCDKNKV[ HWPEVKQP UWEJ CU VJG )CWUUKCP RTQDCDKNKV[ HWPEVKQP $GECWUG VJG EQFGDQQM KU GSWKXCNGPV VQ C RQQN QH TGHGTGPEG XGEVQTU VJG .83*// J[DTKF PCVWTCNN[ WUGU VJG FKUETGVG *// HQT KVU KORNGOGPVCVKQP =? (KIWTG KNNWUVTCVGU CP .83*// J[DTKF URGGEJ RCVVGTP ENCUUKſGT 6JG U[UVGO EQPUKUVU QH C EQFGDQQM CPF C UVCVG VTCPUKVKQP /CTMQX EJCKP 6JG EQFGDQQM EQPVCKPU C PWODGT QH RCKTU QH EQFGU U[ODQNU CPF EQFG XGEVQTU 6JG EQFG XGEVQTU CTG UGV KP VJG UCOG XGEVQT URCEG CU CP KPRWV HGCVWTG XGEVQT CPF VJG[ CTG WUGF VQ GPEQFG VJG
E\&5&3UHVV//&
'LVFUHWH+00
&ODVV
6SHHFK
/94WUDLQHG FRGHYHFWRUV
FIGURE 4.2 Block diagram of LVQ/HMM hybrid classifier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ſECVKQP QH RJQPGOG WPKVU QT UWDRJQPGOKE WPKVU VJG RQUVGPF *// OQFWNG ECP HWPFCOGP VCNN[ OCMG OQTG CEEWTCVG ENCUUKſECVKQP FGEKUKQPU HQT VJG GPVKTG KPRWV WVVGTCPEG YQTF QT YQTF UGSWGPEG RCVVGTP +P =? .83*// YCU KORNGOGPVGF D[ WUKPI .83 CPF KVU JKIJ FKUETKOKPCVKXG RQYGT YCU UWEEGUUHWNN[ FGOQPUVTCVGF 4.3.2.4 HMM/LVQ Hybrid Classifier &KUETKOKPCVKXG RQYGT UJQWNF DG KPEQTRQTCVGF KP C UVCIG VJCV KU CU ENQUG CU RQUUKDNG VQ VJG ſPCN ENCUUKſECVKQP FGEKUKQP QH C TGEQIPK\GT 1DXKQWUN[ C RJQPGOG ENCUUKſGT UJQWNF JCXG JKIJ FKUETKOKPCVKXG RQYGT HQT RJQPGOG ENCUUKſECVKQP C YQTF ENCUUKſGT UJQWNF JCXG JKIJ FKUETKOKPCVKXG RQYGT HQT YQTF ENCUUKſECVKQP CPF C YQTF UGSWGPEG ENCUUKſGT UJQWNF JCXG JKIJ FKUETKOKPCVKXG RQYGT HQT YQTF UGSWGPEG ENCUUKſECVKQP +P NKIJV QH VJKU VJG .83*// ENCUUKſGT KU KPUWHſEKGPV HQT OCMKPI VJG DGUV WUG QH VJG
E\&5&3UHVV//&
&ODVV
&RPSDULVRQ
/94WUDLQHG UHIHUHQFHYHFWRUV
7LPHQRUPDOL]DWLRQEDVHGRQ WKHDYHUDJLQJZLWKLQRQHVWDWH
+00EDVHGVHJPHQWDWLRQ
&
&
&0
6SHHFK
FIGURE 4.3 Block diagram of HMM/LVQ hybrid classifier.
.83ŏU FKUETKOKPCVKXG ECRCDKNKV[ 6JKU EQPEGTP UWIIGUVU C TGXGTUCN QH VJG J[DTKF KFGC KG *//.83 =? (KIWTG KNNWUVTCVGU CP *//.83 J[DTKF ENCUUKſGT *// KU WUGF JGTG VQ PQTOCN K\G VJG PQPNKPGCT VGORQTCN UVTWEVWTG QH URGGEJ KPRWVU CPF CP .83VTCKPGF FKUVCPEG ENCUUKſGT YQTMU CU C U[UVGO QH ENCUUKH[KPI C FWTCVKQPPQTOCNK\GF KPRWV RCVVGTP 6JG U[UVGO KPENWFGU C UGV QH UVCVG *//U QPG *// RGT URGGEJ WPKV VQ DG ENCUUKſGF GI YQTF ENCUU KP CFFKVKQP VQ VJG .83 ENCUUKſGT 4GHGTGPEG XGEVQTU QH VJG .83 ENCUUKſGT CTG UGV KP VJG XGEVQT URCEG QH VJG FWTCVKQPPQTOCNK\GF KPRWV RCVVGTPU 6JG VTCKPKPI RTQEGFWTG QH VJG U[UVGO KU FKXKFGF KPVQ VJG HQNNQYKPI VYQ UWDUGSWGPV RTQEGFWTGU VGORQTCN PQTOCNK\CVKQP CPF .83 VTCKPKPI QH TGHGTGPEG XGEVQTU 6JG OGEJCPKUO QH VGORQTCN PQTOCNK\CVKQP KU CU HQNNQYU *//U CTG VTCKPGF KP C TGIWNCT OCPPGT WUWCNN[ DCUGF QP VJG /.' OGVJQF HQT VJGKT EQTTGURQPFKPI ENCUUGU
GCEJ *// OCMGU UVCVGDCUGF UGIOGPVCVKQP D[ WUKPI VJG 8KVGTDK UGIOGPVCVKQP
VJG CXGTCIG XGEVQT KU ECNEWNCVGF QXGT VJG HGCVWTG XGEVQTU CUUKIPGF VQ QPG UVCVG UQ VJCV KP VJG ECUG QH UVCVG *// CP KPRWV FKOGPUKQPCN XGEVQT UGSWGPEG KU OCRRGF VQ CP FKOGPUKQPCN XGEVQT RTGEKUGN[ OCVTKZ CPF CUUWOG VJCV ENCUU *//U CTG CXCKNCDNG # VKOGPQTOCNK\GF XGEVQT KU CUUKIPGF C ENCUU
E\&5&3UHVV//&
NCDGN QH VJG *// WUGF HQT KVU IGPGTCVKQP 6JGP VKOGPQTOCNK\GF XGEVQTU CTG IGPGTCVGF GCEJ VTGCVGF CU C FKHHGTGPV ENCUU VQMGP .83 VTCKPKPI GURGEKCNN[ .83 KP =? FGUKIPU TGHGTGPEG XGEVQTU UQ VJCV VJG ENCUUKſGT ECP EQTTGEVN[ ENCUUKH[ CNN QH VJG IGPGTCVGF VQMGPU +P VJG ENCUUKſECVKQP UVCIG HQT CP WPMPQYP KPRWV VJG KPRWV KU ſTUV EQPXGTVGF VQ VQMGPU GCEJ NCDGNGF CU C FKHHGTGPV ENCUU 6JGP HQT GXGT[ VQMGP VJG PGCTGUV TGHGT GPEG XGEVQT KU UGNGEVGF CPF VJG FKUVCPEG DGVYGGP VJG PGCTGUV XGEVQT CPF VJG VQMGP KU ECNEWNCVGF $GECWUG VQMGPU CTG IGPGTCVGF JGTG FKUVCPEG XCNWGU CTG ECNEW NCVGF (KPCNN[ C ENCUUKſECVKQP FGEKUKQP KU OCFG HQT VJG KPRWV URGGEJ RCVVGTP D[ WUKPI C YGKIJVGF UWO QH VJG FKUVCPEG XCNWGU 6JG JKIJ FKUETKOKPCVKXG RQYGT QH VJKU J[DTKF ENCUUKſGT YCU UJQYP KP C EQPHWUCDNG #OGTKECP 'TJ[OG RJQPGOG TGEQIPKVKQP VCUM =?
4.3.3 Squared Error Minimization 4.3.3.1 Training Using the Squared Error Loss (QT VJG GZGORNCT
ENCUU VCUM VJG USWCTGF GTTQT NQUU KU FGſPGF CU
YJGTG KU C VGCEJKPI VCTIGV UKIPCN VJCV KU WUWCNN[ UGV HQT C VTCKPKPI UCORNG
VQ
QVJGTYKUG
6JG NQUU KU VJGP TGYTKVVGP CU
*GTG QPG ECP ſPF VJCV VJG DQVVQO NKPG GZRTGUUKQP QH ECP DG VTGCVGF CU C MKPF QH OKUENCUUKſECVKQP OGCUWTG
+V VWTPU QWV VJCV VJG TGFWEVKQP QH VJG USWCTGF GTTQT NQUU TGUWNVU KP VJG TGFWEVKQP QH VJKU OKUENCUUKſECVKQP OGCUWTG CPF CNUQ VJCV VJG OKPKOK\CVKQP QH VJG USWCTGF GTTQT NQUU QH KU GSWKXCNGPV VQ VJG OKPKOK\CVKQP QH C UKORNG NKPGCT NQUU KP YJKEJ VJG
E\&5&3UHVV//&
EGJ RXWSXWOD\HU
FODVVDFWLYDWLRQ OD\HU VKLIWLQJRYHULQSXW
DEVWUDFW OD\HU
VKLIWLQJRYHULQSXW
LQSXWIHDWXUH YHFWRUVHTXHQFH
VKLIWLQJRYHULQSXW
FIGURE 4.4 Architecture of time-delay neural network [27].
OKUENCUUKſECVKQP OGCUWTG KU GODGFFGF 6JWU VJG VTCKPKPI WUKPI VJG NQUU KU EGTVCKPN[ FKUETKOKPCVKXG DWV QDXKQWUN[ FKHHGTGPV HTQO VJG CVVGORV VQ CEJKGXG VJG OKPKOWO ENCUUKſECVKQP GTTQT EQPFKVKQP D[ WUKPI VJG UOQQVJGF GTTQT EQWPV NQUU 5GG =? HQT FGVCKNGF FKUEWUUKQPU 4.3.3.2 Time-delay Neural Network 6JG VKOGFGNC[ PGWTCN PGVYQTM 6&00 KU QPG QH VJG ENCUUKE 00 CRRNKECVKQPU VQ URGGEJ RCVVGTP TGEQIPKVKQP =? 6JG VKOGFGNC[ CTEJKVGEVWTG KU KPEQTRQTCVGF VQ EQRG YKVJ VJG URGGEJ UKIPCN F[PCOKEU D[ WUKPI OWNVKRNG HGCVWTG XGEVQTU GCEJ DGKPI IGP GTCVGF D[ UJKHVKPI C VKOG YKPFQY QXGT CP KPRWV URGGEJ RCVVGTP 6JCV KU VJKU FGUKIP KU C UJKHVVQNGTCPV CTEJKVGEVWTG CPF KV YCU C OQFGN HQT VJG FGXGNQROGPV QH VJG UJKHV VQNGTCPV .83 ENCUUKſGT =? (KIWTG KNNWUVTCVGU C V[RKECN GZCORNG QH VJG 6&00 CTEJKVGEVWTG 6JG 6&00 ENCU UKſGT CUUWOGU CP KPRWV URGGEJ RCVVGTP VQ DG C UGSWGPEG QH HGCVWTG XGEVQTU +V WUGU C NKOKVGFNGPIVJ UVTGCO QH VJG HGCVWTG XGEVQTU CU KVU KPRWV CPF KP QTFGT VQ HGGF HQTYCTF VJG KPHQTOCVKQP QH C UJQTV HQEWUGF UGIOGPV KV EQPUVTCKPVU PGVYQTM EQPPGEVKQPU VQ C UOCNNGT PWODGT QH PQFGU VJCP VJG PWODGT QH HTCOGU KP VJG YJQNG NKOKVGF UVTGCO
HQT GZRNCPCVKQP RWTRQUGU NGV WU ECNN VJKU UGV QH NKOKVGF EQPPGEVKQPU C FGNC[ITQWR 6JG ENCUUKſGT VJGP CEEWOWNCVGU VJG KPHQTOCVKQP HGF HQTYCTF HTQO VJG NQYGT PGVYQTM NC[GTU D[ UJKHVKPI VJG FGNC[ITQWR QXGT VJG KPRWV 6TCKPKPI QH VJG 6&00 ENCUUKſGT KU RGTHQTOGF YKVJ VJG USWCTGF GTTQT NQUU # VTCKPKPI
E\&5&3UHVV//&
c
c c YGKIJVKPI QTŎOKPKOW Oŏ UGCTEJ
FKOQWVRWV
c
c
FKOKPRWV
c
FKOTGHGTGPEG XGEVQT
FIGURE 4.5 Schematic description of distance classifier as a single intermediate layer network (2-dimensional input, 3 references/class, 3 classes).
VCTIGV KU C XGEVQT QH YJKEJ VJG EQORQPGPV HQT VJG EQTTGEV ENCUU KU QPG CPF QH YJKEJ EQORQPGPVU HQT QVJGT ENCUUGU CTG UGV VQ \GTQ 6JGP VJG NQUU KU FGſPGF DGVYGGP VJG VCTIGV CPF CP QWVRWV QH VJG ENCUUKſGT #EEQTFKPIN[ VJG OKPKOK\CVKQP QH VJKU USWCTGF GTTQT NQUU YQTMU UQ VJCV VJG ENCUUKſGT QWVRWV TGUGODNGU VJG VTCKPKPI VCTIGV +P VJG ENCUUKſECVKQP UVCIG HQT WPMPQYP RCVVGTPU VJG ENCUU JCXKPI VJG NCTIGUV PGVYQTM QWVRWV KU UGNGEVGF CU C ENCUUKſECVKQP TGUWNV #U UJQYP KP (KIWTG VJG 6&00 ENCUUKſGT JCU OWNVKRNG KPVGTOGFKCVG NC[GTU 6JKU OWNVKKPVGTOGFKCVGNC[GT UVTWEVWTG KU C URGEKCN HGCVWTG QH OWNVKNC[GT PGVYQTMU UWEJ CU /.2 CPF KV KU ENGCTN[ FKUVKPEV HTQO VJG UVTWEVWTG QH VTCFKVKQPCN ENCUUKſGTU UWEJ CU C FKUVCPEG ENCUUKſGT QPN[ WUKPI TGHGTGPEG XGEVQTU CU ENCUU OQFGNU YJKEJ ECP DG EQPUKFGTGF C UKPING KPVGTOGFKCVG NC[GT PGVYQTM (KIWTG +P CFFKVKQP VJG TGCFGT EQWNF TGECNN VJG ECUG QH CP .83 ENCUUKſGT VJCV JCU QPN[ TGHGTGPEG XGEVQTU # RTKP EKRCN HWPEVKQP QH VJG OWNVKRNG KPVGTOGFKCVG NC[GTU KU KPHQTOCVKQP CDUVTCEVKQP +P VJG 6&00 ENCUUKſGT KV KU GZRGEVGF VJCV KPHQTOCVKQP GUUGPVKCN HQT ENCUUKſECVKQP YJKEJ KU CEEQTFKPIN[ UJKHVVQNGTCPV ECP DG GZVTCEVGF VJTQWIJ VJGUG JKFFGP NC[GTU +P =? JKIJN[ KPƀWGPVKCN GZRGTKOGPVCN TGUWNVU QH WUKPI VJG 6&00 ENCUUKſGT HQT RJQPGOG ENCUUKſECVKQP CTG TGRQTVGF 4.3.3.3 Multi-state Time-delay Neural Network 6Q EQRG YKVJ VJG F[PCOKE PCVWTG QH NQPIGT URGGEJ WPKVU VJG EQPEGRV QH 6&00 YCU FKTGEVN[ GZVGPFGF VQ C OWVNKUVCVG VKOGFGNC[ PGWTCN PGVYQTM /56&00 =? 6JKU
E\&5&3UHVV//&
FGXGNQROGPV EQPVTCUVU YKVJ VJCV QH VJG J[DTKF U[UVGOU WUKPI .83 CPF *// +V ECP DG UCKF VJCV /56&00 KPEQTRQTCVGU VJG UVCVG VTCPUKVKQP UVTWEVWTG YJKEJ YCU YKFGN[ WUGF KP ITCRJKECN OQFGNU UWEJ CU *// KP KVU U[UVGO UVTWEVWTG KP RNCEG QH UKORN[ EQODKPKPI GZKUVKPI EQPEGRVU UWEJ CU 6&00 CPF *// #P /56&00 ENCUUKſGT EQPUKUVU QH C PWODGT QH NQECN 6&00 U[UVGOU GCEJ QH YJKEJ HQTOU C UVCVG CPF KU FGUKIPGF HQT VJG ENCUUKſECVKQP QH RJQPGOGU QT UWDRJQPGOKE WPKVU CPF KV QWVRWVU C ENCUUKſECVKQP TGUWNV HQT VJG GPVKTG URGGEJ KPRWV GI YQTF UGSWGPEG ENCUUKſECVKQP TGUWNV VJTQWIJ VJG &2DCUGF UGCTEJ QH VJG DGUV RJQPGOG UVCVG UGSWGPEG 6TCKPKPI QH VJG ENCUUKſGT KU FQPG YKVJ VJG USWCTGF GTTQT NQUU CU HQT VJG QTKIKPCN 6&00 ENCUUKſGTU *QYGXGT VJG NQUU JGTG KU FGſPGF CV VJG NGXGN QH VJG ſPCN ENCUU QWVRWVU QH VJG VCUM KG YQTF QT YQTF UGSWGPEG ENCUUGU KP RNCEG QH VJG UJQTV RJQPGOG NGXGN CPF CFFKVKQPCNN[ VJG &2DCUGF UGIOGPVCVKQP QXGT VJG GPVKTG VTCKPKPI KPRWV RCVVGTP KU GODGFFGF KP VJG NQUU OKPKOK\CVKQP %NGCTN[ VJG VTCKPKPI QH /56&00 KU UWKVCDNG HQT VJG ENCUUKſECVKQP QH NQPIGT URGGEJ WPKVU +PFGGF CU CP CNVGTPCVKXG VQ VJG UVCP FCTF *//DCUGF ENCUUKſGT VJG /56&00 ENCUUKſGT JCU DGGP WUGF KP XCTKQWU URGGEJ RCVVGTP TGEQIPKVKQP VCUMU =?
4.3.4 Cross Entropy Minimization 4.3.4.1 Training Using the Cross Entropy Loss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
YJGTG KU CUUWOGF VQ DGNQPI VQ CPF KU CP KPRWV VQ VJG VJ QWVRWV PQFG YJKEJ EQTTGURQPFU VQ CPF QH VJG ENCUUKſGT +V KU CU UWOGF VJCV VJKU PQFG RTQFWEGU C UQHVOCZ XCNWG 1PG OC[ PQVG JGTG VJCV VJG UWDVTCEVKQP HQTO KP VJG DQVVQO NKPG QH ECP DG
E\&5&3UHVV//&
EQPUKFGTGF VQ DG C V[RG QH OKUENCUUKſECVKQP OGCUWTG CU HQNNQYU
TGRTGUGPVU CP QRGTCVKQP QH FKUETKOKPCPV HWPEVKQP EQORCTKUQP QXGT VJG RQUUKDNG ENCUUGU CPF VJGTGHQTG KV ECP EGTVCKPN[ DG EQPUKFGTGF C FKUETKOKPCPV HWPEVKQP 6JGP VJG CXGTCIG NQUU VQ DG OKPKOK\GF DGEQOGU
1DUGTXCVKQP JGTG VGNNU QPG VJCV VJKU NQUU ECP CNUQ DG VTGCVGF CU C ECUG QH CRRN[KPI VJG NKPGCT KPFKXKFWCN NQUU HWPEVKQP VQ VJG OKUENCUUKſECVKQP OGCUWTG &WG VQ VJG GORNQ[OGPV QH VJG NKPGCT NQUU VJGTG UGGOU VQ DG C FKUETGRCPE[ DGVYGGP VJG OKPK OK\CVKQP QH VJG ETQUU GPVTQR[ NQUU CPF VJG CEJKGXGOGPV QH VJG OKPKOWO ENCUUKſECVKQP GTTQT RTQDCDKNKV[ EQPFKVKQP 4.3.4.2 Unidirectional Network Classifier 6JG ETQUU GPVTQR[ NQUU YCU WUGF KP CP GCTN[ UVWF[ VJCV RKQPGGTGF VJG CRRNKECVKQP QH TGEWTTGPV PGWTCN PGVYQTMU VQ URGGEJ RCVVGTP TGEQIPKVKQP =? 0CVWTCN UKIPCNU UWEJ CU URGGEJ UKIPCNU QDG[ VJG NCY QH ECWUCNKV[ 6JWU KP RTKPEKRNG VGORQTCN KPHQTOCVKQP IQGU HTQO VJG RCUV VQ VJG HWVWTG CPF CEEQTFKPIN[ C WPKFKTGE VKQPCN RCUVVQHWVWTG QT NGHVVQTKIJV PGVYQTM UVTWEVWTG KU QHVGP WUGF CU C RTKOCT[ UGNGEVKQP HQT OQFGNKPI UWEJ UKIPCNU # V[RKECN UVTWEVWTG QH C WPKFKTGEVKQPCN PGVYQTM KU KNNWUVTCVGF KP (KIWTG +P VJKU ſIWTG CV VKOG KPFGZ KPRWV CEQWUVKE XGEVQT Ù KU RTGUGPVGF VQ VJG PGVYQTM CNQPI YKVJ VJG UVCVG XGEVQT × CPF VJGUG VYQ XGEVQTU RTQ FWEG VJG QWVRWV XGEVQT Ý CPF VJG PGZV UVCVG XGEVQT × # FGUKIP IQCN JGTG KU VQ FGVGTOKPG VYQ YGKIJV OCVTKEGU Ï CPF Î UQ VJCV Ý ECP UCVKUH[ C RTGUGV FGUKIP QDLGEVKXG +P =? VJG ETQUU GPVTQR[ NQUU YCU CRRNKGF VQ C TGEWTTGPV PGVYQTM WUGF CU C NKMGNK JQQF GUVKOCVQT KP EQPLWPEVKQP YKVJ CP *//DCUGF URGGEJ RCVVGTP ENCUUKſGT 6JG PGVYQTM VTCKPKPI HQT OKPKOK\KPI VJKU NQUU YCU FQPG YKVJ C UVCPFCTF VTCKPKPI OGVJQF HQT TGEWTTGPV PGVYQTMU KG VJG DCEMRTQRCICVKQP VJTQWIJ VKOG OGVJQF VJCV GZRCPFU C TGEWTTGPV PGVYQTM KP VKOG QT KP QVJGT YQTFU EQPUKFGTU C TGEWTTGPV PGVYQTM HQT CNN VKOG KPFKEGU CU C UKPING XGT[ NCTIG PGVYQTM YKVJ CP KPRWV CPF QWVRWV CV GCEJ VKOG KPFGZ CPF UJCTGF YGKIJVU QXGT CNN VKOG KPFKEGU 4.3.4.3 Bidirectional Network Classifier +P RTKPEKRNG VJG VGORQTCN EQTTGNCVKQP DCUGF QP ECWUCNKV[ KU TGRTGUGPVGF KP VJG HQTYCTF WPKFKTGEVKQPCN KPHQTOCVKQP ƀQY +P CFFKVKQP VQ VJKU EQTTGNCVKQP URGGEJ UKIPCNU WUWCNN[ RQUUGUU DCEMYCTF FKTGEVKQPCN VGORQTCN EQTTGNCVKQP 6JG URGGEJ UKIPCN KU CP QWVRWV QH C RJ[UKQNQIKECN CTVKEWNCVKQP U[UVGO VJCV KU EQPVTQNNGF D[ C URGGEJ RTQFWEVKQP RNCP YJKEJ RTGRCTGU HWVWTG CTVKEWNCVKQP CPF CEEQTFKPIN[ JCU DCEMYCTF KPƀWGPEG QP VJG
E\&5&3UHVV//&
ut
st
W
yt
V
s(t+1)
Time delay
ut : Input vector st : State vector yt : Output vector
FIGURE 4.6 Architecture of unidirectional network [23]. RCUV CEQWUVKE UVCVG QH VJG URGGEJ UKIPCN 6Q TGRTGUGPV VJGUG VYQ KPHQTOCVKQP ƀ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ſECVKQP
4.4 Fusion of Multiple Classification Decisions 4.4.1 Principles 7UKPI OCP[ FGEKUKQPU KU IGPGTCNN[ OQTG UVCDNG CPF QHVGP OQTG WUGHWN KP VGTOU QH TQ DWUVPGUU VQ WPMPQYP RCVVGTP UCORNGU VJCV FQ PQV CRRGCT KP VJG VTCKPKPIFGUKIP UVCIG
E\&5&3UHVV//&
)RUZDUG VWDWHV
%DFNZDUG VWDWHV W
W
W
2XWSXWQHXURQJURXS +LGGHQVWDWH QHXURQJURXS ,QSXWV *URXSRIZHLJKWV ZLWKLQIRUPDWLRQIORZ
FIGURE 4.7 Architecture of bi-directional network [25].
VJCP WUKPI C UKPING FGEKUKQP +P TGCNYQTNF RCVVGTP ENCUUKſECVKQP RTQDNGOU VTWG UCO RNG FKUVTKDWVKQPU CTG IGPGTCNN[ WPQDUGTXCDNG CPF CEJKGXCDNG ENCUUKſECVKQP FGEKUKQPU CTG OGTGN[ GUVKOCVGU QH VJG VTWG ENCUUKſECVKQP FGEKUKQPU GCEJ TGN[KPI QP KVU EQTTG URQPFKPI VTWG ENCUU DQWPFCTKGU 6JGTGHQTG VJG RTQXGTD DCUKECNN[ JQNFU VTWG GXGP KP VJG UEKGPVKſE HTCOGYQTM QH RCVVGTP ENCUUKſECVKQP CPF VJG EQPEGRV QH FGEKUKQP HW UKQP KG OCMKPI C ſPCN FGEKUKQP D[ EQODKPKPI OWNVKRNG RTGFGEKUKQPU JCU TGEGPVN[ CVVTCEVGF OCP[ TGUGCTEJGTUŏ KPVGTGUVU 6JG UKORNGUV CPF OQUV DCUKE YC[ QH FGEKUKQP HWUKQP KU VQ WUG VJG CXGTCIG QH OWNVKRNG RTGFGEKUKQPU +PVWKVKQP UWIIGUVU VJCV VJG FGEKUKQP OCFG D[ CXGTCIKPI KPFGRGPFGPV RTGFGEKUKQPU KU OQTG UVCDNG KPUGPUKVKXG VQ VJG UGNGEVKQP QH VTCKPKPI UCORNGU QT TQ DWUV VQ WPMPQYP VGUVKPI UCORNGU VJCP VJG KPFKXKFWCN RTGFGEKUKQPU KPENWFGF VJGTGKP #EVWCNN[ DCUGF QP VJG VTCEVCDNG PCVWTG QH VJG USWCTG GTTQT NQUU VJG FGEKUKQP HWUKQP OGVJQFU WUKPI VJKU NQUU HWPEVKQP JCXG DGGP GZVGPUKXGN[ GZRNQTGF KP VJG NKVGTCVWTG GI =? CPF KV JCU DGGP UJQYP VJCV KP RTKPEKRNG VJG FGEKUKQP HWUKQP UEJGOG TGFWEGU GT TQTU KP TGITGUUKQP GUVKOCVKQP #UUWOKPI VJG VCTIGV VQ GUVKOCVG VQ DG VJG OWNVKPQOKCN FKUVTKDWVKQP HWPEVKQP VJCV TGRTGUGPVU VJG ENCUU KPFGZ KPHQTOCVKQP QH UCORNGU VJG CPCN [UKU TGUWNVU QH VJG FGEKUKQP HWUKQP OGEJCPKUO HQT TGITGUUKQP ECUGU ECP DG CRRNKGF VQ VJG ECUGU QH ENCUUKſECVKQP *QYGXGT KV UJQWNF DG TGECNNGF VJCV VJGTG KU C FKUETGRCPE[ DGVYGGP VJG VTCKPKPI YKVJ VJG OKPKOK\CVKQP QH VJG USWCTGF GTTQT NQUU CPF VJG CEJKGXG OGPV QH VJG OKPKOWO ENCUUKſECVKQP GTTQT RTQDCDKNKV[ EQPFKVKQP 6JWU KV UGGOU VJCV
E\&5&3UHVV//&
&ODVVERXQGDU\
&ODVVLILHU
&ODVVLILHU
7UDLQLQJVHW
7UDLQLQJVHW
&ODVVERXQGDU\=
&ODVVERXQGDU\
&ODVVLILHU=
7UDLQLQJVHW=
0RWKHUVHWRISDWWHUQVDPSOHV
D &ODVVERXQGDU\
&ODVVERXQGDU\
&ODVVLILHU
&ODVVLILHU
7UDLQLQJVHW
&ODVVERXQGDU\=
&ODVVLILHU=
0RWKHUVHWRISDWWHUQVDPSOHV
E &ODVVERXQGDU\
&ODVVERXQGDU\
&ODVVLILHU
&ODVVLILHU
)HDWXUHV
)HDWXUHV
7UDLQLQJVHW
&ODVVERXQGDU\=
&ODVVLILHU= )HDWXUHV=
0RWKHUVHWRISDWWHUQVDPSOHV
F
FIGURE 4.8 Typical classifier design schemes of averaging-based decision fusion. VJG UKORNG CRRNKECVKQP QH VJG TGUWNVU QH TGITGUUKQP RTQDNGOU VQ ENCUUKſECVKQP RTQDNGOU KU KPUWHſEKGPV CPF VJCV KV KU PGEGUUCT[ VQ HWTVJGT CPCN[\G FGEKUKQP HWUKQP HQTOCNKUOU VJCV ECP DG OQTG FKTGEVN[ CRRNKGF VQ VJG OKPKOK\CVKQP QH ENCUUKſECVKQP GTTQTU #U QPG OKIJV KOCIKPG HTQO VJG FKUEWUUKQPU KP VJG CDQXG RCTCITCRJ VJG FGEKUKQP HWUKQP CRRTQCEJ VQ URGGEJ RCVVGTP TGEQIPKVKQP JCU DGGP VGUVGF KP UQOGYJCV JGWTKU VKE UV[NGU /QUV ECUGU QH VJG CRRTQCEJ JCXG GORKTKECNN[ GORNQ[GF VJG CXGTCIKPI QT YGKIJVGF CXGTCIKPI UEJGOG QH RTGFGEKUKQPU UKORN[ GZRGEVKPI VJCV VJG TGUWNVKPI FG EKUKQPU YQWNF DG OQTG TQDWUV VQ WPMPQYP RCVVGTP UCORNGU (KIWTG KNNWUVTCVGU VJTGG OCLQT V[RGU QH GODQFKOGPVU QH VJG CXGTCIKPIDCUGF FGEKUKQP HWUKQP HQT ENCU UKſECVKQP C FGUKIP QH OWNVKRNG UWDENCUUKſGTU WUKPI FKHHGTGPV UGVU QH VTCKPKPI FCVC
D FGUKIP QH OWNVKRNG UWDENCUUKſGTU WUKPI C UKPING VTCKPKPI FCVC UGV CPF E FGUKIP QH OWNVKRNG UWDENCUUKſGTU WUKPI FKHHGTGPV V[RGU QH HGCVWTG TGRTGUGPVCVKQP +P CNN VJTGG ECUGU C ſPCN FGEKUKQP KU OCFG VJTQWIJ CXGTCIKPI RTGFGEKUKQPU GCEJ OCFG YKVJ KVU
E\&5&3UHVV//&
EQTTGURQPFKPI UWDENCUUKſGT +P QVJGT YQTFU VJG ſPCN GUVKOCVGF ENCUU DQWPFCT[ KU FGVGTOKPGF D[ CXGTCIKPI VJG RTGDQWPFCTKGU GCEJ RTQFWEGF D[ KVU EQTTGURQPFKPI UWDENCUUKſGT 6JG RCVVGTP UCORNG UGVU HQT VTCKPKPI CTG VJG QPGU GZVTCEVGF HTQO VJG OQVJGT UCORNG UGV YJKEJ KU WUWCNN[ WPQDUGTXCDNG +P C CNN QH VJG VTCKPKPI UCORNG UGVU WUG C EQOOQP OGVJQF QH HGCVWTG TGRTGUGPVCVKQP VJCV KU CNN QH VJG RCVVGTPU QH VJG VTCKPKPI UCORNG UGVU CTG TGRTGUGPVGF KP CP KFGPVKECN HGCVWTG URCEG *GTG GCEJ VTCKPKPI UCORNG UGV KU GZVTCEVGF KPFGRGPFGPVN[ HTQO VJG OQVJGT FCVC UGV CPF C UWD ENCUUKſGT KU FGUKIPGF D[ WUKPI QPN[ KVU EQTTGURQPFKPI VTCKPKPI FCVC UGV +P EQPVTCUV KP D OWNVKRNG UWDENCUUKſGTU CTG FGUKIPGF QXGT C UKPING VTCKPKPI FCVC UGV QH UQOG ſPKVG UK\G CKOKPI VQ OCKPVCKP VJG KPFGRGPFGPE[ COQPI VJG FGUKIP RTQEGFWTGU QH VJG FKHHGTGPV UWDENCUUKſGTU +P E C FKHHGTGPV UWDENCUUKſGT KU FGUKIPGF D[ WUKPI C FKHHGTGPV HGCVWTG TGRTGUGPVCVKQP *GTG VJG FKHHGTGPV V[RGU QH HGCVWTGU CTG ECNEWNCVGF QXGT C UKPING VTCKPKPI UCORNG UGV QH UQOG ſPKVG UK\G $GECWUG VJG GHHGEV QH VJG CX GTCIKPI UEJGOG CUUWOGU VJCV VJG UWDENCUUKſGTU CTG FGUKIPGF KPFGRGPFGPVN[ FGUKIP RTQEGFWTGU YJKEJ CTG KNNWUVTCVGF D[ CTTQYU KP VJG ſIWTG UJQWNF DG DCUKECNN[ CU KPFG RGPFGPV CU RQUUKDNG +P VJKU NKIJV ECUG C UGGOU VQ DG VJG OQUV GHHGEVKXG HQT WVKNK\KPI VJG XCNWG QH VJG FGEKUKQP HWUKQP UEJGOG *QYGXGT RTGRCTKPI OWNVKRNG KPFGRGPFGPV VTCKPKPI UCORNG UGVU KU QHVGP EQUVN[ CPF OQUV ECUGU QH URGGEJ RCVVGTP TGEQIPKVKQP JCXG GORNQ[GF ECUGU D CPF E +P VJG TGOCKPKPI RCIGU QH VJKU UGEVKQP YG UJCNN KPVTQFWEG UGXGTCN GZGORNCT GODQFKOGPVU QH VJG CXGTCIKPIDCUGF FGEKUKQP HWUKQP CRRTQCEJ VQ URGGEJ RCVVGTP TGEQIPKVKQP
4.4.2 Examples of Embodiment 4.4.2.1 Multi-codebook Classifier Designed with GPD 1PG QH VJG OQUV UVTCKIJVHQTYCTF GODQFKOGPVU QH ECUG D KP (KIWTG YCU FGXGN QRGF KP =? +P VJKU TGRQTV VJG )&2 VTCKPKPI YCU CRRNKGF VQ RTQVQV[RG TGHGTGPEG XGEVQT DCUGF FKUVCPEG ENCUUKſGTU YJKEJ CTG DCUKECNN[ VJG UCOG CU CP .83 ENCUUKſGT +P CP GZRGTKOGPVCN VCUM QH ENCUUKH[KPI ENCUU #OGTKECP 'TJ[OG URGGEJ UCORNGU KV YCU UJQYP VJCV VJG WUG QH OWNVKRNG EQFGDQQMU GCEJ FGUKIPGF UGRCTCVGN[ YKVJ )2& EQWNF UWEEGUUHWNN[ KPETGCUG VJG ENCUUKſECVKQP CEEWTCE[ QH VJG DCUGNKPG UKPING )2& VTCKPGF EQFGDQQM ENCUUKſGT VJG CEEWTCE[ QH VJG DCUGNKPG U[UVGO KG YCU KP ETGCUGF VQ YJKEJ YCU COQPI VJG JKIJGUV UEQTGU QP VJG 'TJ[OG FCVC UGV *GTG GCEJ QH VJG OWNVKRNG EQFGDQQMU KU QH VJG UCOG UK\G CU VJG EQFGDQQM QH VJG DCUGNKPG UKPING EQFGDQQM U[UVGO +P =? CPQVJGT GHHGEV QH VJG FGEKUKQP HWUKQP YCU FGOQPUVTCVGF (KIWTG UJQYU ENCUUKſECVKQP CEEWTCE[ UEQTGU CU C HWPEVKQP QH VJG PWODGT QH RTQVQV[RGU RGT ENCUU CPF EQFGDQQM *GTG VJG VQVCN PWODGT QH RTQVQV[RGU RGT ENCUU YCU MGRV EQPUVCPV DWV VJG PWODGT CUUKIPGF RGT EQFGDQQM YCU EJCPIGF (QT GZCORNG YKVJ VJG VQVCN QH RTQVQV[RGU RGT ENCUU QPG EQWNF JCXG EQFGDQQMU YKVJ RTQVQV[RGU RGT ENCUU QT EQFGDQQMU YKVJ RGT ENCUU 6JG EWTXG KP VJG ſIWTG ENGCTN[ UJQYU VJCV VJGTG KU C DCNCPEG DGVYGGP VJG ſPGITCKPGF DQWPFCT[ GUVKOCVKQP QH OCP[ RTQVQV[RGU RGT EQFGDQQM CPF VJG EQCTUGITCKPGF CXGTCIKPI YKVJ OCP[ EQFGDQQMU QH HGYGT RTQVQ V[RGU UWIIGUVKPI VJG WVKNKV[ QH CXGTCIKPI UEJGOG WPFGT C RTCEVKECN EQPFKVKQP YJGTG
E\&5&3UHVV//&
4GEQIPKVKQPCEEWTCE[
2TQVQV[RGU%NCUU%QFGDQQM
FIGURE 4.9 Relation between recognition accuracy and the number of prototypes per class and codebook [3]. CXCKNCDNG U[UVGO TGUQWTEGU UWEJ CU RTQVQV[RGU CTG NKOKVGF 4.4.2.2 Multi-class Classification Based on Support Vector Machine )GPGTCNN[ ENCUUKſECVKQP KP C JKIJFKOGPUKQPCN XGEVQT URCEG KU GCUKGT VJCP KP C NQY FKOGPUKQPCN XGEVQT URCEG $[ HQEWUKPI QP VJKU IGPGTCN PCVWTG QH ENCUUKſECVKQP VJG UWRRQTV XGEVQT OCEJKPG 58/ JCU DGGP CVVTCEVKPI OWEJ TGUGCTEJ KPVGTGUV TGEGPVN[ #U UJQYP KP VJG HQNNQYKPI RCTCITCRJU 58/DCUGF OWNVKENCUU ENCUUKſECVKQP ECP DG EQPUKFGTGF C V[RG QH GODQFKOGPV QH ECUG D 6JG RWTRQUG QH 58/DCUGF ENCUUKſECVKQP KU VQ EQPXGTV C IKXGP VCUM QH ENCUUKH[KPI UCORNGU KP UQOG QTKIKPCNN[ NQYFKOGPUKQPCN XGEVQT URCEG VQ C VCUM QH ENCUUKH[KPI VJG UCORNGU KP C JKIJFKOGPUKQPCN XGEVQT URCEG VJTQWIJ C PQPNKPGCT UCORNG RTQLGEVKQP RGTHQTOGF D[ C PGWTCN PGVYQTMŏU MGTPGN HWPEVKQP 6JG RTQLGEVGF UCORNGU CTG VJGP GP EQFGF CU QPG QH VYQ ENCUUGU D[ WUKPI C NKPGCT FKUETKOKPCPV HWPEVKQP JCXKPI VJG NCTIGUV OCTIKP KP C VYQENCUU DQWPFCT[ TGIKQP 6JWU VJKU CRRTQCEJ KU PCVWTCNN[ GZRGEVGF VQ RQUUGUU VYQ V[RGU QH TQDWUV RTQRGTVKGU QPG DCUGF QP VJG NCTIGOCTIKP ENCUUKſECVKQP KP C JKIJFKOGPUKQPCN XGEVQT URCEG CPF VJG QVJGT DCUGF QP VJG HWUKQP QH VYQENCUU ENCUUKſECVKQP FGEKUKQPU HQT C OWNVKENCUU VCUM UGVVKPI +V YQWNF UGGO VJCV VJG ſTUV V[RG QH TQDWUVPGUU QTKIKPCVGU CU C OGCPU VQ EKTEWOXGPV QWVNKGT UCORNGU VJCV VGPF VQ DG OKUENCUUKſGF CPF VJG UGEQPF QTKIKPCVGU HTQO VJG KPETGCUGF UVCVKUVKECN UVCDKNKV[ KP UGVVKPI ENCUU DQWPFCTKGU #U KVU PCOG KORNKGU 58/ KU C OGVJQF HQT ENCUUKH[KPI ſZGFFKOGPUKQPCN UVCVKE XGEVQT RCVVGTPU 'ZVGPUKQP QH VJKU OGVJQFQNQI[ VQ VJG ENCUUKſECVKQP QH F[PCOKE RCVVGTPU KU
E\&5&3UHVV//&
UVKNN CP QPIQKPI TGUGCTEJ VQRKE CPF VJGTGHQTG CRRNKECVKQPU QH 58/ VQ URGGEJ RCVVGTP TGEQIPKVKQP CTG PQV [GV VJCV OCVWTG +P VJG HQNNQYKPI RCTCITCRJ YG KPVTQFWEG QPG TGEGPV CRRNKECVKQP GZCORNG QH 58/ ENCUUKſGTU KP =? 6JG VCUM VGUVGF KP =? YCU ENCUUKſECVKQP QH RJQPGOG UGIOGPVU VJCV YGTG TGRTGUGPVGF CU UVCVKE HGCVWTG XGEVQTU YJQUG EQORQPGPVU YGTG HQTOCPV CEQWUVKE TGUQPCPEG HTG SWGPE[ XCNWGU QT CXGTCIG EGRUVTCN EQGHſEKGPVU 58/ YCU CRRNKGF VQ VJGUG UVCVKE XGEVQT RCVVGTPU $GECWUG VJG 58/ HQTOCNKUO QTKIKPCNN[ CUUWOGU VJG PWODGT QH ENCUUGU VQ DG VYQ VJGTG CTG VYQ RQUUKDNG EQODKPCVQTKCN HQTOCVKQPU QH ENCUUKſECVKQP
VJG őQPG XU QPGŒ HQTOCVKQP CPF VJG őQPG XU CNNŒ HQTOCVKQP +P VJG QPG XU QPG HQTOCVKQP CP 58/DCUGF UWDENCUUKſGT YCU FGUKIPGF HQT GXGT[ RCKT QH VYQ FKHHGTGPV ENCUUGU #P WPMPQYP VGUV UCORNG YCU RTGENCUUKſGF D[ CNN QH VJG FGUKIPGF UWDENCUUKſGTU CPF YCU VJGP ſPCNN[ ENCUUKſGF YKVJ C XQVKPI UEJGOG QXGT VJG UWD ENCUUKſGTU *GTG VJG XQVKPI UEJGOG KU GSWKXCNGPV VQ CXGTCIKPI QXGT RTGFGEKUKQPU +P VJG QPG XU CNN HQTOCVKQP CP 58/DCUGF UWDENCUUKſGT YCU FGUKIPGF HQT GXGT[ RCKT QH C VCTIGV ENCUU CPF TGOCKPKPI ENCUUGU # VGUV UCORNG YCU RTGENCUUKſGF D[ CNN QH VJG UWDENCUUKſGTU CPF VJGP ſPCNN[ GPEQFGF VQ VJG ENCUU JCXKPI VJG NCTIGUV FKUVCPEG HTQO VJG UGRCTCVKPI J[RGTRNCPG +V UJQWNF DG PQVGF VJCV VJKU UGEQPF UEJGOG FQGU PQV KPENWFG VJG GHHGEV QH VJG CXGTCIKPI 6JGUG VYQ V[RGU QH 58/DCUGF ENCUUKſGTU YGTG EQORCTGF VQ C EQPXGPVKQPCN )CWUUKCP OKZVWTG ENCUUKſGT CPF VJG UWRGTKQTKV[ QH VJG QPG XU QPG HQTOCVKQP WUKPI 58/ YCU FGOQPUVTCVGF 6JG CDQXG QPG XU QPG HQTOCVKQP KPEQTRQTCVGF D[ VJG XQVKPI UEJGOG KU C V[RKECN HTCOGYQTM QH FGEKUKQP HWUKQP *QYGXGT KV KU PQV VTKXKCN VQ UJQY VJG UWRGTKQTKV[ QH VJG HTCOGYQTM VJGQTGVKECNN[ #P KORQTVCPV QPIQKPI TGUGCTEJ KUUWG KU JQY VQ CRRN[ VJG 58/DCUGF VYQENCUU ENCUUKſGT VQ OWNVKENCUU URGGEJ RCVVGTP ENCUUKſECVKQP VCUMU 4.4.2.3 Decision Fusion Using Different Classifiers +P VJG GODQFKOGPVU QH VJG RTGXKQWU UWDUGEVKQPU C UKPING V[RG QH ENCUUKſGT UVTWE VWTG YCU GORNQ[GF (QT GZCORNG CNN QH VJG UWDENCUUKſGTU KP 5GEVKQP CTG RTQVQV[RGDCUGF FKUVCPEG ENCUUKſGTU CPF CNN VJQUG KP 5GEVKQP CTG 58/DCUGF ENCUUKſGTU # FKHHGTGPV V[RG QH ENCUUKſGT EQWNF TGRTGUGPV C FKHHGTGPV V[RG QH ENCUU DQWPFCT[ CEEQTFKPIN[ ENCUUKſECVKQP FGEKUKQPU +P NKIJV QH VJKU CPQVJGT V[RG QH GO DQFKOGPV QH ECUG D KP (KIWTG KG QPG EQODKPKPI FKHHGTGPV V[RGU QH ENCUUKſGTU UWEJ CU *// CPF PGWTCN PGVYQTM JCU DGGP VGUVGF KP CP KORQTVCPV UWDCTGC QH URGGEJ RCVVGTP TGEQIPKVKQP KG URGCMGT TGEQIPKVKQP GI =? +P =? VJTGG V[RGU QH ENCUUKſGTU YGTG WUGF CP *//DCUGF U[UVGO C &2DCUGF FKUVCPEG ENCUUKſGT CPF C PGWTCN VTGG PGVYQTM U[UVGO YJKEJ KU C JKGTCTEJKECN ENCU UKſGT VJCV WUGU C VTGG CTEJKVGEVWTG VQ KORNGOGPV C UGSWGPVKCN NKPGCT FGEKUKQP UVTCV GI[ 6JTQWIJ EQORTGJGPUKXG GZRGTKOGPVCN GXCNWCVKQPU C DCUKE VGPFGPE[ YCU FGOQP UVTCVGF VJG NCTIGT VJG PWODGT QH UWDENCUUKſGTU KU VJG OQTG CEEWTCVG VJG ſPCN HWUGF ENCUUKſECVKQP FGEKUKQPU CTG #P QXGTXKGY QH VJG FGEKUKQP HWUKQP CRRTQCEJ VQ URGCMGT TGEQIPKVKQP ECP DG HQWPF KP VJG NKVGTCVWTG GI UGG =?
E\&5&3UHVV//&
'3PDWFKLQJ WLPHZDUSLQJ &RPELQHG YLVHPHSKRQHPH OD\HU 3KRQHPH OD\HU
9LVHPH OD\HU
+LGGHQ OD\HU
+LGGHQ OD\HU
$FRXVWLF LQSXW
'3PDWFKLQJ WLPHZDUSLQJ &RPELQHG YLVHPHSKRQHPH OD\HU +LGGHQ OD\HU
9LVXDO LQSXW
D
$FRXVWLF LQSXW
'3PDWFKLQJ WLPHZDUSLQJ
F
&RPELQHG YLVHPHSKRQHPH OD\HU +LGGHQ OD\HU
+LGGHQ OD\HU $FRXVWLF LQSXW
9LVXDO LQSXW
9LVXDO LQSXW
E
FIGURE 4.10 Typical block diagrams of the MSTDNN-based audio-visual speech recognition [7].
4.4.2.4 Decision Fusion Using Multi-modal Classifiers 6JG EQPVTKDWVKQP OGEJCPKUO QH VJG FGEKUKQP HWUKQP ECUG E QH (KIWTG KU DCUK ECNN[ FKHHGTGPV HTQO VJCV QH VJG RTGXKQWU VYQ ECUGU KG C CPF D +P ECUG E VJG FGEKUKQP EQODKPCVKQP KU RGTHQTOGF KP FKHHGTGPV HGCVWTG URCEGU CPF VJWU KVU TGUWNVKPI GHHGEV KU PQV VJG UCOG CU VJG CXGTCIG EQORWVCVKQP KP C UKPING HGCVWTG URCEG +P VJKU OWNVKHGCVWTG ECUG VJG UWDENCUUKſGTU GCEJ FGUKIPGF QXGT C FKHHGTGPV HGCVWTG UGV CTG GZRGEVGF VQ OWVWCNN[ EQORGPUCVG VJG YGCMPGUU QH VJGKT EQORGVKPI UWDENCUUKſGTU YJKNG VJG UWDENCUUKſGTU KP ECUGU C CPF D CTG GZRGEVGF VQ RCTVKEKRCVG VJG CXGTCIKPI QRGTCVKQP KP QTFGT VQ TGFWEG VJG UVCVKUVKECN XCTKCPEG QH VJG KPFKXKFWCN RTGFGEKUKQPU #OQPI VJG OCP[ RQUUKDKNKVKGU QH VJG FGEKUKQP HWUKQP WUKPI FKHHGTGPV V[RGU QH HGCVWTGU VJG WUG QH C XKUWCN HCEG KOCIG CU YGNN CU KVU EQTTGURQPFKPI CEQWUVKE URGGEJ UKIPCN JCU DGGP TCRKFN[ ITQYKPI CU CP GOGTIKPI TGUGCTEJ VQRKE 2GTHGEV TGEQIPKVKQP QH URGGEJ UKIPCNU KU GUUGPVKCNN[ FKHſEWNV FWG VQ UGXGTCN TGCUQPU UWEJ CU VJG KPEQORNGVGPGUU QH CTVKEWNCVKQP CPF CEQWUVKE FKUVQTVKQP QXGT URGGEJ VTCPUOKUUKQP EJCPPGNU 'XGP HQT JW OCPU JGCTKPI QXGT C VGNGRJQPG KU WUWCNN[ OQTG FKHſEWNV VJCP JGCTKPI KP HCEGVQHCEG EQOOWPKECVKQP RTQDCDN[ FWG VQ VJG NCEM QH HCEG KPHQTOCVKQP +P VJG HQNNQYKPI YG KPVTQFWEG VYQ TGEGPV GZCORNGU QH CWFKQXKUWCN URGGEJ TGEQIPKVKQP VJCV WUGU NKR UJCRG KPHQTOCVKQP KP CFFKVKQP VQ VJG UVCPFCTF QDUGTXCVKQP QH C URGGEJ UKIPCN = ?
E\&5&3UHVV//&
$XGLR+00 $XGLROLNHOLKRRG
$XGLR IHDWXUHV
6WUHDP ZHLJKWV
&ODVV LQGH[
9LVXDO+00 9LVXDO IHDWXUHV
9LVXDOOLNHOLKRRG
FIGURE 4.11 Block diagram of the twofold-HMM-based audio-visual speech recognition [21]. +P =? /56&00 ENCUUKſGTU YGTG WUGF VQ ENCUUKH[ CP KPRWV UVTGCO VJCV EQPUKUVGF QH CP CEQWUVKE URGGEJ UKIPCN CPF KVU EQTTGURQPFKPI XKUWCN NKR UJCRG KOCIG UKIPCN $CUGF QP VJG UVTWEVWTCN ƀGZKDKNKV[ QH /56&00 VJGTG CTG UGXGTCN RQUUKDKNKVKGU QH UKIPCN EQODK PCVKQP 6JTGG V[RGU QH EQODKPCVKQP KNNWUVTCVGF KP (KIWTG YGTG CEVWCNN[ GZCO KPGF C FCVC HWUKQP QH EQODKPKPI CWFKQXKUWCN FCVC QP VJG XKUGOGRJQPGOG NC[GT
D FCVC HWUKQP QH EQODKPKPI CWFKQXKUWCN FCVC QP JKFFGP NC[GT CPF E FCVC HWUKQP QH EQODKPKPI CWFKQXKUWCN FCVC QP VJG KPRWV NC[GT #OQPI VJGUG KP RTKPEKRNG VJG UVTWE VWTG QH C KU VJG DGUV UWKVGF HQT KPFGRGPFGPV FGUKIP QH VJG VYQ UWDENCUUKſGTU GCEJ HQT C FKHHGTGPV V[RG QH OQFCN FCVC +PFGRGPFGPE[ JGTG CNNQYU QPG VQ ECTGHWNN[ VTCKP GCEJ UWDENCUUKſGT TGƀGEVKPI VJG PCVWTG QH GCEJ OQFCN FCVC KP KVU EQTTGURQPFKPI FGUKIP QH VJG UWDENCUUKſGT %QORTGJGPUKXG GZRGTKOGPVCN GXCNWCVKQPU ENGCTN[ FGOQPUVTCVGF VJG GHHGEV QH VJG FCVC HWUKQP QXGT VJG OWNVKOQFCN FCVC UVTGCO URGEKCNN[ VJG CRRTQCEJ QH WUKPI VJG V[RG C PGVYQTM +P =? KP RNCEG QH VJG UQECNNGF PGWTCN PGVYQTMU *//ŏU YGTG WUGF CU UWDENCUUKſGTU QPG HQT VJG CEQWUVKE URGGEJ UVTGCO CPF QPG HQT VJG XKUWCN NKR UJCRG UVTGCO +V UJQWNF DG TGECNNGF VJCV *// ECP DG FGſPGF CU C V[RG QH 00 U[UVGO 6JG DNQEM FKCITCO QH VJG ENCUUKſGT WUGF KU KNNWUVTCVGF KP (KIWTG &GUKIP MG[U KP VJKU FCVC HWUKQP UEJGOG CTG VJG UGNGEVKQPU QH FGUKIP OGVJQFU HQT VJG CWFKQ *// VJG XKUWCN *// CPF VJG UVTGCO YGKIJV VJCV KU WUGF HQT EQODKPKPI VJG QWVRWVU QH VJG VYQ FKHHGTGPV OQFCN *// UWDENCUUKſGTU # EQPXGPVKQPCN UGNGEVKQP QH VJG FGUKIP OGVJQFU HQT VJG CWFKQXKUWCN *// UWDENCUUKſGTU KU VJG /.' OGVJQF CPF CP KOOGFKCVG TGCUQPCDNG UGNGEVKQP HQT VJG UVTGCO YGKIJV KU VJG )2& OGVJQF WUKPI VJG ENCUUKſECVKQP GTTQT NQUU 6JTQWIJ GZRGTKOGPVCN GXCNWCVKQPU KP C URGCMGTKPFGRGPFGPV KUQNCVGF YQTF TGEQIPK VKQP VCUM VJG CWVJQTU FGOQPUVTCVGF VJCV WUKPI )2& VTCKPKPI HQT CNN VJTGG VTCKPCDNG OQFWNGU KG VJG CWFKQ *// UWDENCUUKſGT VJG XKUWCN *// UWDENCUUKſGT CPF VJG UVTGCO YGKIJV CEJKGXGF C UKIPKſECPV GTTQT TGFWEVKQP WR VQ QXGT VJG ENCUUKſGT
E\&5&3UHVV//&
KP YJKEJ VYQ UWDENCUUKſGTU YGTG VTCKPGF YKVJ VJG /.' OGVJQF
4.5 Concluding Remarks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ſECVKQP OGCUWTG VQ KPVTQFWEG C NQUU HWPEVKQP VJCV GPCDNGU QPG VQ GXCNWCVG VJG TGEQIPKVKQP TGUWNV QH C IKXGP VTCKPKPI UCORNG CPF VQ QRVKOK\G VJG VTCKPCDNG U[UVGO RCTCOGVGTU QH VJG TGEQIPK\GT CV JCPF D[ WUKPI VJG RTGUGV NQUU HWPEVKQP +P VJG VTCKPKPI RTQEGFWTG VJG UGNGEVKQP QH VJG NQUU KU ETWEKCN 8CTKQWU UGNGE VKQPU JCXG DGGP GZCOKPGF UQ HCT KPENWFKPI VJG ENCUUKſECVKQP GTTQT NQUU VJG USWCTGF GTTQT NQUU CPF VJG ETQUU GPVTQR[ NQUU (TQO VJG XKGYRQKPV QH VJG )2& HQTOCNKUO YG UJQYGF VJG FKTGEVPGUU QH VJG ENCUUKſECVKQP GTTQT NQUU VQ VJG QRVKOCN OKPKOWO GTTQT TCVG ENCUUKſECVKQP QT KP QVJGT YQTFU VJG OKPKOWO ENCUUKſECVKQP GTTQT RTQDCDKNKV[ EQPFKVKQP +P VJG EJCRVGT YG CNUQ UWOOCTK\GF VJG HTCOGYQTM QH VJG HWUKQP QH OWNVKRNG ENCUUK ſECVKQP FGEKUKQPU #U CP CRRTQCEJ VQ KPETGCUKPI VJG FGUKIP TQDWUVPGUU VQ WPMPQYP UCORNGU FGEKUKQP HWUKQP GURGEKCNN[ VJG CXGTCIKPI UEJGOG QXGT UWDENCUUKſGT FGEK UKQPU JCU CVVTCEVGF VJG TGEGPV TGUGCTEJGTUŏ KPVGTGUVU 6JGTG CTG UGXGTCN RQUUKDKNKVKGU HQT VJG GODQFKOGPVU QH VJG HTCOGYQTM 9G RCTVKEWNCTN[ HQEWUGF QP VYQ ECUGU FGUKIPKPI OWNVKRNG UWDENCUUKſGTU QXGT C UKPING VTCKPKPI UGV YKVJ C UKPING OGVJQF QH HGCVWTG TGRTGUGPVCVKQP CPF FGUKIPKPI OWNVKRNG UWDENCUUKſGTU QXGT C UKPING VTCKP KPI UGV YKVJ OWNVKRNG OGVJQFU QH HGCVWTG TGRTGUGPVCVKQP 6JG WVKNKV[ QH VJG FGEKUKQP HWUKQP CRRTQCEJ JCU DGGP ENGCTN[ FGOQPUVTCVGF KP OCP[ GZRGTKOGPVCN UVWFKGU DWV KVU VJGQTGVKECN CPCN[UKU KU UVKNN KPUWHſEKGPV FWG VQ VJG NCEM QH C OCVJGOCVKECN HTCOGYQTM HQT CPCN[\KPI VJG UVCVKUVKECN EJCTCEVGTKUVKEU QH ENCUU DQWPFCT[ GUVKOCVGU (QNNQYKPI VJG EQPXGPVKQPCN VCZQPQO[ YG JCXG VQ FKUVKPIWKUJ DGVYGGP *// CPF 00 *QYGXGT YG DCUKECNN[ VJKPM VJCV VJGTG CTG PQ UKIPKſECPV FKHHGTGPEGU KP U[UVGO UVTWEVWTG DGVYGGP VJGUG VYQ V[RGU QH U[UVGOU #EVWCNN[ DQVJ CTG GODQFKOGPVU QH C YKFGT EQPEGRV QH C ITCRJKECN OQFGN +P VJKU NKIJV YG CNUQ VJKPM VJCV KP ENCUUKſGT FGUKIP QPG UJQWNF PQV CFJGTG VQ GORKTKECN UGNGEVKQPU QH U[UVGO UVTWEVWTG UWEJ CU C J[DTKF U[UVGO VJCV UKORN[ EQODKPGU GZKUVKPI U[UVGOU 6JKU KU YJ[ YG JCXG QPN[ DTKGƀ[ FKUEWUUGF VJG KUUWG QH 00ŏU CTEJKVGEVWTG UGNGEVKQP 0GGFNGUU VQ UC[ C RTQRGT FGUKIP QH 00 UVTWEVWTG KU KORQTVCPV +V UJQWNF RTQDCDN[ TGƀGEV VJG PCVWTG QH RCVVGTPU VJCV PGGF VQ DG ENCUUKſGF +P CFFKVKQP VTCKPKPI OGVJQFU UWEJ CU FKUETKOKPCVKXG VTCKP KPI UJQWNF DG HWTVJGT GXQNXGF UQ CU VQ EQXGT VJG FGVGTOKPCVKQP QH ENCUUKſGT UVTWEVWTG
E\&5&3UHVV//&
CU YGNN CU VJG CFLWUVOGPV QH RTGUGV U[UVGO RCTCOGVGTU +P 5GEVKQP YG KPVTQFWEGF VJG INQDCN UEQRG FGUKIP OGVJQF VJCV WUGU VJG EJCKP TWNG QH FKHHGTGPVKCN ECNEWNWU 1PG OC[ PQVKEG VJCV VJG OGEJCPKUO QH VJG INQDCN VTCKP KPI KU GSWKXCNGPV KP QRGTCVKQP VQ VJG GTTQT DCEMRTQRCICVKQP CNIQTKVJO FGXGNQRGF HQT /.2 PGVYQTMU VJQWIJ VJG OGEJCPKUO UJQYP KU OQTG IGPGTCN $[ WUKPI VJG INQDCN FGUKIP UVTCVGI[ QPG ECP KP RTKPEKRNG FGUKIP HGCVWTG GZVTCEVQTU VJCV CTG DGVVGT UWKVGF VQ ENCUUKſECVKQP VJCP VJQUG FGVGTOKPGF KP GORKTKECN YC[U # HWVWTG TGUGCTEJ GHHQTV YKNN DG VQ FKUEQXGT C HGCVWTG GZVTCEVQT VJCV OQFGNU VJG UCNKGPV PCVWTG QH URGGEJ UKIPCNU HQT ENCUUKſECVKQP #U UWOOCTK\GF CDQXG VJG UGNGEVKQP QH VJG NQUU HWPEVKQP KU KORQTVCPV CPF CNUQ VJG ENCUUKſECVKQP GTTQT EQWPV NQUU RQUUGUUGU VJG URGEKCN HGCVWTG QH FKTGEVPGUU VQ VJG OKP KOWO GTTQTTCVG ENCUUKſECVKQP (WTVJGT CPCN[UGU QP VJKU NQUU UGNGEVKQP CTG ENGCTN[ FGUKTGF HQT CFXCPEKPI RCVVGTP TGEQIPKVKQP OGVJQFQNQI[ CU YGNN CU 00DCUGF URGGEJ TGEQIPKVKQP VGEJPQNQI[
References =? 5 #OCTK ő# VJGQT[ QH CFCRVKXG RCVVGTP ENCUUKſGTUŒ +''' 6TCPU '% XQN '% RR =? 2 %NCTMUQP CPF 2 /QTGPQ ő1P VJG WUG QH UWRRQTV XGEVQT OCEJKPGU HQT RJQ PGVKE ENCUUKſECVKQPŒ 2TQE +%#552 RCRGT PQ =? # &WEJQP CPF 5 -CVCIKTK ő+PETGCUKPI VJG 4QDWUVPGUU QH )2&$CUGF #NIQ TKVJOUŒ 2TQE #5, 5RTKPI %QPHGTGPEG RR =? 4 &WFC CPF 2 *CTV ő2CVVGTP %NCUUKſECVKQP CPF 5EGPG #PCN[UKUŒ 0GY ;QTM 9KNG[ =? - 4 (CTTGNN ő/QFGN %QODKPCVKQP CPF 9GKIJV 5GNGEVKQP %TKVGTKC HQT 5RGCMGT 8GTKſECVKQPŒ 0GWTCN 0GVYQTMU HQT 5KIPCN 2TQEGUUKPI +: +''' RR
=? - (CTTGNN ő0GVYQTMU HQT 5RGCMGT 4GEQIPKVKQPŒ KP ő*CPFDQQM QH 0GWTCN 0GV YQTMU HQT 5RGGEJ 2TQEGUUKPI GF 5 -CVCIKTKŒ #TVGEJ *QWUG =? , (TKVUEJ * *KNF 7 /GKGT CPF # 9CKDGN ő6KOG&GNC[ 0GWTCN 0GV YQTMU CPF 00*// *[DTKFU # (COKN[ QH %QPPGEVKQPKUV %QPVKPWQWU5RGGEJ 4GEQIPKVKQP 5[UVGOUŒ KP ő*CPFDQQM QH 0GWTCN 0GVYQTMU HQT 5RGGEJ 2TQEGUU KPI 'F 5 -CVCIKTKŒ RR $QUVQP #TVGEJ *QWUG =? 2 *CHHPGT / (TCP\KPK CPF # 9CKDGN ő+PVGITCVKPI VKOG CNKIPOGPV CPF PGWTCN PGVYQTMU HQT JKIJ RGTHQTOCPEG EQPVKPWQWU URGGEJ TGEQIPKVKQPŒ 2TQE +%#552 RR
E\&5&3UHVV//&
=? * +YCOKFC 5 -CVCIKTK ' /E>OQVV CPF ; 6QJMWTC ő# *[DTKF 5RGGEJ 4GEQIPKVKQP 5[UVGO 7UKPI *//U YKVJ CP .836TCKPGF %QFGDQQMŒ , #EQWUV 5QE ,RP ' 8QN 0Q RR =? $* ,WCPI CPF 5 -CVCIKTK ő&KUETKOKPCVKXG NGCTPKPI HQT OKPKOWO GTTQT ENCU UKſECVKQPŒ +''' 6TCPU 52 XQN PQ RR =? $* ,WCPI 9 %JQW CPF %* .GG ő/KPKOWO ENCUUKſECVKQP GTTQT TCVG OGVJQFU HQT URGGEJ TGEQIPKVKQPŒ +''' 6TCPU 5#2 XQN RR =? 5 -CVCIKTK %* .GG CPF $* ,WCPI ő0GY FKUETKOKPCVKXG VTCKPKPI CNIQ TKVJOU DCUGF QP VJG IGPGTCNK\GF RTQDCDKNKUVKE FGUEGPV OGVJQFŒ +''' 0GWTCN 0GVYQTMU HQT 5KIPCN 2TQEGUUKPI RR =? 5 -CVCIKTK CPF %* .GG ő# PGY J[DTKF CNIQTKVJO HQT URGGEJ TGEQIPKVKQP DCUGF QP *// UGIOGPVCVKQP CPF NGCTPKPI XGEVQT SWCPVK\CVKQPŒ +''' 6TCPU 5#2 XQN RR =? 5 -CVCIKTK $* ,WCPI CPF # $KGO ő&KUETKOKPCVKXG HGCVWTG GZVTCEVKQPŒ KP #TVKſEKCN 0GWTCN 0GVYQTMU HQT 5RGGEJ CPF 8KUKQP 4 /COOQP 'F .QPFQP 7- %JCROCP CPF *CNN RR =? 5 -CVCIKTK $* ,WCPI CPF %* .GG ő2CVVGTP TGEQIPKVKQP WUKPI C HCOKN[ QH FGUKIP CNIQTKVJOU DCUGF WRQP VJG IGPGTCNK\GF RTQDCDKNKUVKE FGUEGPV OGVJQFŒ 2TQE +''' XQN PQ RR =? 5 -CVCIKTK 'Fő*CPFDQQM QH 0GWTCN 0GVYQTMU HQT 5RGGEJ 2TQEGUUKPIŒ $QUVQP #TVGEJ *QWUG =? 5 -CVCIKTK ő/KPKOWO %NCUUKſECVKQP 'TTQT 0GVYQTMUŒ KP ő*CPFDQQM QH 0GW TCN 0GVYQTMU HQT 5RGGEJ 2TQEGUUKPI 'F 5 -CVCIKTKŒ RR $QUVQP #TVGEJ *QWUG =? 6 -QJQPGP ) $CTPC CPF 4 %JTKUNG[ ő5VCVKUVKECN RCVVGTP TGEQIPKVKQP YKVJ PGWTCN PGVYQTMU DGPEJOCTMKPI UVWFKGUŒ 2TQE QH +%00 XQN RR + +
=? 6 -QJQPGP ő5GNH1TICPK\KPI (GCVWTG /CRUŒ 0GY ;QTM 5RTKPIGT8GTNCI
=? ' /E>OQVV CPF 5 -CVCIKTK ő.83DCUGF UJKHVVQNGTCPV RJQPGOG TGEQIPK VKQPŒ +''' 6TCPU 52 XQN RR =? % /K[CLKOC - 6QMWFC CPF 6 -KVCOWTC ő#WFKQXKUWCN 5RGGEJ 4GEQIPK VKQP 7UKPI /KPKOWO %NCUUKſECVKQP 'TTQT 6TCKPKPIŒ KP ő0GWTCN 0GVYQTMU HQT 5KIPCN 2TQEGUUKPI :Œ +''' =? . 4CDKPGT CPF $* ,WCPI ő(WPFCOGPVCNU QH 5RGGEJ 4GEQIPKVKQPŒ 'PING YQQF %NKHHU 2TGPVKEG*CNN
E\&5&3UHVV//&
=? # , 4QDKPUQP ő#P CRRNKECVKQP QH TGEWTTGPV PGVU VQ RJQPG RTQDCDKNKV[ GUVKOC VKQPŒ +''' 6TCPU 00 XQN PQ RR =? & ' 4WOGNJCTV ) ' *KPVQP CPF 4 , 9KNNKCOU ő.GCTPKPI +PVGTPCN 4GRTG UGPVCVKQPU D[ 'TTQT 2TQRCICVKQPŒ KP ő2CTCNNGN &KUVTKDWVGF 2TQEGUUKPI 'ZRNQ TCVKQPU KP VJG /KETQUVTWEVWTG QH %QIPKVKQP & ' 4WOGNJCTV GV CN 'FUŒ /+6 2TGUU =? / 5EJWUVGT CPF - - 2CNKYCN ő$KFKTGEVKQPCN TGEWTTGPV PGWTCN PGVYQTMUŒ +''' 6TCPU 52 XQN PQ RR =? 8 6TGUR ő%QOOKVVGG /CEJKPGUŒ KP ő*CPFDQQM QH 0GWTCN 0GVYQTM 5KIPCN 2TQEGUUKPI GFU ;* *W CPF ,0 *YCPIŒ $QEC 4CVQP %4% 2TGUU =? # 9CKDGN 6 *CPC\CYC ) *KPVQP - 5JKMCPQ CPF - .CPI ő2JQPGOG TGEQIPKVKQP WUKPI VKOGFGNC[ PGWTCN PGVYQTMUŒ +''' 6TCPU #552 XQN RR
4.6 Appendix: Maximizing Mutual Information 6JGTG KU CPQVJGT RQUUKDNG FGſPKVKQP QH VJG NQUU VJCV KU TGNCVGF VQ VJG ETQUU GPVTQR[ NQUU 6JKU JCU DGGP WUGF CU C NQUU HQT VJG FKUETKOKPCVKXG VTCKPKPI QH *// URGGEJ ENCUUKſGTU DWV VQ VJG DGUV QH VJG CWVJQTUŏ MPQYNGFIG KV JCU PQV DGGP WUGF HQT VJG FGUKIP QH 00DCUGF URGGEJ ENCUUKſGTU (QT TGHGTGPEG RWTRQUGU YG KPVTQFWEG VJKU CNVGTPCVKXG EJQKEG 6JG NQUU KU TGHGTTGF VQ CU OWVWCN KPHQTOCVKQP NQUU CPF KV KU FGſPGF CU
YJGTG KV KU CUUWOGF VJCV C VTCKPKPI UCORNG DGNQPIU VQ 6TCKPKPI WUKPI VJKU OGCUWTG CKOU CV EQTTGEV ENCUUKſECVKQP D[ OCZKOK\KPI VJG OWVWCN KPHQTOCVKQP QXGT VJG RQUUKDNG ENCUUGU QT KP QVJGT YQTFU KPETGCUKPI VJG UGRCTCDKNKV[ QH VJG ENCUUGU %NGCTN[ VJKU VTCKPKPI HQNNQYU VJG FKUETKOKPCVKXG VTCKPKPI EQPEGRV (QT FKUEWUUKQP RWTRQUGU YG EQPUKFGT VJG PGICVKXG OWVWCN KPHQTOCVKQP CPF TGCEJ VJG HQNNQYKPI KPGSWCNKV[ VJTQWIJ UKORNG TGYTKVKPI QRGTCVKQPU
E\&5&3UHVV//&
*GTG CUUWOKPI VJG NQICTKVJOKE NKMGNKJQQF VQ DG VJG FKUETKOKPCPV HWPEVKQP QPG ECP VTGCV VJG DQVVQO NKPG GZRTGUUKQP QH CU C MKPF QH OKUENCU UKſECVKQP OGCUWTG
6JGP VJG KPGSWCNKV[
JQNFU VTWG %NGCTN[ OCZKOK\KPI VJG OWVWCN KPHQTOCVKQP NGCFU CV NGCUV VQ OKPKOK\KPI VJG OKUENCUUKſECVKQP OGCUWTG %QPUGSWGPVN[ VTCKPKPI DCUGF QP VJG OCZKOK\C VKQP QH VJG OWVWCN KPHQTOCVKQP KU EQPUKFGTGF FKUETKOKPCVKXG VTCKPKPI VJCV WUGU VJG NKPGCT NQUU CPF VJG OKUENCUUKſECVKQP OGCUWTG .KMG VJG ECUG QH VJG USWCTGF GTTQT NQUU VJKU VTCKPKPI KU EGTVCKPN[ C V[RG QH FKUETKOK PCVKXG VTCKPKPI *QYGXGT FWG VQ VJG FKUETGRCPE[ DGVYGGP VJG UOQQVJGF GTTQT EQWPV NQUU CPF VJG NKPGCT NQUU WUGF JGTG VJKU VTCKPKPI ECPPQV IWCTCPVGG VJCV VJG OKPKOWO ENCUUKſECVKQP GTTQT EQPFKVKQP YKNN DG CEJKGXGF
E\&5&3UHVV//&
5 Large Vocabulary Speech Recognition Based on Statistical Methods Jean-Luc Gauvain and Lori Lamel LIMSI, France
CONTENTS
+PVTQFWEVKQP 1XGTXKGY .CPIWCIG /QFGNKPI 2TQPWPEKCVKQP /QFGNKPI #EQWUVKE /QFGNKPI &GEQFKPI +PFKECVKXG 2GTHQTOCPEG .GXGNU 2QTVCDKNKV[ CPF .CPIWCIG &GRGPFGPEKGU 4GHGTGPEGU
5.1 Introduction 5RGGEJ TGEQIPKVKQP KU EQPEGTPGF YKVJ EQPXGTVKPI VJG URGGEJ YCXGHQTO CP CEQWUVKE UKIPCN KPVQ C UGSWGPEG QH YQTFU 6QFC[ŏ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
E\&5&3UHVV//&
5.2 Overview (TQO C UVCVKUVKECN RQKPV QH XKGY URGGEJ KU CUUWOGF VQ DG IGPGTCVGF D[ C NCPIWCIG OQFGN YJKEJ RTQXKFGU GUVKOCVGU QH HQT CNN RQUUKDNG YQTF UVTKPIU ½ ¾ CPF CP CEQWUVKE OQFGN TGRTGUGPVGF D[ C RTQDCDKNKV[ FGPUKV[ HWPEVKQP GP EQFKPI VJG OGUUCIG KP VJG UKIPCN 6JG IQCN QH URGGEJ TGEQIPKVKQP KU IGPGTCNN[ FGſPGF CU ſPFKPI VJG OQUV NKMGN[ YQTF UGSWGPEG IKXGP VJG QDUGTXGF CEQWUVKE UKIPCN IKXGP VJG URGGEJ UKIPCN QT GSWKXCNGPVN[ KG QH OCZKOK\KPI VJG RTQDCDKNKV[ QH OCZKOK\KPI VJG RTQFWEV .8%54 U[UVGOU WUG CEQWUVKE WPKVU EQTTGURQPFKPI VQ RJQPGU QT RJQPGUKPEQPVGZV YJGTG GCEJ YQTF KU FGUETKDGF D[ QPG QT OQTG RJQPG VTCPUETKRVKQPU #UUWOKPI VJCV VJG URGGEJ UKIPCN FGRGPFU QPN[ QP VJG WPFGTN[KPI RJQPG UGSWGPEG YJGTG VJG ½ ¾ VJGP ECP DG TGYTKVVGP CU UWOOCVKQP KU VCMGP QXGT VJG UGV RTQPWPEKCVKQPU EQTTGURQPFKPI VQ VJG YQTF UGSWGPEG +P RTCEVKEG VJKU UGV KU TGCUQPCDN[ UOCNN CU VJG CXGTCIG PWODGT QH RTQPWPEKC VKQP XCTKCPVU RGT YQTF KU NGUU VJCP VYQ 6JG WPFGTN[KPI URGGEJ IGPGTCVKQP OQFGN KU KNNWUVTCVGF KP (KIWTG 6JG YQTF UGSWGPEG RTQFWEGF D[ VJG NCPIWCIG OQFGN KU UWE EGUUKXGN[ VTCPUHQTOGF D[ VYQ VTCPUFWEGTU VJG RTQPWPEKCVKQP OQFGN CPF VJG CEQWUVKE OQFGN VQ [KGNF VJG URGGEJ UKIPCN 6JKU HQTOWNCVKQP QH VJG .8%54 RTQDNGO NGCFU VQ VJG HQNNQYKPI HQWT OCKP EQPUKFGT CVKQPU
6JG NCPIWCIG OQFGNKPI RTQDNGO KG EQORWVKPI VJG C RTKQTK RTQDCDKNKV[ +V KU WUWCNN[ GUVKOCVGF HTQO TGNCVKXG nITCO HTGSWGPEKGU KP VTCPUETKRVKQPU QH URGGEJ FCVC CU YGNN CU TGNCVGF VGZV EQTRQTC
6JG RTQPWPEKCVKQP OQFGNKPI RTQDNGO KG VJG EQORWVCVKQP QH 6JKU TGNKGU QP C RTQPWPEKCVKQP FKEVKQPCT[ YJKEJ OC[ KPENWFG GUVKOCVGU QH VJG YQTF RTQPWPEKCVKQP RTQDCDKNKVKGU 6JG CEQWUVKE OQFGNKPI RTQDNGO KG FGVGTOKPKPI VJG UVTWEVWTG QH VJG RTQDC DKNKV[ FGPUKV[ HWPEVKQP CPF GUVKOCVKPI KVU UVCVKUVKECN RCTCOGVGTU HTQO URGGEJ UCORNGU 6JG OQUV RTGFQOKPCPV CRRTQCEJ WUGU EQPVKPWQWU FGPUKV[ JKF FGP /CTMQX OQFGNU *// VQ TGRTGUGPV EQPVGZVFGRGPFGPV RJQPGU
6JG UGCTEJ RTQDNGO KG FGVGTOKPKPI VJG DGUV YQTF J[RQVJGUKU HQT VJG URGGEJ FCVC IKXGP VJG OQFGNU 6JKU KU C DKI EJCNNGPIG HQT .8%54 FWG VQ VJG NCTIG XQECDWNCT[ CPF NCPIWCIG OQFGN UK\G
+P VJKU EJCRVGT VJG VGTO RJQPG KU WUGF VQ TGHGT VQ CEQWUVKE WPKVU YKVJQWV CVVGORVKPI VQ NCDGN VJGO CU RJQPGOKE TGHGTTKPI VQ VJG GNGOGPVCT[ CPF FKUVKPEVKXG UQWPFU KP VJG NCPIWCIG QT RJQPGVKE VJG QDUGTXGF TGCNK\CVKQP QH VJG GNGOGPVCT[ UQWPFU %QPVGZVWCN RJQPG WPKVU RJQPGUKPEQPVGZV KORNKEKVN[ OQFGN YJCV ECP DG EQPUKFGTGF CNNQRJQPGU KG EQPVGZVWCN RJQPGVKE XCTKCPVU QH VJG WPFGTN[KPI RJQPGOG
E\&5&3UHVV//&
2 9
H :^*
2 *^9 9
.CPIWCIG /QFGN
2TQPWPEKCVKQP /QFGN
YQTFUGSWGPEG
*
#EQWUVKE /QFGN
RJQPGUGSWGPEG
:
URGGEJUKIPCN
FIGURE 5.1 LVCSR speech generation model: The word sequence produced by the language model is successively transformed by the pronunciation model ( ) and the acoustic model ( ), resulting in the speech signal .
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
5.3 Language Modeling .CPIWCIG OQFGNU ./U ECRVWTG TGIWNCTKVKGU KP URQMGP NCPIWCIG CPF CTG WUGF KP URGGEJ TGEQIPKVKQP VQ GUVKOCVG VJG RTQDCDKNKV[ QH YQTF UGSWGPEGU 9JKNG ITCOOCV KECN EQPUVTCKPVU FGUETKDGF D[ JCPFETCHVGF EQPVGZVHTGG ITCOOCTU JCXG DGGP WUGF HQT UOCNN VQ OGFKWO UK\G XQECDWNCT[ VCUMU .8%54 KU GUUGPVKCNN[ CNYC[U DCUGF QP FCVC FTKXGP CRRTQCEJGU 6JG OQUV RQRWNCT UVCVKUVKECN OGVJQF KU VJG UQ ECNNGF nITCO OQFGN YJKEJ CVVGORVU VQ ECRVWTG VJG U[PVCEVKE CPF UGOCPVKE EQPUVTCKPVU QH VJG NCP IWCIG D[ GUVKOCVKPI VJG HTGSWGPEKGU QH UGSWGPEGU QH n YQTFU 6JG CUUWORVKQP KU ½ OCFG VJCV VJG RTQDCDKNKV[ QH C IKXGP YQTF UVTKPI ¾ ECP DG CR
E\&5&3UHVV//&
6GZV %QTRWU
5RGGEJ %QTRWU
6TCKPKPI
&GEQFKPI
0QTOCNK\CVKQP
/CPWCN 6TCPUETKRVKQP
(GCVWTG 'ZVTCEVKQP
0ITCO 'UVKOCVKQP
6TCKPKPI .GZKEQP
*// 6TCKPKPI
.CPIWCIG /QFGN
4GEQIPK\GT .GZKEQP
#EQWUVKE /QFGNU
2 9
5RGGEJ 5CORNG
;
#EQWUVKE (TQPVGPF
2 *^9
:
H :^*
9
&GEQFGT
5RGGEJ 6TCPUETKRVKQP
FIGURE 5.2 System diagram of a generic speech recognizer based on statistical models, including training and decoding processes and the main knowledge sources. RTQZKOCVGF D[ VJG HQNNQYKPI HQTYCTF UGSWGPVKCN FGEQORQUKVKQP
VJGTGD[ TGFWEKPI VJG YQTF JKUVQT[ VQ VJG RTGEGFKPI YQTFU +V UJQWNF DG PQVGF VJCV QVJGT FGEQORQUKVKQPU QH ECP CNUQ DG CRRTQRTKCVG HQT GZCORNG C DCEMYCTF FGEQORQUKVKQP YKNN NGCF VQ C DCEMYCTF nITCO OQFGN # RTGTGSWKUKVG HQT GUVKOCVKPI n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ſPGF CU 2Z ½
E\&5&3UHVV//&
½
HQT C IKXGP VGZV ½ CPF C VTKITCO ./ KG FGPQVGU VJG NCPIWCIG OQFGN GUVKOCVG QH VJG VGZV RTQDCDKNKV[ 6JG RGTRNGZKV[ FGRGPFU QP DQVJ VJG NCPIWCIG DGKPI OQFGNGF CPF VJG OQFGN KG KV IKXGU C EQODKPGF GUVKOCVG QH JQY IQQF VJG OQFGN KU CPF JQY EQORNGZ VJG NCPIWCIG KU =? +H VJG NGHV QWV FCVC UGV KU TGRTGUGPVCVKXG QH VJG OQFGN VJG RGTRNGZKV[ ECP DG UGGP CU C OGCUWTG QH VJG CXGTCIG DTCPEJKPI HCEVQT KG VJG XQECDWNCT[ UK\G QH C OGOQT[NGUU WPKHQTO NCPIWCIG OQFGN YKVJ UCOG GPVTQR[ CU VJG NCPIWCIG OQFGN WPFGT EQPUKFGTCVKQP
5.3.1 Text Preparation #NVJQWIJ KFGCN NCPIWCIG OQFGN VTCKPKPI FCVC YQWNF EQPUKUV QH NCTIG EQTRQTC QH VTCP UETKDGF CWFKQ FCVC TGRTGUGPVCVKXG QH VJG VCTIGVGF VCUM KP RTCEVKEG UWEJ FCVC CTG FKHſ EWNV VQ QDVCKP 6JGTGHQTG C XCTKGV[ QH QVJGT OQTG QT NGUU ENQUGN[ TGNCVGF VGZV OCVGTKCNU CTG WUWCNN[ WUGF HQT NCPIWCIG OQFGN VTCKPKPI )KXGP C NCTIG VGZV EQTRWU KV OC[ UGGO TGNCVKXGN[ UVTCKIJVHQTYCTF VQ EQPUVTWEV nITCO NCPIWCIG OQFGNU /QUV QH VJG UVGRU CTG RTGVV[ UVCPFCTF CPF OCMG WUG QH VQQNU VJCV EQWPV YQTF UGSWGPEG QEEWTTGPEGU =? 6JG OCKP EQPUKFGTCVKQPU CTG VJG EJQKEG QH VJG XQECDWNCT[ VJG FGſPKVKQP QH YQTFU VTGCVOGPV QH EQORQWPF YQTFU CPF CETQP[OU CPF VJG EJQKEG QH VJG ./ DCEMQHH UVTCVGI[ EH 5GEVKQP 6JGTG KU JQYGXGT C UKIPKſECPV COQWPV QH GHHQTV PGGFGF VQ RTQEGUU QT PQTOCNK\G VJG VGZVU DGHQTG VJG[ ECP DG WUGF 1PG OQVKXCVKQP HQT VJG PQTOCNK\CVKQP KU VQ TGFWEG NGZKECN XCTKCDKNKV[ UQ CU VQ KPETGCUG VJG EQXGTCIG HQT C ſZGF UK\G VCUM XQECDWNCT[ 6JG RTQEGUUKPI FGEKUKQPU CTG IGPGTCNN[ NCPIWCIGURGEKſE 0WOGTKECN GZRTGUUKQPU CPF FCVGU CTG V[RKECNN[ GZRCPFGF VQ CRRTQZKOCVG VJG URQMGP one hundred fifty dollars HQTO CPF VQ TGFWEG VJG NGZKECN XCTKGV[ nineteen ninety one QT one thousand nine hundred and ninety one 5QOG GZCORNG VTCPUHQTOCVKQPU CTG UJQYP KP (KIWTG CNQPI YKVJ VJG TWNG RTQDCDKNKVKGU (QT GZCO RNG VJG YQTF hundred HQNNQYGF D[ C PWODGT ECP DG TGRNCEGF D[ hundred and QH VJG VKOG CPF QH VJG UGSWGPEG million dollars CTG TGRNCEGF YKVJ LWUV VJG YQTF million =?
*70&4'& PD
10' '+)*6* %14214#6+10 +0%14214#6'& 10' *70&4'& /+..+10 &1..#45 $+..+10 &1..#45
*70&4'& #0& PD #0 '+)*6* %142 +0% # *70&4'& /+..+10 $+..+10
FIGURE 5.3 Some example transformation rules applied during text normalization with associated probabilities.
E\&5&3UHVV//&
(WTVJGT UGOKCWVQOCVKE RTQEGUUKPI KU PGEGUUCT[ VQ EQTTGEV HTGSWGPV GTTQTU KPJGTGPV KP VJG VGZVU UWEJ CU QDXKQWU OKUURGNNKPIU million officials QT CTKUKPI HTQO RTQEGUUKPI YKVJ VJG FKUVTKDWVGF VGZV RTQEGUUKPI VQQNU 5QOG PQTOCNK\CVKQPU ECP DG EQPUKFGTGF CU őFGEQORQWPFKPIŒ TWNGU KP VJCV VJG[ OQFKH[ VJG YQTF DQWPFCTKGU CPF VJG VQVCN PWODGT QH YQTFU 6JGUG EQPEGTP VJG RTQEGUUKPI QH CODKIWQWU RWPEVWCVKQP OCTMGTU
UWEJ CU J[RJGP CPF CRQUVTQRJG VJG RTQEGUUKPI QH FKIKV UVTKPIU CPF VTGCVOGPV QH # $ % & +P CIINWVKPCVKXG NCPIWCIGU CDDTGXKCVKQPU CPF CETQP[OU #$%& UWEJ CU )GTOCP FGEQORQWPFKPI TWNGU ECP DG WUGF VQ TGFWEG VJG NGZKECN XCTKGV[ (QT GZCORNG VJG [GCT YJKEJ KU YTKVVGP KP UVCPFCTF )GTOCP CU neunzehnhunderteinundneunzig ECP DG VTCPUHQTOGF KPVQ VJG YQTF UGSWGPEG neunzehn hundert ein und neunzig &GRGPFKPI WRQP VJG VCTIGV CRRNKECVKQP VJG TGEQIPK\GT J[RQVJGUGU OC[ PGGF VQ DG OCRRGF VQ C OQTG CRRTQRTKCVG YTKVVGP HQTO 1VJGT PQTOCNK\CVKQPU UWEJ CU UGPVGPEG KPKVKCN ECRKVCNK\CVKQP CPF ECUG FKUVKPEVKQP MGGR VJG VQVCN PWODGT QH YQTFU WPEJCPIGF DWV TGFWEG ITCRJGOKE XCTKCDKNKV[ +P IGPGTCN C EQORTQOKUG KU OCFG DG VYGGP RTQFWEKPI CP QWVRWV ENQUG VQ VJG UVCPFCTF YTKVVGP HQTO QH VJG NCPIWCIG CPF VJG NGZKECN EQXGTCIG YKVJ VJG ſPCN EJQKEG DGKPI NCTIGN[ CRRNKECVKQPFTKXGP
5.3.2 Vocabulary Selection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ŏU QWVQHXQECDWNCT[ 118 TCVG QP VJG ./ FGXGNQROGPV FCVC 6JGTGHQTG LWFKEKQWU UGNGEVKQP QH VJG FGXGNQROGPV FCVC KU KORQTVCPV 6JG DGUV NGZKECN EQXGTCIG OC[ DG QDVCKPGF D[ UGNGEVKPI VJG XQECD WNCT[ WUKPI QPN[ C UWDUGV QH VJG VTCKPKPI FCVC UWEJ CU VJG OQUV TGEGPV FCVC QT FCVC QP C IKXGP VQRKE KPUVGCF QH WUKPI CNN VJG CXCKNCDNG FCVC = ? #P QDXKQWU YC[ VQ TGFWEG VJG GTTQT TCVG FWG VQ 118U KU VQ KPETGCUG VJG UK\G QH VJG NGZKEQP 7UKPI C XGT[ NCTIG NGZKEQP JCU DGGP UJQYP VQ KORTQXG RGTHQTOCPEG FGURKVG VJG RQVGPVKCN QH KPETGCUGF EQPHWUKDKNKV[ QH VJG NGZKECN GPVTKGU =?
5.3.3 N-gram Estimation
7UKPI VJG OCZKOWO NKMGNKJQQF /. ETKVGTKQP VJG ITCO RTQDCDKNKVKGU CTG GUVK OCVGF HTQO VJG HTGSWGPEKGU QH VJG YQTF UGSWGPEGU QH NGPIVJ KP VJG VTCKPKPI EQTRWU
VGZVU QT URGGEJ VTCPUETKRVKQPU (QT GZCORNG VJG /. GUVKOCVG QH VJG VTKITCO RTQDC DKNKV[ KU IKXGP D[
¾ ½ ¾ ½ YJGTG FGPQVGU VJG PWODGT QH VKOGU VJG ITCO CRRGCTU KP VJG VTCKPKPI FCVC ¾
E\&5&3UHVV//&
½
(QT NCTIG XQECDWNCT[ UK\GU OCP[ QH VJG RQUUKDNG ITCOU YKNN PQV QEEWT KP GXGP C XGT[ NCTIG VTCKPKPI EQTRWU &WG VQ VJG URCTUGPGUU QH VJG FCVC OCZKOWO NKMGNKJQQF GUVKOCVGU CTG ENGCTN[ KPCFGSWCVG CPF PGGF VQ DG UOQQVJGF &KHHGTGPV CRRTQCEJGU JCXG DGGP KPXGUVKICVGF VQ UOQQVJ VJG GUVKOCVGU QH VJG RTQDCDKNKVKGU QH TCTG nITCOU = ? 6JG OQUV EQOOQP CRRTQCEJ KU VQ WUG C DCEMQHH OGEJCPKUO =? YJKEJ TGNKGU QP C NQYGT QTFGT nITCO +H VJGTG KU PQV GPQWIJ FCVC VQ QDVCKP C TQDWUV GUVKOCVG HTQO VJG nITCO EQWPVU C HTCEVKQP QH VJG RTQDCDKNKV[ OCUU KU VCMGP HTQO VJG QDUGTXGF n ITCOU D[ FKUEQWPVKPI VJG /. GUVKOCVGU = ? 6JG RTQDCDKNKVKGU QH VJG TCTG ITCOU CTG VJGP GUVKOCVGF HTQO VJG ITCO RTQDCDKNKVKGU KP C TGEWTUKXG OCPPGT CU UJQYP JGTG HQT C VTKITCO OQFGN
¾ ½ ½ $ ¾ ½ YJGTG $ ¾ ½ KU C DCEMQHH EQGHſEKGPV PGGFGF VQ GPUWTG VJCV VJG RTQDCDKNKV[ UWO HQT C IKXGP EQPVGZV KU GSWCN VQ QPG %QORWVKPI VJG DKITCO GUVKOCVG ½
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ſEKGPVU 9QTF ENCUU QT ECVGIQT[DCUGF NCPIWCIG OQFGNU ECP DG WUGF VQ TGFWEG VJG FGRGP FGPE[ QP VJG VTCKPKPI FCVC )KXGP UQOG VTCKPKPI FCVC CPF C OCRRKPI YJKEJ CUUKIPU GCEJ YQTF VQ C WPKSWG ECVGIQT[ VJG VTCKPKPI VGZV ECP DG VCIIGF CPF VJG ITCO RTQDCDKNKVKGU ·½ ½ YJKEJ CTG QHVGP CRRTQZKOCVGF D[ ·½ ½ ECP DG GUVKOCVGF HTQO VJG TGN CVKXG HTGSWGPEKGU KP VJG UCOG OCPPGT CU C TGIWNCT YQTF ITCO 6JG ENCUU CUUKIP OGPV KU QHVGP QDVCKPGF D[ OKPKOK\KPI VJG RGTRNGZKV[ QH C DKITCO ECVGIQT[ OQFGN HQT C IKXGP PWODGT QH YQTF ECVGIQTKGU = ? +V KU CNUQ EQOOQP RTCEVKEG VQ KPVGTRQNCVG VJG ECVGIQT[ ./ YKVJ VJG ITCO ./ KP QTFGT VQ QDVCKP C NQYGT RGTRNGZKV[ VJCP VJCV QH VJG TGIWNCT ITCO OQFGN 6JG TGUWNVKPI VTKITCO RTQDCDKNKV[ GUVKOCVGU CTG
¾
½
¾
½
¾ ½
1VJGT UVCVKUVKECN NCPIWCIG OQFGNU JCXG DGGP KPXGUVKICVGF D[ OCRRKPI VJG YQTF JKU VQT[ ½ ½ QPVQ GSWKXCNGPEG ENCUUGU QVJGT VJCP VJG ENCUUKECN ITCOU
E\&5&3UHVV//&
*QYGXGT VJGUG OQFGNKPI VGEJPKSWGU UWEJ CU FGEKUKQP VTGG OQFGNU OCZKOWO GP VTQR[ OQFGNU QT NKPIWKUVKECNN[ OQVKXCVGF OQFGNU RTQDCDKNKUVKE EQPVGZVHTGG CPF NKPM ITCOOCTU JCXG DGGP WUGF YKVJ OQFGTCVG UWEEGUU NGCFKPI VQ UOCNN ICKPU QXGT VJG OWEJ UKORNGT ITCO OQFGN =?
5.3.4 LM Adaptation .8%54 U[UVGOU WUG QPG QT OQTG NCPIWCIG OQFGNU DWV VJGUG ./U CTG WUWCNN[ UVCVKE GXGP VJQWIJ VJG EJQKEG QH YJKEJ OQFGN VQ WUG ECP DG F[PCOKE FGRGPFGPV HQT GZ CORNG QP VJG FKCNQI UVCVG .CPIWCIG OQFGN CFCRVCVKQP KU QH KPVGTGUV HQT KORTQXKPI VJG OQFGN CEEWTCE[ CPF HQT MGGRKPI VJG OQFGNU WRVQFCVG 8CTKQWU CRRTQCEJGU JCXG DGGP VCMGP VQ CFCRV VJG NCPIWCIG OQFGN DCUGF QP VJG QDUGTXGF VGZV UQ HCT KPENWFKPI VJG WUG QH C cache model = ? C trigger model =? QT topic coherence modeling =? 6JG ECEJG OQFGN KU DCUGF QP VJG KFGC VJCV YQTFU CRRGCTKPI KP C FQEWOGPV YKNN JCXG CP KPETGCUGF RTQDCDKNKV[ QH CRRGCTKPI CICKP KP VJG UCOG FQEWOGPV (QT UJQTV FQEWOGPVU VJG PWODGT QH YQTFU CRRGCTKPI KU NKOKVGF CPF CU C EQPUGSWGPEG VJG DGPGſ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nITCO OQFGNU
5.4 Pronunciation Modeling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
E\&5&3UHVV//&
Phone Example Phone Example 8QYGNU (TKECVKXGU K DGGV U UWG DKV \ \QQ G DCKV UJQG D'V OGCUWTG ¿ DCV H HCP DWV X XCP DQVV VJKP Q DQCV 2NQUKXGU W DQQV D DGV DQQM F FGDV DKTF
IGV &KRJVJQPIU R RGV DKVG V VCV DQ[ M ECV DQWV #HHTKECVGU 4GFWEGF 8QYGNU EJGCR
ZDQWV LGGR FCVGF 0CUCNU DWVVGT O OGV 5GOKXQYGNU P PGV N NGF VJKPI T TGF 5[NNCDKEU Y YGF O DQVV QO [ [GV P DWVVQP J JCV N DQVVNG FIGURE 5.4 Set of 45 phone symbols for English with illustrative words, with the portion corresponding to the phone sound underlined.
NGCUV RCTVKCNN[ ETGCVGF OCPWCNN[ UGXGTCN CRRTQCEJGU VQ CWVQOCVKECNN[ NGCTP CPF IGP GTCVG YQTF RTQPWPEKCVKQPU JCXG DGGP KPXGUVKICVGF 5WEJ CRRTQCEJGU YJKNG RTQOKU KPI JCXG VQ FCVG IKXGP QPN[ UOCNN RGTHQTOCPEG KORTQXGOGPVU GXGP YJGP VTCKPGF QP OCPWCN VTCPUETKRVKQPU =? 2TQPWPEKCVKQP XCTKCPVU ECP DG QDUGTXGF HQT C XCTKGV[ QH YQTFU #NVGTPCVKXG RTQPWPEKC VKQPU CTG QDXKQWUN[ PGGFGF HQT JQOQITCRJU YQTFU URGNNGF VJG UCOG DWV RTQPQWPEGF FKHHGTGPVN[ YJKEJ TGƀGEV FKHHGTGPV RCTVU QH URGGEJ XGTD QT PQWP UWEJ CU excuse, record, moderate 5QOG HTGSWGPV CHſZGU UWEJ CU anti-, bi-, multi-, -ization ECP DG RTQPQWPEGF YKVJ C FKRJVJQPI QT C UJQTV XQYGN QT 6JG WRRGT RCTV QH (KIWTG IKXGU UQOG GZCORNG YQTFU YKVJ OWNVKRNG RTQPWPEKCVKQPU CPF VJGKT CUUQEK CVGF RTQDCDKNKVKGU 7UKPI C UGV QH CNNQRJQPG OQFGNU EH 5GEVKQP VJG RTQPWP EKCVKQP RTQDCDKNKVKGU CTG GUVKOCVGF D[ ſTUV CNKIPKPI VJG TGHGTGPEG YQTF VTCPUETKRVKQP
E\&5&3UHVV//&
%17210 14)#0+<#6+10 *70&4'& /1&'4#6' 61 + &10 ŏ 6 -019 &10 ŏ 6 -019 &+& ;17 )1+0) 61
MWRP M[WRP TIP\GP TIP \GP JPFF JPFTF JPF JPTF OFV OFGV V VW FQPPQ FQPVPQ FPQ FPQ
FQPPQ FQPVPQ FPQ F+F[W F+ F+F[ IQ V IQ VW IP IEP
FIGURE 5.5 Some example lexical entries and their pronunciations along with estimate probabilities. For the compound words, the original concatenated pronunciation is given in the 1st line and the reduced forms are given in the 2nd line.
YKVJ VJG CWFKQ UKIPCN WUKPI C NGZKEQP EQPVCKPKPI GSWCNN[ NKMGN[ CNVGTPCVKXG RTQPWP EKCVKQPU NGVVKPI VJG 8KVGTDK CNIQTKVJO EJQQUG VJG DGUV RTQPWPEKCVKQP HQT GCEJ YQTF 6JG RTQDCDKNKVKGU CTG VJGP GUVKOCVGF HTQO VJG TGNCVKXG HTGSWGPEKGU QH GCEJ XCTKCPV 9QTFU QH HQTGKIP QTKIKP RCTVKEWNCTN[ RTQRGT PCOGU OC[ JCXG FKHHGTGPV RTQPWPEKC VKQPU FGRGPFKPI WRQP VJG URGCMGTŏU HCOKNKCTKV[ YKVJ VJG QTKIKPCN NCPIWCIG +V KU CNUQ EQOOQP HQT OWNVKU[NNCDNKE YQTFU VQ DG RTQPQWPEGF YKVJ FKHHGTGPV PWODGTU QH U[N NCDNGU (QT GZCORNG CDQWV QH VJG QEEWTTGPEGU QH interest CPF conference CPF QH company CTG URQMGP YKVJ VYQ U[NNCDNGU KPUVGCF QH VJTGG +H CEQWUVKE OQFGN VTCKPKPI KU ECTTKGF QWV YKVJQWV CNNQYKPI HQT CRRTQRTKCVG RTQPWPEKCVKQP XCTKCPVU VJGTG YKNN PGEGUUCTKN[ DG C OKUCNKIPOGPV QH QPG QT OQTG RJQPGU OCMKPI VJG RJQPG OQF GNU NGUU CEEWTCVG 'ZRGTKGPEG JCU UJQYP VJCV ECTGHWN NGZKECN FGUKIP KORTQXGU URGGEJ TGEQIPKVKQP U[UVGO RGTHQTOCPEG =? +P URGGEJ HTQO HCUV URGCMGTU QT URGCMGTU YKVJ TGNCZGF URGCMKPI UV[NGU KV KU EQOOQP VQ QDUGTXG RQQTN[ CTVKEWNCVGF QT UMKRRGF WPUVTGUUGF U[NNCDNGU RCTVKEWNCTN[ KP NQPI YQTFU YKVJ UGSWGPEGU QH WPUVTGUUGF U[NNCDNGU #NVJQWIJ UWEJ NQPI YQTFU CTG V[RK ECNN[ YGNN TGEQIPK\GF QHVGP C PGCTD[ HWPEVKQP YQTF KU FGNGVGF 6Q TGFWEG VJGUG MKPFU QH GTTQTU CNVGTPCVG RTQPWPEKCVKQPU KP VJG NGZKEQP ECP CNNQY UEJYCFGNGVKQP QT U[N NCDKE EQPUQPCPVU KP WPUVTGUUGF U[NNCDNGU %QORQWPF YQTFU JCXG CNUQ DGGP WUGF CU C YC[ VQ TGRTGUGPV TGFWEGF HQTOU HQT EQOOQP YQTF UGSWGPEGU UWEJ CU don’t know did you CPF going to 5QOG QH VJG TGFWEGF HQTOU CTG UQ HTGSWGPV VJCV VJG[ JCXG C EQOOQPN[ CEEGRVGF YTKVVGP HQTO gonna, dunno 5QOG GZCORNG EQORQWPF YQTFU CTG UJQYP KP VJG NQYGT RCTV QH (KIWTG CNQPI YKVJ GUVKOCVGU QH VJG RTQPWPEKCVKQP RTQDCDKNKVKGU HQT VJG FKHHGTGPV XCTKCPVU 6JGUG GZCORNGU KNNWUVTCVG VJG KPVGTGUV KP WUKPI EQORQWPF YQTFU KP TGEQIPKVKQP NGZKEQPU (NWGPV URGGEJ GHHGEVU ECP CNVGTPCVKXGN[ DG
E\&5&3UHVV//&
OQFGNGF WUKPI RJQPQNQIKECN TWNGU = ? 6JG RTKPEKRNG DGJKPF VJG RJQPQNQIK ECN TWNGU KU VQ OQFKH[ VJG CNNQYCDNG RJQPG UGSWGPEGU VQ VCMG KPVQ CEEQWPV GZRGEVGF XCTKCVKQPU 6JGUG TWNGU CTG QRVKQPCNN[ CRRNKGF FWTKPI VTCKPKPI CPF TGEQIPKVKQP 7UKPI RJQPQNQIKECN TWNGU FWTKPI VTCKPKPI TGUWNVU KP DGVVGT CEQWUVKE OQFGNU CU VJG[ CTG NGUU őRQNNWVGFŒ D[ YTQPI VTCPUETKRVKQPU 6JGKT WUG FWTKPI TGEQIPKVKQP TGFWEGU VJG PWODGT QH OKUOCVEJGU 6JG UCOG OGEJCPKUO JCU DGGP WUGF VQ JCPFNG NKCKUQPU OWVGG CPF ſPCN EQPUQPCPV ENWUVGT TGFWEVKQP HQT (TGPEJ #U URGGEJ TGEQIPKVKQP TGUGCTEJ JCU OQXGF HTQO TGCF URGGEJ VQ HQWPF CWFKQ FCVC VJG RJQPG UGV JCU DGGP GZRCPFGF VQ KPENWFG PQPURGGEJ GXGPVU 6JGUG ECP EQTTGURQPF VQ PQKUGU RTQFWEGF D[ VJG URGCMGT DTGCVJ PQKUG EQWIJKPI UPGG\KPI NCWIJVGT GVE QT ECP EQTTGURQPF VQ GZVGTPCN UQWTEGU OWUKE OQVQT VCRRKPI GVE
5.5 Acoustic Modeling 1PG QH VJG OCKP EJCNNGPIGU QH CEQWUVKE OQFGNKPI KU VQ JCPFNG VJG XCTKCDKNKV[ RTGUGPV KP VJG URGGEJ UKIPCN 8CTKCDKNKV[ ECP CTKUG HTQO VJG NKPIWKUVKE EQPVGZV QT ECP DG CUUQEKCVGF YKVJ VJG PQPNKPIWKUVKE EQPVGZV UWEJ CU VJG URGCMGT GI RJ[UKECN EJCTCE VGTKUVKEU URGCMKPI UV[NG OQQF GVE CPF VJG CEQWUVKE GPXKTQPOGPV GI DCEMITQWPF PQKUG OWUKE CPF TGEQTFKPI EJCPPGN GI FKTGEV OKETQRJQPG VGNGRJQPG /QUV UVCVG QHVJGCTV .8%54 U[UVGOU OCMG WUG QH JKFFGP /CTMQX OQFGNU *//U HQT CEQWUVKE OQFGNKPI = ? YJKEJ EQPUKUVU QH OQFGNKPI VJG RTQDCDKNKV[ FGPUKV[ HWPEVKQP QH C UGSWGPEG QH CEQWUVKE HGCVWTG XGEVQTU 1VJGT CRRTQCEJGU KPENWFG UGIOGPV DCUGF OQFGNU = ? CPF PGWTCN PGVYQTMU = ? VQ GUVKOCVG VJG CEQWUVKE QDUGTXC VKQP NKMGNKJQQFU 9KVJ GZEGRVKQP QH VJG CEQWUVKE NKMGNKJQQF EQORWVCVKQP CNN U[UVGOU OCMG WUG QH VJG *// HTCOGYQTM VQ EQODKPG NKPIWKUVKE CPF CEQWUVKE KPHQTOCVKQP KP C UKPING PGVYQTM TGRTGUGPVKPI CNN RQUUKDNG UGPVGPEGU
5.5.1 Acoustic Front-end 6JG ſ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ſEKGPVU QDVCKPGF YKVJ C /GN (TG SWGPE[ %GRUVTCN /(% CPCN[UKU =? QT YKVJ C 2GTEGRVWCN .KPGCT 2TGFKEVKQP 2.2
E\&5&3UHVV//&
CPCN[UKU =? %GRUVTCN RCTCOGVGTU CTG NGUU EQTTGNCVGF VJCP FKTGEV URGEVTCN EQORQ PGPVU YJKEJ UKORNKſGU GUVKOCVKQP QH VJG CEQWUVKE OQFGN RCTCOGVGTU D[ TGFWEKPI VJG PGGF HQT OQFGNKPI VJG FGRGPFGPE[ DGVYGGP HGCVWTGU +P DQVJ ECUGU C /GN UECNG UJQTV VGTO RQYGT URGEVTWO KU GUVKOCVGF QP C ſZGF YKPFQY WUWCNN[ KP VJG TCPIG QH VQ OU +P QTFGT VQ CXQKF URWTKQWU JKIJ HTGSWGPE[ EQORQPGPVU KP VJG URGEVTWO FWG VQ FKUEQPVKPWKVKGU ECWUGF D[ YKPFQYKPI VJG UKIPCN KV KU EQOOQP VQ WUG C VCRGTGF YKP FQY UWEJ CU C *COOKPI YKPFQY 6JG YKPFQY KU VJGP UJKHVGF CPF VJG PGZV HGCVWTG XGEVQT EQORWVGF 6JG OQUV EQOOQPN[ WUGF QHHUGV KU OU 6JKU CEQWUVKE RCTCOGVGT K\CVKQP EQPXGTVU VJG URGGEJ UKIPCN KPVQ C UGSWGPEG QH HGCVWTG XGEVQTU GCEJ XGEVQT TGRTGUGPVKPI C OU KPVGTXCN TGHGTTGF VQ CU C HTCOG QT C HGCVWTG XGEVQT
ܽ ܾ ÜÌ 6JG /GN UECNG CRRTQZKOCVGU VJG HTGSWGPE[ TGUQNWVKQP QH VJG JWOCP CWFKVQT[ U[UVGO DGKPI NKPGCT KP VJG NQY HTGSWGPE[ TCPIG DGNQY *\ CPF NQICTKVJOKE CDQXG *\ 6JG EGRUVTCN RCTCOGVGTU CTG QDVCKPGF D[ VCMKPI CP KPXGTUG VTCPUHQTO QH VJG NQI QH VJG ſNVGTDCPM RCTCOGVGTU +P VJG ECUG QH VJG /(% EQGHſEKGPVU C EQUKPG VTCPUHQTO KU CRRNKGF VQ VJG NQI RQYGT URGEVTWO YJGTGCU C TQQV.KPGCT 2TGFKEVKXG %QFKPI .2% CPCN[UKU KU WUGF VQ QDVCKP VJG 2.2 EGRUVTWO EQGHſ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ſGF HQTO QH EGRUVTCN UWDVTCEVKQP ECP DG ECTTKGF QWV YJGTG C TWPPKPI OGCP KU EQORWVGF HTQO VJG 0 NCUV HTCOGU 0 KU QHVGP QP VJG QTFGT QH EQTTGURQPFKPI VQ U QH URGGEJ +V KU CNUQ EQOOQP VQ PQTOCNK\G VJG HGCVWTG XCTKCPEG UQ VJCV GCEJ TGUWNVKPI EGRUVTCN EQGHſEKGPV JCU C WPKV[ XCTKCPEG +P QTFGT VQ ECRVWTG VJG F[PCOKE PCVWTG QH VJG URGGEJ UKIPCN VJG HGCVWTG XGEVQT KU WUWCNN[ CWIOGPVGF YKVJ őFGNVCŒ RCTCOGVGTU 6JG FGNVC RCTCOGVGTU CTG EQORWVGF D[ VCMKPI VJG ſTUV CPF UGEQPF FKHHGTGPEGU QH VJG HGCVWTGU KP UWEEGUUKXG HTCOGU #U C TGUWNV C V[RKECN HGCVWTG XGEVQT Ü Ø YKNN KPENWFG EGRUVTWO EQGHſEKGPVU RNWU VJG PQT OCNK\GF NQIGPGTI[ CNQPI YKVJ VJG ſTUV CPF UGEQPF QTFGT FGTKXCVKXGU KG C VQVCN QH EQORQPGPVU +PUVGCF QH WUKPI VJGUG ſ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ſTUV FGEQFKPI RCUU VTCP UETKRVKQP CPF UQOG CEQWUVKE OQFGNU =? 86.0 OWUV CNUQ DG CRRNKGF FWTKPI VJG VTCKPKPI RTQEGUU VQ QDVCKP OQFGNU UWKVGF VQ FGEQFG VJG PQTOCNK\GF VGUV FCVC 6JKU
E\&5&3UHVV//&
C
C
C
C
C
FIGURE 5.6 A simple 3-state left-to-right HMM topology commonly used for allophone modeling in LVCSR. The model generates at least 3 speech frames per allophone, resulting in a minimal phone segment duration of 30ms for frame rate of 100Hz.
PQTOCNK\CVKQP JCU DGGP UJQYP VQ IKXG UKIPKſECPV GTTQT TCVG TGFWEVKQP KP RCTVKEWNCT QP VGNGRJQPG EQPXGTUCVKQPCN URGGEJ =?
5.5.2 Modeling Allophones /QFGNKPI CNNQRJQPGU YKVJ *KFFGP /CTMQX OQFGNU KU RQRWNCT DGECWUG VJGUG OQFGNU YQTM TGCUQPCDN[ YGNN CPF VJGKT RCTCOGVGTU ECP DG GHſ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ſIWTCVKQPU JCXG VQ GOKVVKPI UVCVGU RGT CNNQRJQPG OQFGN YJGTG VJG PWODGT QH UVCVGU KORQUGU C OKPKOCN FWTCVKQP HQT VJG RJQPG 5QOG EQPſIWTCVKQPU CNNQY EGTVCKP UVCVGU VQ DG UMKRRGF VJGTGD[ TGFWEKPI VJG TGSWKTGF OKPKOCN FWTCVKQP 6JG RTQDCDKNKV[ QH CP QDUGTXCVKQP KG C URGGEJ XGEVQT KU CUUWOGF VQ DG FGRGPFGPV QPN[ QP VJG EWTTGPV UVCVG
)KXGP CP UVCVG *// YKVJ RCTCOGVGT XGEVQT VJG *// UVQEJCUVKE RTQEGUU KU FGUETKDGF D[ VJG HQNNQYKPI LQKPV RTQDCDKNKV[ FGPUKV[ HWPEVKQP QH VJG QDUGTXGF UKIPCN ܽ Ü CPF VJG WPQDUGTXGF UVCVG UGSWGPEG
¼
½
Ü
YJGTG KU VJG KPKVKCN RTQDCDKNKV[ QH UVCVG KU VJG VTCPUKVKQP RTQDCDKNKV[ HTQO UVCVG VQ UVCVG CPF KU VJG GOKVVKPI 2&( CUUQEKCVGF YKVJ GCEJ UVCVG (KIWTG UJQYU VJG VTCPUKVKQP UVTWEVWTG QH C UVCVG NGHVVQTKIJV *// VQRQNQI[ EQOOQPN[ WUGF HQT CNNQRJQPG OQFGNKPI KP .8%54
6JG OQUV HTGSWGPVN[ WUGF UVCVG QWVRWV 2&( HQT URGCMGTKPFGRGPFGPV U[UVGOU KU C
E\&5&3UHVV//&
5+56'4 /sst/ VTKRJQPGU s(*,) (s,s) s(,t) t(s,) (t,*) SWKPRJQPGU s(*,s) (s,st) s(s,t) t(s,) (st,*) FIGURE 5.7 Examples of allophonic transcriptions in terms of intra-word triphones and quinphones. Each contextual unit is defined by the central phone followed by its phone context shown in parentheses (left-context, right-context). * is a wildcard signifying any context. OKZVWTG QH )CWUUKCPU YKVJ VQ EQORQPGPVU Ü
Ü
YJGTG CPF FGPQVG TGURGEVKXGN[ VJG OGCP XGEVQT VJG EQXCTKCPEG OCVTKZ CPF VJG OKZVWTG YGKIJV QH VJG VJ )CWUUKCP EQORQPGPV QH UVCVG 6Q TGFWEG VJG PWODGT QH RCTCOGVGTU CPF VJG KPJGTGPV GUVKOCVKQP RTQDNGO NKPMGF VQ HWNN EQXCTKCPEG OCVTKEGU VJG EQXCTKCPEG OCVTKEGU CTG WUWCNN[ CUUWOGF VQ DG FKCIQPCN 4GEGPVN[ KV JCU DGGP FGOQPUVTCVGF VJCV PQPFKCIQPCN EQXCTKCPEG OCVTKEGU ECP DG WUGF YJKNG MGGRKPI VJG GUVKOCVKQP RTQDNGO OCPCIGCDNG = ? 2JQPG DCUGF OQFGNU QHHGT VJG CFXCPVCIG VJCV TGEQIPKVKQP NGZKEQPU ECP DG FGUETKDGF WUKPI VJG GNGOGPVCT[ WPKVU QH VJG IKXGP NCPIWCIG EH CPF VJWU ECP DGPGſV HTQO OCP[ NKPIWKUVKE UVWFKGU +V KU QH EQWTUG RQUUKDNG VQ RGTHQTO URGGEJ TGEQIPKVKQP YKVJ QWV WUKPI C RJQPGOKE NGZKEQP GKVJGT D[ WUG QH YQTF OQFGNU CU YCU VJG OQTG EQO OQPN[ WUGF CRRTQCEJ [GCTU CIQ QT C FKHHGTGPV OCRRKPI UWEJ CU HGPQPGU YJKEJ CTG UOCNN FCVCFTKXGP CEQWUVKE WPKVU =? %QORCTGF VQ YQTF OQFGNU UWDYQTF WPKVU TGFWEG VJG PWODGT QH RCTCOGVGTU GPCDNG ETQUU YQTF OQFGNKPI CPF HCEKNKVCVG RQTVKPI VQ PGY XQECDWNCTKGU (GPQPGU QHHGT VJG CFFKVKQPCN CFXCPVCIG QH CWVQOCVKE VTCKPKPI DWV NCEM VJG CDKNKV[ VQ KPENWFG a priori NKPIWKUVKE MPQYNGFIG # IKXGP *// ECP TGRTGUGPV C RJQPG YKVJQWV EQPUKFGTCVKQP QH KVU PGKIJDQTU EQPVGZV KPFGRGPFGPV OQFGN QT C RJQPG KP C RCTVKEWNCT EQPVGZV CNNQRJQPG OQFGN 8CTKQWU V[RGU QH EQPVGZVU JCXG DGGP KPXGUVKICVGF HTQO C UKPING RJQPG EQPVGZV TKIJV QT NGHV EQPVGZV NGHV CPF TKIJVEQPVGZV VTKRJQPG RQUKVKQPFGRGPFGPV VTKRJQPGU ETQUUYQTF CPF YKVJKP YQTF VTKRJQPGU HWPEVKQP YQTF VTKRJQPGU CPF SWKPRJQPGU =? 6JG EQPVGZV OC[ QT OC[ PQV KPENWFG VJG RQUKVKQP QH VJG RJQPG YKVJKP VJG YQTF YQTF RQUKVKQP FGRGPFGPV CPF YQTFKPVGTPCN CPF ETQUUYQTF EQPVGZVU OC[ DG OGTIGF QT EQPUKFGTGF CU UGRCTCVG OQFGNU &KHHGTGPV CRRTQCEJGU CTG WUGF VQ UGNGEV VJG EQPVGZ VWCN WPKVU DCUGF QP HTGSWGPE[ QH QEEWTTGPEG CPF ENWUVGTKPI VGEJPKSWGU 6JG QRVKOCN UGV QH OQFGNGF EQPVGZVU KU WUWCNN[ VJG TGUWNV QH C VTCFGQHH DGVYGGP TGUQNWVKQP CPF TQ DWUVPGUU CPF KU JKIJN[ FGRGPFGPV QP VJG CXCKNCDNG VTCKPKPI FCVC 6JKU QRVKOK\CVKQP KU IGPGTCNN[ FQPG D[ OKPKOK\KPI VJG TGEQIPK\GT GTTQT TCVG QP UQOG FGXGNQROGPV FCVC 7UKPI EQPVGZVWCN RJQPG OQFGNU ECP DG UGGP CU TGRNCEKPI VJG RJQPG VTCPUETKRVKQP CU URGEKſGF KP VJG RTQPWPEKCVKQP FKEVKQPCT[ D[ C VTCPUETKRVKQP KP VGTOU QH CNNQRJQPGU
E\&5&3UHVV//&
Position: UVCVGRQUKVKQP YQTFDGIKP YQTFGPF OQPQRJQPG General classes: XQYGN EQPUQPCPV EQPVKPWCPV UQPQTCPV XQKEGFEQPUQPCPV XQKEGNGUU HTKECVKXG UVTKFGPV UVQR PCUCN UGOKXQYGN CURKTCVGF CPVGTKQT JKIJ EQTQPCN UNCEM TQWPFGF VGPUG TGVTQƀGZ U[NNCDKE ſNNGTU Vowel classes: JKIJXQYGN NQYXQYGN TQWPFGFXQYGN VGPUGXQYGN TGFWEGF FKRJVJQPI HTQPVXQYGN DCEMXQYGN NQPIXQYGN UJQTVXQYGN TGVTQƀGZXQYGN FKRJVJQPI(WR FKRJVJQPI(FQYP Consonant classes: NCDKCN FGPVCN CNXGQNCT RCNCVCN XGNCT CHHTKECVG Individual phones: UGG (KIWTG
FIGURE 5.8 Example questions used for decision tree clustering.
(KIWTG IKXGU VJG VTKRJQPG CPF SWKPRJQPG VTCPUETKRVKQPU HQT VJG YQTF 5+56'4 WU KPI QPN[ YQTF KPVGTPCN WPKVU KG VJG CNNQRJQPKE VTCPUETKRVKQP KU KPFGRGPFGPV QH VJG YQTF EQPVGZV 9JGP WUKPI ETQUUYQTF VTKRJQPGU VJG OQFGNU WUGF HQT VJG ſTUV CPF NCUV RJQPG QH GCEJ YQTF QT VJG ſTUV CPF NCUV VYQ RJQPGU KP VJG ECUG QH SWKPRJQPGU FGRGPF QP VJG YQTF EQPVGZV OCMKPI VJG FGEQFKPI RTQDNGO UKIPKſECPVN[ OQTG EQO RNGZ # RQYGTHWN VGEJPKSWG VQ MGGR VJG OQFGNU VTCKPCDNG YKVJQWV UCETKſEKPI OQFGN TGUQ NWVKQP KU VQ VCMG CFXCPVCIG QH VJG UVCVG UKOKNCTKV[ COQPI FKHHGTGPV OQFGNU QH C IKXGP RJQPG D[ V[KPI VJG *// UVCVG FKUVTKDWVKQPU 6JKU DCUKE KFGC KU WUGF KP OQUV EWT TGPV U[UVGOU CNVJQWIJ VJGTG CTG UNKIJV FKHHGTGPEGU KP VJG KORNGOGPVCVKQP CPF KP VJG PCOKPI QH VJG TGUWNVKPI ENWUVGTGF UVCVGU senones =? genones =? PELs =? tiedstates =? +P RTCEVKEG DQVJ CIINQOGTCVKXG ENWUVGTKPI CPF FKXKUKXG ENWUVGTKPI JCXG DGGP HQWPF VQ [KGNF OQFGN UGVU YKVJ EQORCTCDNG RGTHQTOCPEG &KXKUKXG FGEKUKQP VTGG ENWUVGTKPI KU RCTVKEWNCTN[ KPVGTGUVKPI YJGP VJGTG CTG C XGT[ NCTIG PWODGT QH UVCVGU VQ ENWUVGT UKPEG KV KU CV VJG UCOG VKOG DQVJ HCUVGT CPF OQTG TQDWUV VJCP C DQVVQO WR ITGGF[ CNIQTKVJO CPF VJGTGHQTG OWEJ GCUKGT VQ VWPG +P CFFKVKQP *// UVCVG V[KPI DCUGF QP FGEKUKQP VTGG ENWUVGTKPI JCU VJG CFXCPVCIG QH RTQXKFKPI C OGCPU VQ DWKNF OQFGNU HQT WPUGGP EQPVGZVU KG VJQUG EQPVGZVU VJCV FQ PQV QEEWT KP VJG VTCKP KPI FCVC = ? 6JG UGV QH SWGUVKQPU V[RKECNN[ EQPEGTP VJG RJQPG RQUKVKQP VJG FKUVKPEVKXG HGCVWTGU CPF KFGPVKVKGU QH VJG RJQPG CPF VJG PGKIJDQTKPI RJQPGU =? CU UJQYP KP (KIWTG 6JG OQUV HTGSWGPVN[ WUGF SWGUVKQPU HQT C NCTIG #OGTKECP 'PINKUJ OQFGN UGV CTG IKXGP KP (KIWTG
5.5.3 HMM Parameter Estimation #EQWUVKE OQFGN VTCKPKPI EQPUKUVU QH GUVKOCVKPI VJG RCTCOGVGTU QH GCEJ *// HTQO VJG CXCKNCDNG VTCKPKPI FCVC (QT )CWUUKCP OKZVWTG *//U VJKU TGSWKTGU GUVKOCVKPI VJG OGCPU CPF EQXCTKCPEG OCVTKEGU VJG OKZVWTG YGKIJVU CPF VJG VTCPUKVKQP RTQDCDKNKVKGU +H KU VJG RCTCOGVGT XGEVQT QH VJG *//U VQ DG VTCKPGF QP UQOG FCVC VJG OCZK
E\&5&3UHVV//&
Question
Log likelihood gain Question
XQYGN= ? UQPQTCPV= ? UQPQTCPV=? HTQPVXQYGN= ? UGOKXQYGN= ? XQKEGFEQPUQPCPV= ? YQTFDQF[RQU=? PCUCN= ? XQKEGNGUU= ? YQTFDGIKPRQU=?
RJQPGT= ? RJQPG*= ? UVTKFGPV= ? RJQPGN PCUCN=? XQYGN=? JKIJXQYGN= ? XQKEGNGUU=? RJQPGP= ? RJQPGU= ??
Log likelihood gain
FIGURE 5.9 The most frequently used decision tree questions for an American English broadcast news transcription system [40]. The [+1] and [-1] indicate that the question has been applied to the right or left context respectively, and [0] to the phone itself. OWO NKMGNKJQQF /. GUVKOCVG KU
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ſPFKPI VJG VTWG /. RCTCOGVGT XCNWGU CPF GXGP YJGP VJG VTWG /. GUVKOCVGU CTG QDVCKPGF VJG[ OC[ PQV DG VJG DGUV QPGU HQT URGGEJ TGEQIPKVKQP 6JGTGHQTG UQOG KORNGOGPVCVKQP FGVCKNU UWEJ CU C RTQRGT KPKVKCNK\CVKQP RTQEGFWTG CPF VJG WUG QH EQPUVTCKPVU QP VJG RCTCOGVGT XCNWGU CTG SWKVG KORQTVCPV 5KPEG VJG IQCN QH VTCKPKPI KU VQ ſPF VJG DGUV OQFGN VQ CEEQWPV HQT VJG QDUGTXGF FCVC VJG RGTHQTOCPEG QH VJG TGEQIPK\GT KU ETKVKECNN[ FGRGPFGPV WRQP VJG TGRTGUGPVCVKXGN[ QH VJG VTCKPKPI FCVC 5QOG OGVJQFU VQ TGFWEG VJKU FGRGPFGPE[ CTG FKUEWUUGF DGNQY KP VJG UWDUGEVKQP QP *// CFCRVCVKQP 5RGCMGTKPFGRGPFGPEG KU QDVCKPGF D[ GUVKOCV KPI VJG RCTCOGVGTU QH VJG CEQWUVKE OQFGNU QP NCTIG URGGEJ EQTRQTC EQPVCKPKPI FCVC HTQO C NCTIG URGCMGT RQRWNCVKQP 5KPEG VJGTG CTG UWDUVCPVKCN FKHHGTGPEGU KP URGGEJ
E\&5&3UHVV//&
HTQO OCNG CPF HGOCNG VCNMGTU KV KU EQOOQP RTCEVKEG VQ WUG UGRCTCVG OQFGNU HQT OCNG CPF HGOCNG URGGEJ KP QTFGT VQ KORTQXG TGEQIPKVKQP RGTHQTOCPEG 6JGUG FKHHGTGPEGU ECP DG CVVTKDWVGF VQ CPCVQOKECN FKHHGTGPEGU QP CXGTCIG HGOCNGU JCXG C UJQTVGT XQECN VTCEV NGPIVJ TGUWNVKPI KP JKIJGT HQTOCPV HTGSWGPEKGU CU YGNN CU C JKIJGT HWPFCOGPVCN HTGSWGPE[ CPF UQEKCN QPGU HGOCNG XQKEG KU QHVGP őDTGCVJKGTŒ ECWUGF D[ KPEQORNGVG ENQUWTG QH VJG XQECN HQNFU 6JG IGPFGTFGRGPFGPV OQFGNU CTG QHVGP QDVCKPGF HTQO URGCMGTKPFGRGPFGPV UGGF OQFGNU WUKPI /CZKOWO A Posteriori GUVKOCVQTU =? EH VJG PGZV UGEVKQP QP *// CFCRVCVKQP 6JG IGPFGTFGRGPFGPV OQFGNU ECP DG HWTVJGT CFCRVGF VQ GCEJ URGEKſE URGCMGT )GPFGTFGRGPFGPV OQFGNKPI KU LWUV QPG GZCORNG QH VJG HCOKN[ QH CFCRVKXG VTCKPKPI UEJGOGU YJKEJ CTG RCTVKEWNCTN[ YGNNUWKVGF VQ JGV GTQIGPGQWU VTCKPKPI FCVC UWEJ CU DTQCFECUV PGYU TGEQTFKPIU YJKEJ KPENWFG C YKFG XCTKGV[ QH CEQWUVKECN EQPFKVKQPU URGCMGT V[RGU CPF URGCMKPI UV[NGU #FCRVKXG VTCKP KPI OCMGU WUG QH *// CFCRVCVKQP VGEJPKSWGU CHVGT RCTVKVKQPKPI VJG VTCKPKPI FCVC CEEQTFKPI VQ CEQWUVKE EQPFKVKQPU CPF URGCMGT ENWUVGTU 5KPEG CP *// KU HCT HTQO DGKPI VJG EQTTGEV OQFGN QH VJG QDUGTXGF FCVC CPF VJGTG KU QPN[ C NKOKVGF COQWPV QH FCVC CXCKNCDNG VQ GUVKOCVG KVU RCTCOGVGTU KV ECP DG CFXCP VCIGQWU VQ TGRNCEG /. VTCKPKPI YKVJ CP CNVGTPCVKXG FKUETKOKPCVKXG VTCKPKPI UEJGOG 6GEJPKSWGU HQT NCTIGUECNG FKUETKOKPCVKXG VTCKPKPI QH VJG CEQWUVKE OQFGNU WUKPI VJG /CZKOWO /WVWCN +PHQTOCVKQP 'UVKOCVKQP //+' ETKVGTKQP KP RNCEG QH EQPXGPVKQPCN /. GUVKOCVKQP JCXG DGGP UVWFKGF +V JCU DGGP FGOQPUVTCVGF VJCV //+'DCUGF U[U VGOU ECP NGCF VQ UK\CDNG YQTF GTTQT TCVG TGFWEVKQPU QP VJG VTCPUETKRVKQP QH EQPXGTUC YKVJ VTCPUETKRVKQP VKQPCN VGNGRJQPG URGGEJ =? (QT C IKXGP VTCKPKPI UGSWGPEG
C VTCKPKPI EQTRWU KU EQORQUGF QH OCP[ QH UWEJ VTCKPKPI UGSWGPEGU VJG //+' ETKVGTKQP YKVJ ſZGF NCPIWCIG OQFGN EQPUKUVU QH OCZKOK\KPI VJG RQUVGTKQT RTQDCDKNKV[ QH VJG YQTF UGSWGPEG KG
Ï
¼
¼
¼
YJGTG VJG UWOOCVKQP KP VJG FGPQOKPCVQT KU VCMGP QXGT CNN RQUUKDNG YQTF UGSWGPEGU (QT .8%54 VJG ECNEWNCVKQP QH VJG FGPQOKPCVQT VGTOU KU EQORWVCVKQPCNN[ GZRGPUKXG UQ KV KU WUWCNN[ CRRTQZKOCVGF D[ EQPUKFGTKPI QPN[ VJG OQUV NKMGN[ YQTF J[RQVJGUGU IKXGP KP VJG HQTO QH C YQTF NCVVKEG EH 5GEVKQP CPF (KIWTG (QT OQTG FGVCKNU CDQWV FKUETKOKPCVKXG VTCKPKPI VJG TGCFGT KU TGHGTTGF VQ =?
5.5.4 HMM Adaptation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
E\&5&3UHVV//&
VJG DCEMITQWPF PQKUG EJCTCEVGTKUVKEU QT VJG URGCMGT KU CXCKNCDNG CFCRVCVKQP JCU VQ DG RGTHQTOGF WUKPI QPN[ VJG VGUV FCVC KP CP WPUWRGTXKUGF OCPPGT (QWT EQOOQPN[ WUGF UEJGOGU VQ CFCRV VJG RCTCOGVGTU QH C URGGEJ *// ECP DG FKUVKPIWKUJGF $C[GUKCP CFCRVCVKQP =? CFCRVCVKQP DCUGF QP NKPGCT VTCPUHQTOC VKQPU =? FCVC ENWUVGTKPI DCUGF CFCRVCVKQP = ? CPF OQFGN EQORQUKVKQP VGEJ PKSWGU =? $C[GUKCP GUVKOCVKQP CNUQ ECNNGF /#2 GUVKOCVKQP ECP DG UGGP CU C YC[ VQ KPEQTRQTCVG RTKQT MPQYNGFIG KPVQ VJG VTCKPKPI RTQEGFWTG D[ CFFKPI RTQDCDKNKUVKE EQPUVTCKPVU QP VJG OQFGN RCTCOGVGTU 6JG FKHHGTGPEG DGVYGGP /#2 VTCKPKPI CPF UVCPFCTF /. VTCKPKPI NKGU KP VJG CUUWORVKQP QH CP CRRTQRTKCVG RTKQT FKUVTKDWVKQP QH VJG RCTCOGVGTU VQ DG GUVKOCVGF +H KU VJG RCTCOGVGT XGEVQT QH VJG *// VQ DG VTCKPGF QP UQOG FCVC KU YKVJ C VTCPUETKRVKQP CPF KH KU VJG RTKQT 2&( QH VJGP VJG /#2 GUVKOCVG FGſPGF CU VJG OQFG QH VJG RQUVGTKQT 2&( QH KG
6JG *// RCTCOGVGTU CTG UVKNN GUVKOCVGF YKVJ VJG '/ CNIQTKVJO DWV WUKPI VJG /#2 TGGUVKOCVKQP HQTOWNCU =? 6JKU NGCFU VQ VJG /#2 CFCRVCVKQP VGEJPKSWG YJGTG EQP UVTCKPVU QP VJG *// RCTCOGVGTU CTG GUVKOCVGF DCUGF QP VJG RCTCOGVGTU QH CP GZKUVKPI OQFGN 5RGCMGTKPFGRGPFGPV CEQWUVKE OQFGNU ECP UGTXG CU UGGF OQFGNU HQT IGPFGT QT URGCMGT CFCRVCVKQP WUKPI VJG IGPFGTURGCMGT URGEKſE FCVC KG KP VJG CDQXG YJGTG KU VJG RCTCOGVGT XGEVQT QH VJG UGGF OQFGNU GSWCVKQP KU TGRNCEGF D[
/#2 CFCRVCVKQP ECP DG WUGF VQ CFCRV VJG OQFGNU VQ CP[ FGUKTGF EQPFKVKQP HQT YJKEJ UWHſEKGPV NCDGNGF VTCKPKPI FCVC CTG CXCKNCDNG /#2 GUVKOCVKQP JCU VJG UCOG CU[OR VQVKE RTQRGTVKGU CU /. GUVKOCVKQP DWV YJGP KPFGRGPFGPV RTKQTU CTG WUGF HQT FKHHGTGPV RJQPG OQFGNU VJG CFCRVCVKQP TCVG OC[ DG XGT[ UNQY RCTVKEWNCTN[ HQT NCTIG OQFGNU +V KU VJGTGHQTG CFXCPVCIGQWU VQ TGRTGUGPV EQTTGNCVKQPU DGVYGGP OQFGN RCTCOGVGTU KP VJG HQTO QH LQKPV RTKQT FKUVTKDWVKQPU = ? .KPGCT VTCPUHQTOU CTG RQYGTHWN VQQNU HQT RGTHQTOKPI WPUWRGTXKUGF URGCMGT CPF GP XKTQPOGPVCN CFCRVCVKQP 6JG /. NKPGCT TGITGUUKQP /..4 VGEJPKSWG = ? KU RCTVKEWNCTN[ YGNNUWKVGF VQ WPUWRGTXKUGF CFCRVCVKQP 5KPEG VJG PWODGT QH VTCPUHQTOC VKQP RCTCOGVGTU KU UOCNN KV KU RQUUKDNG VQ CFCRV NCTIG OQFGNU YKVJ UOCNN COQWPVU QH FCVC +V EQPUKUVU QH ſPFKPI VJG VTCPUHQTOCVKQP WUWCNN[ CP CHſPG VTCPUHQTOCVKQP QH VJG *// )CWUUKCP OGCPU YJKEJ OCZKOK\GU VJG NKMGNKJQQF QH VJG CFCRVCVKQP FCVC HQT C IKXGP J[RQVJGUK\GF VTCPUETKRVKQP KG
Ê
6JG VTCPUHQTO RCTCOGVGTU A CPF b CTG UJCTGF D[ VJG FKHHGTGPV RJQPG WPKVU CPF CTG VJGTGHQTG TQDWUV VQ TGEQIPKVKQP GTTQTU 6Q QDVCKP VJG /. CU[ORVQVKE RTQRGTVKGU KV KU PGEGUUCT[ VQ WUG OWNVKRNG NKPGCT VTCPUHQTOU CPF VQ CFLWUV VJG PWODGT QH NKPGCT VTCPU HQTOCVKQPU VQ VJG COQWPV QH CXCKNCDNG CFCRVCVKQP FCVC 6JKU ECP DG FQPG GHſEKGPVN[ D[ CTTCPIKPI VJG OKZVWTG EQORQPGPVU KPVQ C VTGG CPF F[PCOKECNN[ FGſPKPI VJG TGITGU UKQP ENCUUGU +P CFFKVKQP VQ VJG )CWUUKCP OGCPU /..4 CFCRVCVKQP KU QHVGP CRRNKGF VQ VJG XCTKCPEG RCTCOGVGTU 6JKU CFCRVCVKQP RTQEGFWTG ECP DG CRRNKGF VQ DQVJ VJG
E\&5&3UHVV//&
VGUV FCVC CPF VTCKPKPI FCVC # PCVWTCN GZVGPUKQP QH VJKU CRRTQCEJ URGCMGT CFCRVKXG VTCKPKPI 5#6 KPEQTRQTCVGU UWRGTXKUGF /..4 KP VJG VTCKPKPI RTQEGFWTG CPF LQKPVN[ GUVKOCVGU VJG VTCKPKPI URGCMGT /..4 VTCPUHQTOU CPF VJG *// RCTCOGVGTU =? 6JG TGUWNVKPI 5#6 OQFGNU CTG DGVVGT UWKVGF VQ /..4 URGCMGT CFCRVCVKQP )KXGP VJG UOCNN PWODGT QH RCTCOGVGTU HQT VJG /..4 VTCPUHQTOCVKQP QP VJG QTFGT QH RCTCOGVGTU HQT C UKPING TGITGUUKQP ENCUU YKVJ C DNQEM FKCIQPCN OCVTKZ VJKU CFCRVCVKQP VGEJPKSWG KU UVKNN UWKVCDNG YKVJ CU NKVVNG CU U QH CFCRVCVKQP FCVC KG QPN[ CDQWV HTCOGU +H NGUU FCVC KU CXCKNCDNG QVJGT CFCRVCVKQP VGEJPKSWGU WUKPI C UOCNNGT PWODGT QH CFCRVCVKQP RCTCOGVGTU CTG TGSWKTGF &CVC ENWUVGTKPI DCUGF CFCR VCVKQP OGVJQFU UWEJ CU VJG GKIGPXQKEGU UEJGOG =? CPF VJG ENWUVGT CFCRVKXG VTCKP KPI =? CTG UWEJ VGEJPKSWGU 6JG[ DQVJ WUG C YGKIJVGF UWO QH ECPQPKECN URGCMGT ENWUVGT OQFGNU VQ GUVKOCVG VJG )CWUUKCP OGCP XGEVQTU 6JGUG CFCRVCVKQP UEJGOGU ECP CNUQ DG EQODKPGF YKVJ UVCPFCTF /..4 CPF /#2 CFCRVCVKQP /QFGN EQORQUKVKQP KU OQUVN[ WUGF VQ EQORGPUCVG HQT CFFKVKXG PQKUG D[ GZRNKEKVN[ OQFGNKPI VJG DCEMITQWPF PQKUG WUWCNN[ YKVJ C UKPING )CWUUKCP CPF EQODKPKPI VJKU OQFGN YKVJ VJG ENGCP URGGEJ OQFGN =? (QT RTCEVKECN TGCUQPU KV KU IGPGTCNN[ CU UWOGF VJCV VJG PQKUG FGPUKV[ KU )CWUUKCP CPF VJCV VJG PQKUG EQTTWRVGF URGGEJ OQFGN JCU VJG UCOG UVTWEVWTG CPF PWODGT QH RCTCOGVGTU CU VJG ENGCP URGGEJ OQFGN Ō V[R KECNN[ C EQPVKPWQWU FGPUKV[ *// YKVJ )CWUUKCP OKZVWTG 8CTKQWU VGEJPKSWGU JCXG DGGP RTQRQUGF VQ GUVKOCVG VJG PQKU[ URGGEJ OQFGNU KPENWFKPI VJG NQIPQTOCN CRRTQZ KOCVKQP CRRTQCEJ C PWOGTKECN KPVGITCVKQP CRRTQCEJ CPF C FCVC FTKXGP CRRTQCEJ =? /QFGN EQORQUKVKQP JCU VJG CFXCPVCIG QH FKTGEVN[ OQFGNKPI VJG PQKU[ EJCPPGN CU QR RQUGF VQ CRRN[KPI DNKPF CFCRVCVKQP VGEJPKSWGU VQ VJG UCOG RTQDNGO
5.6 Decoding 6JG .8%54 FGEQFKPI RTQDNGO KU VJG FGUKIP QH CP GHſEKGPV UGCTEJ CNIQTKVJO VQ FGCN YKVJ VJG JWIG UGCTEJ URCEG QDVCKPGF D[ EQODKPKPI VJG CEQWUVKE CPF NCPIWCIG OQF GNU 5VTKEVN[ URGCMKPI VJG CKO QH VJG FGEQFGT KU VQ FGVGTOKPG VJG OQUV NKMGN[ YQTF UGSWGPEG IKXGP VJG NCPIWCIG OQFGN VJG RTQPWPEKCVKQP FKEVKQPCT[ CPF VJG CEQWU VKE OQFGNU KG
Ï Ï
YJGTG VJG UWOOCVKQP KU VCMGP QXGT CNN RQUUKDNG RTQPWPEKCVKQPU CPF CNN RQUUKDNG *// UVCVG UGSWGPEGU EQTTGURQPFKPI VQ VJG YQTF UGSWGPEG +P RTCEVKEG JQYGXGT KV KU EQOOQP VQ UGCTEJ HQT VJG OQUV NKMGN[ *// UVCVG UGSWGPEG 6JKU OCZKOWO CRRTQZ KOCVKQP CNUQ TGHGTTGF VQ CU 8KVGTDK UGCTEJ NGCFU VQ C UKORNKſGF XKGY QH VJG FGEQFKPI RTQDNGO
E\&5&3UHVV//&
6JKU KU CP GCUKGT VCUM EQPUKUVKPI QH ſPFKPI VJG DGUV RCVJ VJTQWIJ C VTGNNKU VJG UGCTEJ URCEG YJGTG GCEJ PQFG TGRTGUGPVU CP *// UVCVG CV C IKXGP VKOG +V JCU DGGP UJQYP VJCV GXGP VJQWIJ VJG 8KVGTDK FGEQFKPI IKXGU QPN[ C ETWFG CRRTQZKOCVKQP QH VJG NKMG NKJQQF QH VJG YQTF UGSWGPEG VJG VYQ YQTF J[RQVJGUGU CTG CNOQUV CNYC[U XGT[ ENQUG 5QOG UKORNG GZVGPUKQPU QH VJG 8KVGTDK UGCTEJ CTG CDNG VQ EQORGPUCVG HQT OQUV QH FGEQFKPI CRRTQZKOCVKQPU KP RCTVKEWNCT VQ CXQKF RGPCNK\KPI YQTFU YKVJ OCP[ RTQPWP EKCVKQPU +P OCP[ URGGEJ TGEQIPKVKQP U[UVGOU VJG ſTUV UVGR QH FGEQFKPI KU KFGPVKH[KPI VJG URGGEJ RQTVKQPU QH VJG CWFKQ UKIPCN 6JKU RTQEGUU KU FGUETKDGF KP VJG PGZV UWDUGE VKQP HQNNQYGF D[ OQTG FGVCKNU QP FGEQFKPI UVTCVGIKGU
5.6.1 Speech/Non-speech Detection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ſECPVN[ KORTQXGF (KPCNN[ GNKOKPCVKPI PQPURGGEJ UGIOGPVU CPF FKXKFKPI VJG FCVC KPVQ UJQTVGT UGIOGPVU YJKEJ ECP UVKNN DG UGXGTCN OKPWVGU NQPI UWDUVCPVKCNN[ TGFWEGU VJG EQORWVCVKQP VKOG CPF UKORNKſGU FG EQFKPI 8CTKQWU CRRTQCEJGU JCXG DGGP RTQRQUGF VQ RCTVKVKQP C EQPVKPWQWU UVTGCO QH CWFKQ FCVC /QUV QH VJGUG CRRTQCEJGU TGN[ QP C VYQ UVGR RTQEGFWTG YJGTG VJG CWFKQ UVTGCO KU ſTUV UGIOGPVGF KP QTFGT VQ NQECVG CEQWUVKE EJCPIGU YJKEJ CTG CUUWOGF VQ DG CUUQ EKCVGF YKVJ EJCPIGU KP URGCMGT DCEMITQWPF QT GPXKTQPOGPVCN EQPFKVKQP CPF EJCPPGN EQPFKVKQP 6JG UGIOGPVCVKQP RTQEGFWTGU ECP DG ENCUUKſGF CU DGKPI DCUGF QP RJQPG FGEQFKPI = ? FKUVCPEGDCUGF UGIOGPVCVKQPU = ? QT QP J[RQVJGUKU VGUVKPI = ? 6JG TGUWNVKPI UGIOGPVU CTG VJGP ENWUVGTGF WUWCNN[ WUKPI )CWU UKCP OQFGNU YJGTG GCEJ ENWUVGT KU CUUWOGF VQ KFGPVKH[ C URGCMGT QT OQTG RTGEKUGN[ C URGCMGT KP C IKXGP CEQWUVKE EQPFKVKQP #P CNVGTPCVKXG NCPIWCIGKPFGRGPFGPV CR RTQCEJ TGNKGU QP CP CWFKQ UVTGCO OKZVWTG OQFGN =? 'CEJ EQORQPGPV CWFKQ UQWTEG
E\&5&3UHVV//&
TGRTGUGPVKPI C URGCMGT KP C RCTVKEWNCT DCEMITQWPF CPF EJCPPGN EQPFKVKQP KU KP VWTP OQFGNGF D[ C OKZVWTG QH )CWUUKCPU 6JG UGIOGPV DQWPFCTKGU CPF NCDGNU CTG LQKPVN[ KFGPVKſGF XKC CP KVGTCVKXG OCZKOWO NKMGNKJQQF UGIOGPVCVKQPENWUVGTKPI RTQEGFWTG WUKPI )CWUUKCP OKZVWTG OQFGNU CPF CIINQOGTCVKXG ENWUVGTKPI
5.6.2 Decoding Strategies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ſEKGPV RTWPKPI VGEJPKSWGU KP QTFGT VQ QDVCKP C UKPING RCUU FGEQFGT VJCV ECP RTQXKFG VJG CPUYGT WUKPI CNN VJG CXCKNCDNG KP HQTOCVKQP KG VJCV KP VJG OQFGNU KP C UKPING HQTYCTF FGEQFKPI RCUU QXGT QH VJG URGGEJ UKIPCN 6JKU MKPF QH FGEQFGT UWEJ CU VJG UVCEM FGEQFGT =? DCUGF QP VJG # CNIQTKVJO QT VJG QPGRCUU HTCOG U[PEJTQPQWU F[PCOKE PGVYQTM FGEQFGT =? KU XGT[ CVVTCEVKXG HQT TGCNVKOG CRRNKECVKQPU 5VCVKE FGEQFGTU TGSWKTG OWEJ OQTG OGOQT[ VJCP F[PCOKE FGEQFGTU YJGP WUGF YKVJ NQPI URCP NCPIWCIG OQFGNU ITCO QT JKIJGT QTFGT CPF CU C EQPUGSWGPEG VJG[ CTG OQUVN[ WUGF YKVJ UOCNNGT NCPIWCIG OQFGNU WUWCNN[ ITCOU QT EQPUVTCKPGF ITCO OCTU +V JCU DGGP TGEGPVN[ UJQYP VJCV D[ RTQRGT QRVKOK\CVKQP QH C ſPKVGUVCVG CW VQOCVQPÝ EQTTGURQPFKPI VQ C TGEQIPK\GT *// PGVYQTM UWDUVCPVKCN TGFWEVKQP QH VJG QXGTCNN PGVYQTM UK\G ECP DG QDVCKPGF GPCDNKPI UVCVKE FGEQFKPI YKVJ NQPI URCP ./U =? *QYGXGT VJG UK\G QH VJG QRVKOK\GF PGVYQTM TGOCKPU RTQRQTVKQPCN VQ VJG ./ UK\G /WNVKRCUU FGEQFKPI ECP DG WUGF VQ RTQITGUUKXGN[ CFF MPQYNGFIG UQWTEGU KP VJG FG EQFKPI RTQEGUU VJWU CNNQYKPI VJG EQORNGZKV[ QH VJG KPFKXKFWCN FGEQFKPI RCUUGU VQ DG TGFWEGF CPF QHVGP TGUWNVKPI KP C HCUVGT QXGTCNN FGEQFGT =? (QT GZCORNG C ſTUV FG EQFKPI RCUU ECP WUG C ITCO NCPIWCIG OQFGN CPF UKORNG CEQWUVKE OQFGNU CPF NCVGT RCUUGU YKNN OCMG WUG QH ITCO CPF ITCO NCPIWCIG OQFGNU YKVJ OQTG EQORNGZ CEQWUVKE OQFGNU 6JKU OWNVKRNG RCUU RCTCFKIO TGSWKTGU C RTQRGT KPVGTHCEG DGVYGGP RCUUGU KP QTFGT VQ CXQKF NQUKPI KPHQTOCVKQP CPF GPIGPFGTKPI UGCTEJ GTTQTU +PHQT OCVKQP KU WUWCNN[ VTCPUOKVVGF XKC YQTF NCVVKEGU Þ QT YQTF ITCRJU UGG (KIWTG CNVJQWIJ UQOG U[UVGOU WUG 0DGUV J[RQVJGUGU YJKEJ CTG C NKUV QH VJG OQUV NKMGN[ Ý #P *//DCUGF URGGEJ TGEQIPK\GT ECP DG UGGP CU C VTCPUFWEVKQP ECUECFG YJKEJ EQPXGTVU VJG QDUGTXGF HGCVWTG XGEVQTU VQ C YQTF UVTKPI YJGTG VQ UQOG CRRTQZKOCVKQP GCEJ VTCPUFWEVKQP RJQPG OQFGN YQTF OQFGN QT NCPIWCIG OQFGN ECP DG TGRTGUGPVGF CU C ſPKVGUVCVG CWVQOCVQP Þ .CVVKEGU CTG ITCRJU YJGTG PQFGU EQTTGURQPF VQ RCTVKEWNCT HTCOGU CPF YJGTG GFIGU TGRTGUGPVKPI YQTF J[RQVJGUKU JCXG CUUQEKCVGF CEQWUVKE CPF NCPIWCIG OQFGN UEQTGU
E\&5&3UHVV//&
#
9#5
)11&
9#5
*'
UKN 9#5
+6
)11&
UKN
9#5
+6
9#5
9#5 9#50ŏ6 9#50ŏ6
#
UKN
)11&
)11&
241)4#/
)11&
UKN 241)4#/
)11& )11& )11&
FIGURE 5.10 Example word lattice generated by a speech recognizer using a bigram language model for a 2.1s utterance. Each graph edge corresponds to a word hypothesis and a time interval (as specified by the time information on the nodes). In this example the word transcription with the highest likelihood is “sil IT WAS A GOOD PROGRAM sil” which happens to be what was said. (The acoustic and language model likelihoods are not given on the figure.)
YQTF UGSWGPEGU YKVJ VJGKT TGURGEVKXG UEQTGU #V VJG RTKEG QH UQOG CEEGRVCDNG CR RTQZKOCVKQPU YQTF NCVVKEGU CPF 0DGUV NKUVU ECP DG IGPGTCVGF YKVJ NKVVNG QXGTJGCF
CDQWV D[ OQFKH[KPI VJG DQQMMGGRKPI QH VJG RCTVKCN J[RQVJGUGU EQPUKFGTGF FWT KPI TGIWNCT FGEQFKPI =? +V ECP UQOGVKOGU DG FKHſEWNV VQ CFF EGTVCKP MPQYNGFIG UQWTEGU KPVQ VJG FGEQFKPI RTQEGUU GURGEKCNN[ YJGP VJG[ FQ PQV ſV KP VJG /CTMQXKCP HTCOGYQTM 6JKU KU VJG ECUG YJGP VT[KPI VQ WUG UGIOGPVCN KPHQTOCVKQP QT VQ WUG ITCOOCVKECN KPHQTOCVKQP HQT NQPI VGTO CITGGOGPV 5WEJ KPHQTOCVKQP ECP DG OQTG GCUKN[ KPVGITCVGF KP C OWNVK RCUU U[UVGO D[ TGUEQTKPI VJG TGEQIPK\GT J[RQVJGUGU CHVGT CRRN[KPI VJG CFFKVKQPCN MPQYNGFIG UQWTEGU 'XKFGPVN[ VJG ſ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
5.6.3 Efficiency #U FKUEWUUGF CDQXG VJGTG CTG OCP[ GHſEKGPV UQNWVKQPU VQ VJG UGCTEJ RTQDNGO JQY GXGT ſPFKPI VJG QRVKOCN UQNWVKQP KU CNYC[U C VTCFGQHH DGVYGGP VJG OQFGN CEEWTCE[ CPF GHſEKGPV RTWPKPI +P IGPGTCN DGVVGT OQFGNU JCXG OQTG RCTCOGVGTU CPF VJGTGHQTG
E\&5&3UHVV//&
TGSWKTG OQTG EQORWVCVKQP *QYGXGT UKPEG VJG OQFGNU CTG OQTG CEEWTCVG KV KU QHVGP RQUUKDNG VQ WUG C VKIJVGT RTWPKPI NGXGN VJWU TGFWEKPI VJG EQORWVCVKQPCN NQCF YKVJQWV CP[ NQUU KP CEEWTCE[ .KOKVCVKQPU QP VJG CXCKNCDNG EQORWVCVKQPCN TGUQWTEGU ECP UKIPKſ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ſEKGPV RTWPKPI UVTCVGIKGU CUUQEKCVGF YKVJ C NCPIWCIG OQFGN NQQMCJGCF = ? /WNVKRCUU CRRTQCEJGU ECP CNUQ DG WUGF UWEEGUUHWNN[ HQT ENQUG VQ TGCNVKOG QRGTCVKQP D[ EJWPMKPI VJG FCVC CPF TWPPKPI VJG FKHHGTGPV RCUUGU KP RCTCNNGN YKVJ C UNKIJV FGNC[ (QT URGCMGTKPFGRGPFGPV .8%54 DCUGF QP )CWUUKCP OKZVWTG *// DGVYGGP CPF QH VJG TGEQIPKVKQP VKOG KU URGPV KP EQORWVKPI VJG *// UVCVG NKMGNKJQQFU YKVJ VJG TGOCKPKPI VKOG EQTTGURQPFKPI VQ VJG UGCTEJ RTQEGFWTG KVUGNH 6JKU KU FWG VQ VJG NCTIG PWODGT QH UVCVGU PGGFGF VQ TGRTGUGPV VJG EQPVGZVFGRGPFGPV RJQPG OQFGNU GXGP YJGP UVCVG V[KPI KU WUGF 6JKU EQORWVCVKQP ECP DG TGFWEGF GKVJGT D[ KORNGOGPVKPI C HCUV UVCVG NKMGNKJQQF EQORWVCVKQP YJKEJ WUWCNN[ TGSWKTGU OCMKPI UQOG CRRTQZKOC VKQPU QT D[ TGFWEKPI VJG OQFGN UK\G YJKEJ JCU VJG CFFKVKQPCN CFXCPVCIG QH TGFWEKPI VJG OGOQT[ TGSWKTGOGPVU # YKFGN[ WUGF VGEJPKSWG HQT URGGFKPI WR VJG UVCVG NKMG NKJQQF EQORWVCVKQP KU XGEVQT SWCPVK\CVKQP QH VJG HGCVWTG XGEVQT URCEG KP QTFGT VQ RTGRCTG C )CWUUKCP UJQTV NKUV HQT GCEJ *// UVCVG CPF GCEJ TGIKQP QH VJG SWCPVKſGF HGCVWTG URCEG =? 9KVJ VJKU VGEJPKSWG VJG PWODGT QH )CWUUKCP NKMGNKJQQFU VQ DG EQORWVGF FWTKPI FGEQFKPI HQT GCEJ KPRWV HTCOG CPF GCEJ UVCVG ECP DG TGFWEGF VQ C HTCEVKQP QH VJG PWODGT QH )CWUUKCPU EQTTGURQPFKPI VQ VJG CEVKXG UVCVGU YKVJ QPN[ C UOCNN NQUU KP CEEWTCE[ /QFGN CPF UVCVG V[KPI CTG EQOOQPN[ WUGF VQ KORTQXG VJG OQFGN CEEWTCE[ DWV QRVKOCN V[KPI HTQO VJG CEEWTCE[ RQKPV QH XKGY ECP UVKNN TGUWNV KP C XGT[ NCTIG OQFGN YKVJ M VQ M UVCVGU YJGP NCTIG COQWPVU QH VTCKPKPI FCVC CTG CXCKNCDNG 2CTCOGVGT V[KPI KU CNUQ RQYGTHWN VGEJPKSWG VQ TGFWEG VJG PWODGT QH RCTCOGVGTU CPF ECP DG CRRNKGF VQ CNN VJG NGXGNU QH VJG OQFGN UVTWEVWTG CNNQRJQPG OQFGN UVCVG CPF )CWUUKCP =? *QYGXGT OQTG ƀGZKDKNKV[ KU CXCKNCDNG HQT )CWUUKCP 2&( V[KPI KP VJCV NCTIG OQFGN TGFWEVKQPU ECP DG QDVCKPGF YKVJQWV UCETKſEKPI VQQ OWEJ KP VGTOU QH U[UVGO CEEWTCE[ 6JKU KU GZGORNKſGF D[ VJG UWDURCEG FKUVTKDWVKQP V[KPI CRRTQCEJ = ? YJKEJ KP KVU OQUV GNGOGPVCT[ KORNGOGPVCVKQP ECP DG UGGP CU C SWCPVK\CVKQP QH VJG OQFGN RCTCOGVGTU 6JG NCPIWCIG OQFGN WUWCNN[ C ITCO QT ITCO DCEMQHH ./ KP UVCVGQHVJGCTV U[UVGOU ECP JCXG C XGT[ NCTIG PWODGT QH RCTCOGVGTU QXGT OKNNKQP CPF VJGTG HQTG OC[ TGSWKTG RTQJKDKVKXG COQWPVU QH OGOQT[ 1PG QH VJG CVVTCEVKXG RTQRGTVKGU QH nITCO OQFGNU KU VJG RQUUKDKNKV[ QH TGN[KPI OQTG QP VJG DCEMQHH EQORQPGPVU D[
E\&5&3UHVV//&
KPETGCUKPI VJG EWVQHHU QP VJG ITCO EQWPVU VJWU TGFWEKPI UKIPKſECPVN[ VJG ./ UK\G
EH 5GEVKQP /QTG GNCDQTCVG ITCO RTWPKPI VGEJPKSWGU JCXG CNUQ DGGP RTQ RQUGF = ? VQ UWDUVCPVKCNN[ TGFWEG VJG ./ UK\G YKVJ PGINKIKDNG NQUU KP CEEWTCE[ #P CNVGTPCVKXG CRRTQCEJ VQ NKOKV VJG OGOQT[ TGSWKTGOGPVU KU VQ MGGR OQUV QH VJG ./ RCTCOGVGTU QP VJG FKUM UKPEG OQUV ITCOU CTG PGXGT WUGF EQODKPGF YKVJ C ECEJG QH VJG UEQTGU HQT CEEGUUGF ./ UVCVGU =?
5.6.4 Confidence Measures %QPſ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ſFGPEG OGCUWTG KG VJG IQCN KU VQ QDVCKP CP GUVKOCVG QH VJG RQUVGTKQT RTQDCDKNKV[ QH VJG VJ YQTF #P GUVKOCVG QH VJKU KP VJG J[RQVJGUK\GF YQTF UVTKPI QT CNVGTPCVKXGN[ NCVVGT RTQDCDKNKV[ ECP DG GHſEKGPVN[ EQORWVGF D[ CRRN[KPI VJG (QTYCTF$CEMYCTF CN IQTKVJO VQ C YQTF ITCRJ IGPGTCVGF D[ VJG URGGEJ TGEQIPK\GT =? *QYGXGT UKPEG VJKU RQUVGTKQT RTQDCDKNKV[ TGNKGU QP KPEQTTGEV OQFGNU KV KU CNUQ EQOOQP VQ WUG CFFK VKQPCN HGCVWTGU UWEJ CU YQTF CPF RJQPG FWTCVKQPU URGCMKPI TCVG CPF UKIPCNVQPQKUG TCVKQ VQ DGVVGT CRRTQZKOCVG VJG YQTF RQUVGTKQT RTQDCDKNKV[ #NN VJG RTGFKE VQTU ECP DG EQODKPGF CPF OCRRGF VQ VJG EQPſFGPEG UEQTG D[ WUKPI GKVJGT C NQIKU VKE TGITGUUKQP =? C IGPGTCNK\GF CFFKVKXG OQFGN =? QT C PGWTCNPGVYQTM =? 6JGUG OQFGNU CTG VTCKPGF QP FGXGNQROGPV FCVC D[ OCZKOK\KPI C EQPſFGPEG UEQTG OGVTKE UWEJ VJG PQTOCNK\GF ETQUU GPVTQR[ 6JG RTQRGT UGV QH HGCVWTGU FGRGPFU QP VJG RCTVKEWNCT CRRNKECVKQP
5.7 Indicative Performance Levels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
E\&5&3UHVV//&
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őYQTF GTTQTŒ TCVG KU C OGCUWTG QH VJG CXGTCIG PWODGT QH GTTQTU VCMKPI KPVQ CEEQWPV VJTGG GTTQT V[RGU YKVJ TGURGEV VQ C TGHGTGPEG VTCPUETKRVKQP substitutions VJG TGHGTGPEG YQTF KU TGRNCEGF D[ CPQVJGT YQTF insertions C YQTF KU J[RQVJGUK\GF VJCV YCU PQV KP VJG TGHGTGPEG CPF deletions C YQTF KP VJG TGHGTGPEG VTCPUETKRVKQP KU OKUUGF 6JG YQTF GTTQT TCVG KU FGſPGF CU UWDU KPU FGN TGHGTGPEG YQTFU CPF KU IGPGTCNN[ EQORWVGF D[ CNKIPKPI VJG TGHGTGPEG CPF J[RQVJGUK\GF VTCPUETKRVKQPU WUKPI C F[PCOKE RTQITCOOKPI CNIQTKVJO YJGTG EQUVU CTG CUUQEKCVGF YKVJ VJG FKHHGT GPV GTTQT V[RGU )KXGP VJKU FGſPKVKQP VJG YQTF GTTQT ECP DG OQTG VJCP 9JKNG VJKU EJCRVGT CFFTGUUGU URGGEJ VTCPUETKRVKQP KG IQKPI HTQO VJG CWFKQ UKIPCN VQ YQTFU KV UJQWNF DG MGRV KP OKPF VJCV CFFKVKQPCN KPHQTOCVKQP ECP DG GZVTCEVGF HTQO VJG CWFKQ UKIPCN 'ZVTCEVKQP QH UQOG QH VJKU UQECNNGF őOGVCFCVCŒ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ſPGF UGOCPVKE UNQVU CPF XCNWGU
5.7.1 Dictation &KEVCVKQP KU VJG OQUV QDXKQWU CWVQOCVKE URGGEJ TGEQIPKVKQP VCUM CPF JCU C NQPI JKUVQT[ QH TGUGCTEJ CPF RTQFWEV FGXGNQROGPV TGUWNVKPI KP NQYEQUV QHHVJGUJGNH U[U VGOU HQT C XCTKGV[ QH RNCVHQTOU CPF NCPIWCIGU 9JKNG HTQO VJG VGEJPQNQIKECN XKGY RQKPV FKEVCVKQP KU WUWCNN[ VJQWIJV QH CU C őUKORNGŒ VTCPUHQTOCVKQP HTQO URGGEJ VQ VGZV VJKU XKGY QXGTNQQMU C XCTKGV[ QH HQTOCVVKPI CPF KPVGITCVKQP KUUWGU YJKEJ CTG KORQTVCPV HQT WUCDKNKV[ 2GTJCRU VJG OQUV PQVCDNG EJCTCEVGTKUVKE QH VJG FKEVCVKQP VCUM KU VJCV VJG URGGEJ FCVC KU RTQFWEGF YKVJ VJG GZRNKEKV IQCN QH DGKPI VTCPUETKDGF D[ C OCEJKPG 6JG URGGEJ FCVC KP C FKEVCVKQP UGUUKQP EQOGU HTQO C UKPING URGCMGT CPF KU TGEQTFGF YKVJ C EQPVTQNNGF UKIPCN CESWKUKVKQP UGVWR 6JG NKPIWKUVKE EQPVGPV KU WUWCNN[ UQOGYJCV NKOKVGF CPF VJG YQTF UVTGCO KU SWKVG ENQUG VQ VJG YTKVVGP HQTO #NVJQWIJ DGPEJOCTMU QH EQOOGTEKCN FKEVCVKQP U[UVGOU CTG PQV RWDNKEN[ CXCKNCDNG FKEVCVKQP JCU UGTXGF CU C DCUGNKPG RGTHQTOCPEG OGCUWTG KP .8%54 OQUV PQVCDN[ KP VJG DGPEJOCTM VGUVU URQPUQTGF D[ VJG 75 *# RTQITCOU CPF EQQTFKPCVGF D[
E\&5&3UHVV//&
0+56 6JG ENQUG TGNCVKQPUJKR DGVYGGP U[UVGO FGXGNQROGPV CPF GXCNWCVKQP TGHGTTGF VQ CU őCUUGUUOGPV FTKXGP VGEJPQNQI[ FGXGNQROGPVŒ JCU NGF VQ NCTIG RGTHQTOCPEG KORTQXGOGPVU KP URKVG QH KPETGCUKPI VCUM FKHſEWNV[ (QT TGCF URGGEJ VJG UVCVGQHVJG CTV KP URGCMGTKPFGRGPFGPV EQPVKPWQWU URGGEJ TGEQIPKVKQP KU GZGORNKſGF D[ VJG NCUV DGPEJOCTM VGUVU QP 0QTVJ #OGTKECP $WUKPGUU 0GYU VCUM = ? 6JG CEQWUVKE VTCKPKPI FCVC YCU EQORTKUGF QH CDQWV J QH TGCF PGYURCRGT VGZVU HTQO UGXGTCN JWPFTGF URGCMGTU CPF VJG NCPIWCIG OQFGN VTCKPKPI OCVGTKCN YCU EQO RTKUGF QH / YQTFU QH PGYURCRGT VGZVU HTQO C XCTKGV[ QH UQWTEGU 1P VGUV FCVC TGEQTFGF YKVJ C ENQUGVCNMKPI OKETQRJQPG YKVJ CP 504 QH CDQWV F$ YQTF GTTQT TCVGU CTQWPF YGTG QDVCKPGF WUKPI C M YQTF XQECDWNCT[ Ü 6JG UCOG TGCF URGGEJ TGEQTFGF YKVJ C VCDNGVQR OKETQRJQPG KP C EQORWVGT TQQOQHſEG GPXKTQPOGPV PQKUG NGXGN F$# 504 CDQWV F$ TGUWNVGF KP C YQTF GTTQT QH CDQWV YKVJ PQKUG EQORGPUCVKQP 9KVJQWV PQKUG EQORGPUCVKQP VJG YQTF GTTQT TCVGU QH U[UVGOU VTCKPGF QP QPN[ ENGCP URGGEJ FCVC YGTG QXGT 6JG YQTF GTTQT HQT TGCF PGYURCRGT VGZVU TGEQTFGF QXGT NQPI FKUVCPEG VGNGRJQPG NKPGU YCU QXGT 5RQPVCPGQWU FKEVCVKQP QH DWUKPGUU CPF ſPCPEKCN PGYU YCU CFFTGUUGF D[ CUMKPI UWDLGEVU YKVJ GZRGTKGPEG KP LQWTPCNKUO VQ TGCF CDQWV C UWDLGEV CPF VJGP FKEVCVG C VGZV 6JG LQWTPCNKUVU YGTG PQV CNNQYGF VQ TGCF HTQO C FTCHV DWV YGTG CNNQYGF VQ TGLGEV KNNHQTOGF UGPVGPEGU =? 6JG YQTF GTTQT QP VJKU FCVC YCU CDQWV #PQVJGT VCUM CFFTGUUGF URGGEJ TGEQIPK VKQP QH PQPPCVKXG VCNMGTU 9KVJ C UGV QH CFCRVCVKQP UGPVGPEGU URGCMGT CFCRVCVKQP TGFWEGF VJG YQTF GTTQT TCVG D[ C HCEVQT QH VYQ HTQO VQ #NVJQWIJ PQV CP QHſEKCN DGPEJOCTM TGUWNV EQORCTCDNG YQTF GTTQT TGFWEVKQPU JCXG DGGP QDVCKPGF HQT PCVKXG URGCMGTU QP QVJGT VCUMU 9JKNG VJG TGUWNVU IKXGP JGTG CTG HQT #OGTKECP 'PINKUJ UQOGYJCV EQORCTCDNG TG UWNVU JCXG DGGP TGRQTVGF D[ XCTKQWU UKVGU HQT QVJGT NCPIWCIGU =? 6JG .4' 5 3#.'
5RGGEJ TGEQIPK\GT 3WCNKV[ #UUGUUOGPV HQT .KPIWKUVKE 'PIKPGGTKPI RTQLGEV =? YJKEJ CKOGF VQ CUUGUU NCPIWCIGFGRGPFGPV KUUWGU KP OWNVKNKPIWCN TGEQIPK\GT GXCNWC VKQP FGOQPUVTCVGF VJCV VJG UCOG TGEQIPKVKQP VGEJPQNQI[ CPF GXCNWCVKQP OGVJQFQNQI[ WUGF HQT #OGTKECP 'PINKUJ EQWNF DG UWEEGUUHWNN[ CRRNKGF VQ C FKEVCVKQP VCUM KP $TKVKUJ 'PINKUJ (TGPEJ CPF )GTOCP
5.7.2 Speech Recognition for Dialog Systems 6JG URGGEJ TGEQIPK\GT KU QHVGP EQPUKFGTGF C ETKVKECN EQORQPGPV QH URQMGP FKCNQI U[UVGOU YJKEJ CKO VQ GPCDNG XQECN CEEGUU VQ UVQTGF KPHQTOCVKQP +P QTFGT VQ RTQXKFG WUGTHTKGPFN[ KPVGTCEVKQP YKVJ C OCEJKPG KV KU PGEGUUCT[ VQ DG CDNG VQ TGEQIPK\G PCVW TCNN[ URQMGP WVVGTCPEGU HTQO WPMPQYP URGCMGTU +P IGPGTCN GCEJ WUGT KPVGTCEVU QPN[ DTKGƀ[ YKVJ VJG OCEJKPG UQ VJGTG KU XGT[ NKVVNG FCVC CXCKNCDNG HQT OQFGN CFCRVCVKQP 6GNGRJQPG UGTXKEGU CTG C PCVWTCN CTGC HQT URQMGP FKCNQI U[UVGOU CU VJG QPN[ OGCPU QH KPVGTCEVKQP YKVJ VJG OCEJKPG CTG XKC XQKEG CPF JCXG VJWU DGGP VJG HQEWU QH OCP[ FGXGNQROGPV GHHQTVU 5KPEG CNN KPVGTCEVKQP YKVJ VJG ECNNGT KU D[ URGGEJ FKCNQI FGUKIP Ü 9KVJ VJG GZEGRVKQP QH VJG VGNGRJQPG TGEQTFKPIU VJG URGCMGTU YGTG CNNQYGF VQ TGRGCV VJGKT TGEQTFKPI KH WPUCVKUſGF YKVJ KV = ?
E\&5&3UHVV//&
CPF TGURQPUG IGPGTCVKQP CTG QH RCTVKEWNCT KORQTVCPEG KP VJG EQPVGZV QH PCVWTCN OKZGF KPKVKCVKXG FKCNQI )TQYKPI KP RQRWNCTKV[ CTG KPHQTOCVKQP MKQUMU =? CPF OWNVKOGFKC YGD KPVGTHCEGU KP YJKEJ FKHHGTGPV OQFCNKVKGU VCEVKNG CPF CWFKQ ECP DG WUGF HQT KPRWV CPF QWVRWV 6JG URGGEJ TGEQIPK\GTU QH FKCNQI U[UVGOU CTG V[RKECNN[ HCEGF YKVJ OQTG EJCNNGPIKPI CEQWUVKE EQPFKVKQPU VJCP HQT FKEVCVKQP VCUMU DGKPI UWDLGEV VQ EJCPPGN FKU VQTVKQPU XCTKGF JCPFUGVU CPF PQKU[ DCEMITQWPF EQPFKVKQPU 6JG ECRCDKNKV[ QH VJG WUGT VQ KPVGTTWRV VJG OCEJKPG KU QHVGP EQPUKFGTGF ETWEKCN HQT WUCDKNKV[ +P EQPVTCUV VQ FKEVCVKQP CRRNKECVKQPU YJGTG KV KU TGNCVKXGN[ UVTCKIJVHQTYCTF VQ QDVCKP NCTIG YTKVVGP EQTRQTC HQT NCPIWCIG OQFGNKPI HQT FKCNQI U[UVGOU KV KU WUWCNN[ PGE GUUCT[ VQ EQNNGEV CRRNKECVKQPURGEKſE FCVC YJKEJ ECP TGRTGUGPV C UKIPKſECPV RQTVKQP QH VJG FGXGNQROGPV GHHQTV =? #ESWKTKPI UWHſEKGPV COQWPVU QH ./ VTCKPKPI FCVC KU OQTG EJCNNGPIKPI VJCP QDVCKPKPI CEQWUVKE FCVC 9KVJ M SWGTKGU TGNCVKXGN[ TQ DWUV CEQWUVKE OQFGNU ECP DG VTCKPGF DWV VJKU PWODGT QH SWGTKGU YKNN V[RKECNN[ EQPVCKP HGYGT VJCP M YQTFU YJKEJ OC[ DG KPUWHſEKGPV HQT YQTF NKUV FGXGNQROGPV CPF HQT VTCKPKPI n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ſEWNV VQ EQORCTG TGUWNVU CETQUU U[UVGOU CPF VCUMU CU FKHHGTGPV VTCPUETKRVKQP EQPXGPVKQPU CPF VGZV PQT OCNK\CVKQPU CTG QHVGP WUGF +V UJQWNF DG PQVGF VJCV TGRQTVKPI YQTF GTTQT TCVGU ECP DG UQOGYJCV OKUNGCFKPI UKPEG CNN FKHHGTGPEGU DGVYGGP VJG exact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
5.7.3 Transcription for Audio Indexation # OQTG TGEGPV CRRNKECVKQP CTGC KU VJG VTCPUETKRVKQP QH IGPGTCN CWFKQ FCVC UWEJ CU TC FKQ CPF VGNGXKUKQP DTQCFECUVU ß QT OGGVKPIU CPF VGNGEQPHGTGPEGU #WVQOCVKE URGGEJ TGEQIPKVKQP KU C MG[ VGEJPQNQI[ HQT CWFKQ CPF XKFGQ KPFGZKPI CPF CP[ MKPF QH CWFKQ FCVC OKPKPI 5GXGTCN EJCTCEVGTKUVKEU QH VJKU V[RG QH CWFKQ FCVC ECP DG PQVGF (KTUV KV ECP DG EQPUKFGTGF őHQWPFŒ FCVC KP VJCV KV KU RTQFWEGF HQT QVJGT TGCUQPU 6Q DG CDNG ß 6JG GCTNKGUV YQTM KP VJKU CTGC VJCV YG CTG CYCTG QH KU VJG 05( + 0(14/'&+# RTQLGEV =? WPFGT VJG &KIKVCN .KDTCTKGU 0GYUQP&GOCPF CEVKQP NKPG # URGEKCN UGEVKQP QH VJG %QOOWPKECVKQPU QH VJG #%/ YCU TGEGPVN[ FGXQVGF VQ VJKU VQRKE =?
E\&5&3UHVV//&
VQ CWVQOCVKECNN[ UVTWEVWTG VJG FCVC HQT QVJGT WUGU KU QPN[ C UGEQPFCT[ DGPGſV 7UKPI VGUV FCVC VCMGP HTQO C TGCN VCUM CU QRRQUGF VQ FCVC TGEQTFGF HQT GXCNWCVKQP RWTRQUGU TGRTGUGPVU C OCLQT UVGR HQT VJG EQOOWPKV[ 5GEQPFN[ VJG FCVC EQPUKUVU QH C EQP VKPWQWU CWFKQ UVTGCO YJGTG VJGTG CTG OWNVKRNG URGCMGT VWTPU OC[DG QXGTNCRRKPI CPF VJGTG KU PQ C RTKQTK UGIOGPVCVKQP KPVQ UGPVGPEGU 6JKTFN[ VJG UKIPCN ECRVWTG CPF DCEMITQWPF GPXKTQPOGPV ECP DG QPN[ OQTG QT NGUU EQPVTQNNGF
6YQ RTKPEKRNG V[RGU QH RTQDNGOU CTG GPEQWPVGTGF KP CWVQOCVKECNN[ VTCPUETKDKPI CWFKQ FCVC UVTGCOU VJQUG TGNCVKPI VQ VJG XCTKGF CEQWUVKE RTQRGTVKGU QH VJG UKIPCN CPF VJQUG TGNCVGF VQ VJG NKPIWKUVKE RTQRGTVKGU QH VJG URGGEJ 0QKUG TQDWUVPGUU KU CNUQ PGGFGF KP QTFGT VQ CEJKGXG CEEGRVCDNG RGTHQTOCPEG NGXGNU +P QTFGT VQ DG TQDWUV YKVJ TGURGEV VQ VJG XCTKGF CEQWUVKE EQPFKVKQPU VJG CEQWUVKE OQFGNU CTG V[RKECNN[ VTCKPGF QP NCTIG EQTRQTC UGXGTCN VGPU QH JQWTU VQ QXGT C JWPFTGF JQWTU EQPVCKPKPI C XCTKGV[ QH FCVC V[RGU 6JG NKPIWKUVKE OQFGNU CTG UKOKNCTN[ VTCKPGF QP NCTIG VGZV EQTRQTC HTQO XCTKQWU UQWTEGU YKVJ FKHHGTGPV NKPIWKUVKE RTQRGTVKGU UWEJ CU PGYURCRGT CPF PGYUYKTG VGZVU +PVGTPGV FCVC EQOOGTEKCN VTCPUETKRVKQPU CPF FGVCKNGF VTCPUETKRVKQPU QH CEQWUVKE FCVC )KXGP VJG URQPVCPGQWU PCVWTG QH RCTVU QH VJG CWFKQ FCVC KV KU KORQTVCPV VQ GZRNKEKVN[ OQFGN GZVTCNKPIWKUVKE RJGPQOGPC UWEJ CU ſNNGT YQTFU CPF DTGCVJ PQKUG
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ſEWNV[ QH QDVCKPKPI EQOOGTEKCN VTCPUETKRVU HQT NCPIWCIG OQFGN GUVKOCVKQP
5WDUVCPVKCNN[ JKIJGT YQTF GTTQT TCVGU CDQXG JCXG DGGP TGRQTVGF HQT VJG VTCP UETKRVKQP QH VGNGRJQPG EQPXGTUCVKQPCN URGGEJ =? WUKPI VJG 5YKVEJDQCTF =? CPF OWNVKNKPWICN %CNNJQOG 5RCPKUJ #TCDKE /CPFCTKP ,CRCPGUG )GTOCP EQTRQTC 6JG %CNNJQOG FCVC KU RCTVKEWNCTN[ EJCNNGPIKPI VQ VTCPUETKDG CU VJG EQPXGTUCVKQPU CTG DG VYGGP VYQ RGQRNG VJCV MPQY GCEJ QVJGT CPF URGCM KP C HCOKNKCT OCPPGT CDQWV UWDLGEVU QH EQOOQP KPVGTGUV
E\&5&3UHVV//&
5.8 Portability and Language Dependencies 5VCVKUVKECNN[DCUGF URGGEJ TGEQIPKVKQP VGEJPQNQI[ JCU DGGP UWEEGUUHWNN[ GORNQ[GF HQT C XCTKGV[ QH VCUMU CPF NCPIWCIGU 6JG RQTVKPI QH C .8%54 U[UVGO VQ C PGY VCUM QT CPQVJGT NCPIWCIG TGSWKTGU VJG CXCKNCDKNKV[ QH UWHſEKGPV COQWPVU QH VTCPUETKDGF VTCKPKPI FCVC CPF KPXQNXGU UWDUVCPVKCN GHHQTV VQ EQPUVTWEV VJG CEQWUVKE CPF NCPIWCIG OQFGNU CPF VQ FGXGNQR VJG TGEQIPKVKQP NGZKEQP 1HVGP JQYGXGT VJG PGEGUUCT[ TG UQWTEGU CTG PQV CXCKNCDNG CPF IGPGTCVKPI VJGO ECP DG NQPI CPF GZRGPUKXG 4GEGPV GHHQTVU JCXG DGGP FKTGEVGF CV FGXGNQRKPI IGPGTKE TGEQIPKVKQP OQFGNU CPF VJG WUG QH WPCPPQVCVGF FCVC HQT VTCKPKPI RWTRQUGU KP CP CKO VQ TGFWEG VJG TGNKCPEG QP OCPWCNN[ CPPQVCVGF VTCKPKPI EQTRQTC CPF TGFWEKPI FGXGNQROGPV EQUVU =? /GVJQFU VQ KORTQXG IGPGTCNKV[ QH VJG OQFGNU CTG WPFGT KPXGUVKICVKQP DWV VJG RTQDNGO KU HCT HTQO DGKPI UQNXGF #NVJQWIJ 'PINKUJ JCU DGGP VJG RTGFQOKPCPV NCPIWCIG HQT VJG EQORWVGT YQTNF VJGTG JCU DGGP C NCTIG ITQYVJ KP VJG COQWPV QH KPHQTOCVKQP CXCKNCDNG KP GNGEVTQPKE HQTO KP OCP[ QH VJG YQTNFŏU NCPIWCIGU $WKNFKPI C TGEQIPK\GT HQT CPQVJGT NCPIWCIG KU PQV UQ FKHHGTGPV VJCP DWKNFKPI C TGEQIPK\GT HQT C PGY VCUM RCTVKEWNCTN[ HQT ENQUG NCPIWCIGU .CPIWCIGFGRGPFGPV U[UVGO EQORQPGPVU UWEJ CU VJG RJQPG UGV VJG PGGF HQT RTQ PWPEKCVKQP CNVGTPCVKXGU QT RJQPQNQIKECN TWNGU GXKFGPVN[ OWUV DG EJCPIGF 1VJGT NCPIWCIG FGRGPFGPV HCEVQTU CTG TGNCVGF VQ VJG FGſPKVKQP CPF CEQWUVKE EQPHWUKDKNKV[ QH VJG YQTFU KP VJG NCPIWCIG UWEJ CU JQOQRJQPG OQPQRJQPG CPF EQORQWPF YQTF TCVGU CPF VJG YQTF EQXGTCIG QH C IKXGP UK\G TGEQIPKVKQP XQECDWNCT[ 6CMKPI KPVQ CEEQWPV NCPIWCIG URGEKſEKVKGU ECP GXKFGPVN[ KORTQXG TGEQIPKVKQP RGTHQTOCPEG (QT GZCORNG VQPCN NCPIWCIGU UWEJ CU %JKPGUG OC[ DGPGſV HTQO GZRNKEKV OQFGNKPI QH RKVEJ YJKEJ KP VWTP OC[ TGSWKTG OQFKſECVKQPU VQ VJG HGCVWTG CPCN[UKU WUGF 6JGTG CTG VYQ RTGFQOKPCPV CRRTQCEJGU HQT DQQVUVTCRRKPI CEQWUVKE OQFGNU HQT CPQVJGT NCPIWCIG 6JG ſTUV KU VQ WUG CEQWUVKE OQFGNU HTQO CP GZKUVKPI TGEQIPK\GT CPF C RTQPWPEKCVKQP FKEVKQPCT[ VQ UGIOGPV OCPWCNN[ CPPQVCVGF VTCKPKPI FCVC HQT VJG VCTIGV NCPIWCIG +H TGEQIPK\GTU HQT UGXGTCN NCPIWCIGU CTG CXCKNCDNG VJG UGGF OQFGNU ECP DG UGNGEVGF D[ VCMKPI VJG ENQUGUV OQFGN KP QPG QH VJG CXCKNCDNG NCPIWCIGURGEKſ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
E\&5&3UHVV//&
WUGF KP RNCEG QH VJG OCPWCN VTCPUETKRVKQPU TGSWKTGF HQT CNKIPOGPV D[ KPEQTRQTCVKPI VJKU KPHQTOCVKQP KP C NCPIWCIG OQFGN 6JKU NCPIWCIG OQFGN ECP DG WUGF YKVJ CEQWU VKE OQFGNU FGXGNQRGF HQT CPQVJGT VCUM VQ CWVQOCVKECNN[ VTCPUETKDG VJG VCUMURGEKſE VTCKPKPI FCVC #NVJQWIJ KP VJG DGIKPPKPI VJG GTTQT TCVG QP PGY FCVC KU NKMGN[ VQ DG TCVJGT JKIJ VJKU URGGEJ FCVC ECP DG WUGF VQ TGVTCKP VJG OQFGNU QH VJG TGEQIPKVKQP U[UVGO #P KVGTCVKXG RTQEGFWTG ECP UWEEGUUKXGN[ TGſPG VJG OQFGNU CPF VJG VTCPUETKR VKQP = ? 6JKU CRRTQCEJ KU RCTVKEWNCTN[ RTQOKUKPI HQT VJG VTCPUETKRVKQP QH TGCFKN[ CXCKNCDNG CWFKQ UQWTEGU UWEJ CU TCFKQ CPF VGNGXKUKQP PGYU DTQCFECUVU VJCV ECP RTQXKFG CP GUUGPVKCNN[ WPNKOKVGF UWRRN[ QH CEQWUVKE VTCKPKPI FCVC
References =? JVVREQTGVGZKVEKV =? & #DDGTNG[ & -KTD[ 5 4GPCNU CPF 6 4QDKPUQP The THISL Broadcast News Retrieval System, 2TQE '5%# '649 QP #EEGUUKPI +PHQTOCVKQP KP 5RQMGP #WFKQ %CODTKFIG 7- #RTKN =? # #PFTGQWO 6 -COO CPF , %QJGP Experiments in Vocal Tract Normalization, 2TQE %#+2 9QTMUJQR (TQPVKGTU KP 5RGGEJ 4GEQIPKVKQP ++ =? 6 #PCUVCUCMQU , /E&QPQWIJ 4 5EJYCTV\ CPF , /CMJQWN A Compact Model for Speaker Adaptation Training, 2TQE +%5.2ŏ 2JKNCFGN RJKC 2# 1EVQDGT =? : #WDGTV One Pass Cross Word Decoding for Large Vocabularies Based on a Lexical Tree Search Organization, 2TQE '5%# 'WTQURGGEJŏ 4 $WFCRGUV *WPICT[ 5GRVGODGT =? 5 #WUVKP 4 5EJYCTV\ CPF 2 2NCEGYC[ The Forward-Backward Search Strategy for Real-Time Speech Recognition, 2TQE +''' +%#552 6QTQPVQ %CPCFC /C[ =? .4 $CJN ,- $CMGT 25 %QJGP 04 &KZQP ( ,GNKPGM 4. /GTEGT CPF *( 5KNXGTOCP Preliminary results on the performance of a system for the automatic recognition of continuous speech, 2TQE +''' +%#552 2JKNCFGN RJKC 2# #RTKN =? .4 $CJN 2 $TQYP 2 FG 5QW\C 4. /GTEGT CPF / 2KEJGP[ Acoustic Markov Models used in the Tangora Speech Recognition System, 2TQE +''' +%#552 1 0GY ;QTM 0; #RTKN =? .4 $CJN ( ,GNKPGM CPF 4. /GTEGT A Maximum Likelihood Approach to Continuous Speech Recognition, +''' 6TCPU 2CVVGTP #PCN[UKU /CEJKPG +P VGNNKIGPEG PAMI-5 /CTEJ
E\&5&3UHVV//&
=? .4 $CJN 28 FG 5QW\C 25 )QRCNCMTKUJPCP & 0CJCOQQ CPF / 2KEJGP[ A Fast Match for Continuous Speech Recognition Using Allophonic Models, 2TQE +''' +%#552 %# 1 5CP (TCPEKUEQ %# /CTEJ =? , $CMGT , $CMGT 2 $CODGTI - $KUJQR . )KNNKEM 8 *GNOCP < *WCPI ; +VQ 5 .QYG $ 2GUMKP 4 4QVJ CPF ( 5ECVVQPG Large Vocabulary Recognition of Wall Street Journal Sentences at Dragon Systems, 2TQE *# 5RGGEJ 0CVWTCN .CPIWCIG 9QTMUJQR *CTTKOCP 0; (GDTWCT[ =? $CWO .' 6 2GVTKG ) 5QWNGU CPF 0 9GKUU A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains #PP /CVJ 5VCV 41 =? ' $QEEJKGTK Vector quantization for efficient computation of continuous density likelihoods, 2TQE +''' +%#552 2 /KPPGCRQNKU /0 /C[ =? ( $TWIPCTC / %GVVQNQ / (GFGTKEQ CPF & )KWNKCPK A Baseline for the Transcription of Italian Broadcast News, 2TQE +''' +%#552 +UVCPDWN 6WTMG[ ,WPG =? . %JCUG Word and acoustic confidence annotation for large vocabulary speech recognition 2TQE '5%# 'WTQURGGEJŏ 4JQFGU )TGGEG 5GRVGODGT =? . %JCUG 4 4QUGPDGTI # *CWRVOCPP / 4CXKUJCPMCT ' 6JC[GT 2 2NCEG YC[ 4 9GKFG CPF % .W Improvements in Language, Lexical and Phonetic Modeling in Sphinx-II, 2TQE #42# 5RQMGP .CPIWCIG 5[UVGOU 6GEJPQNQI[ 9QTMUJQR #WUVKP 6: ,CPWCT[ =? 5( %JGP CPF , )QQFOCP An empirical study of smoothing techniques for language modeling, %QORWVGT 5RGGEJ .CPIWCIG 13 1EVQDGT =? 55 %JGP CPF 25 )QRCNCMTKUJPCP Speaker, Environment and Channel Change Detection and Clustering via the Bayesian Information Criterion 2TQE *# $TQCFECUV 0GYU 6TCPUETKRVKQP 7PFGTUVCPFKPI 9QTMUJQR .CPFUFQYPG 8# (GDTWCT[ =? ;. %JQY 4 5EJYCTV\ 5 4QWMQU 1 -KODCNN 2 2TKEG ( -WDCNC /1 &WPJCO / -TCUPGT CPF , /CMJQWN The Role of Word-Dependent Coarticulatory Effects in a Phoneme-Based Speech Recognition System 2TQE +''' +%#552 3 6QM[Q ,CRCP #RTKN =? 2 %NCTMUQP CPF 4 4QUGPHGNF Statistical Language Modelling using CMUCambridge Toolkit, 2TQE '5%# 'WTQ5RGGEJŏ 4JQFGU )TGGEG 5GRVGODGT =? 5 &CXKU CPF 2 /GTOGNUVGKP Comparison of Parametric Representations of Monosyllabic Word Recognition in Continuously Spoken Sentences, +''' 6TCPU #EQWUVKEU 5RGGEJ 5KIPCN 2TQEGUUKPI 28
E\&5&3UHVV//&
=? &GORUVGT #2 // .CKTF CPF &$ 4WDKP Maximum Likelihood from Incomplete Data via the EM Algorithm ,QWTPCN QH VJG 4Q[CN 5VCVKUVKECN 5QEKGV[ 5GTKGU $ OGVJQFQNQIKECN 39 =? 0 &GUJOWMJ # )CPCRCVJKTCLW 4, &WPECP CPF , 2KEQPG Human Speech Recognition Performance on the 1995 CSR Hub-3 Corpus 2TQE #42# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR *CTTKOCP 0; (GDTWCT[ =? 8 &KICNCMKU CPF * /WTXGKV Genones: Optimization the Degree of Tying in a Large Vocabulary HMM-based Speech Recognizer, 2TQE +''' +%#552 1 #FGNCKFG #WUVTCNKC #RTKN =? 8 &KICNCMKU & 4VKEJGX CPF .) 0GWOG[GT Speaker adaptation using constrained estimation of Gaussian mixtures +''' 6TCPU QP 5RGGEJ #WFKQ 3 5GRVGODGT =? , &TG[HWU)TCH Sonograph and Sound Mechanics, , #EQWUV 5QE #OGTKEC 22 =? * &WFNG[ CPF 5 $CNCUJGM Automatic Recognition of Phonetic Patterns in Speech, , #EQWUV 5QE #OGTKEC 30 =? 9, 'DGN CPF , 2KEQPG Human Speech Recognition Performance on the 1994 CSR Spoke 10 Corpus 2TQE #42# 5RQMGP .CPIWCIG 5[UVGOU 6GEJPQNQI[ 9QTMUJQR #WUVKP 6: ,CPWCT[ =? 5 (WTWK Comparison of speaker recognition methods using statistical features and dynamic features, +''' 6TCPU QP #EQWUVKEU 5RGGEJ 5KIPCN 2TQEGUUKPI ASSP-29 =? /,( )CNGU CPF 5, ;QWPI An improved approach to hidden Markov model decomposition of speech and noise, 2TQE +''' +%#552 5CP (TCPEKUEQ %# /CTEJ =? /,( )CNGU CPF 5, ;QWPI Robust Continuous Speech Recognition using Parallel Model Combination, %QORWVGT 5RGGEJ .CPIWCIG 9 1EVQDGT =? /,( )CNGU Cluster Adaptive Training for Speech Recognition, 2TQE +% 5.2ŏ 5[FPG[ #WUVTCNKC 0QXGODGT =? /,( )CNGU Semi-Tied Covariance Matrices for Hidden Markov Models, +''' 6TCPU QP 5RGGEJ CPF #WFKQ 7 /C[ =? ,. )CWXCKP ) #FFC . .COGN CPF / #FFC&GEMGT Transcribing Broadcast News: The LIMSI Nov96 Hub4 System, 2TQE #42# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR %JCPVKNN[ 8# (GDTWCT[ =? ,. )CWXCKP 5 $GPPCEGH . &GXKNNGTU . .COGN CPF 4 4QUUGV Spoken Language component of the MASK Kiosk KP - 8CTIJGUG 5 2ƀGIGT 'FU *WOCP %QOHQTV CPF UGEWTKV[ QH KPHQTOCVKQP U[UVGOU 5RTKPIGT8GTNCI #NUQ KP
E\&5&3UHVV//&
2TQE *WOCP %QOHQTV CPF 5GEWTKV[ 9QTMUJQR $TWUUGNU $GNIWKO 1EVQDGT =? ,. )CWXCKP ,, )CPIQNH CPF . .COGN Speech Recognition for an Information Kiosk, 2TQE +%5.2ŏ Ō 2JKNCFGNRJKC 2# 1EVQDGT =? ,. )CWXCKP . .COGN CPF ) #FFC Partitioning and Transcription of Broadcast News Data, 2TQE +%5.2ŏ 5 5[FPG[ #WUVTCNKC &GEGODGT =? ,. )CWXCKP .( .COGN CPF / #FFC&GEMGT Developments in Continuous Speech Dictation using the ARPA WSJ Task, 2TQE +''' +%#552 &GVTQKV /+ /C[ =? ,. )CWXCKP CPF %* .GG Maximum a Posteriori Estimation for Multivariate Gaussian Mixture Observations of Markov Chains, +''' 6TCPU 5RGGEJ #WFKQ 2TQEGUUKPI 2 #RTKN =? ,. )CWXCKP . .COGN CPF ) #FFC The LIMSI Broadcast News Transcription System 5RGGEJ %QOOWPKECVKQP 37 /C[ =? . )KNNKEM CPF 4 4QVJ A Rapid Match Algorithm for Continuous Speech Recognition, 2TQE *# 5RGGEJ 0CVWTCN .CPIWCIG 9QTMUJQR *KFFGP 8CNNG[ 2# ,WPG =? . )KNNKEM ; +VQ CPF , ;QWPI A Probabilistic Approach to Confidence Measure Estimation and Evaluation 2TQE +''' +%#552 /WPKEJ )GTOCP[ #RTKN =? ,4 )NCUU 6, *C\GP CPF + . *GVJGTKPIVQP Real-time Telephone-based Speech Recognition in the Jupiter Domain, 2TQE +''' +%#552 1 2JQGPKZ #< /CTEJ =? , )QFHTG[ ' *QNNKOCP CPF , /E&CPKGN SWITCHBOARD: Telephone Speech Corpus for Research and Development, 2TQE +''' +%#552 5CP (TCPEKUEQ %# /CTEJ =? +, )QQF The Population Frequencies of Species and the Estimation of Population Parameters $KQOVGTKMC 40 =? 25 )QRCNCMTKUJPCP .4 $CJN CPF 4. /GTEGT A tree search strategy for large-vocabulary continuous speech recognition, 2TQE +''' +%#552 1 &GVTQKV /+ /C[ =? 4 *CGD7ODCEJ CPF * 0G[ Linear Discriminant Analysis for Improved Large Vocabulary Continuous Speech Recognition, 2TQE +%#552 1 /CTEJ =? 6 *CKP 5' ,QJPUQP # 6WGTM 2% 9QQFNCPF CPF 5, ;QWPI Segment Generation and Clustering in the HTK Broadcast News Transcription System, 2TQE *# $TQCFECUV 0GYU 6TCPUETKRVKQP 7PFGTUVCPFKPI 9QTMUJQR .CPFUFQYPG 8# (GDTWCT[
E\&5&3UHVV//&
=? #) *CWRVOCPP / 9KVDTQEM CPF / %JTKUVGN News-on-Demand-’An Application of Informedia Technology’, &KIKVCN .KDTCTKGU /CIC\KPG 5GRVGODGT =? %6 *GORJKNN ,, )QFHTG[ CPF )4 &QFFKPIVQP The ATIS Spoken Language Systems Pilot Corpus, 2TQE *# 5RGGEJ 0CVWTCN .CPIWCIG 9QTMUJQR 2KVVUDWTIJ 2# ,WPG =? * *GTOCPUM[ Perceptual linear predictive (PLP) analysis of speech, , #EQWUV 5QE #OGTKEC 87 =? // *QEJDGTI 5, 4GPCNU #, 4QDKPUQP CPF & -GTUJCY Large vocabulary continuous speech recognition using a hybrid connectionist-HMM system, 2TQE +%5.2ŏ ;QMQJCOC ,CRCP 5GRVGODGT =? /, *WPV Signal Representation %JCRVGT QH VJG 5VCVG QH VJG #TV KP *WOCP .CPIWCIG 6GEJPQNQI[ %QNG GV CN GFU
JVVRYYYEUGQIKGFW%5.7*.6UWTXG[EJPQFGJVON =? / *YCPI CPF : *WCPI Subphonetic Modeling with Markov States - Senone, 2TQE +''' +%#552 1 5CP (TCPEKUEQ %# /CTEJ =? /; *YCPI : *WCPI CPF ( #NNGXC Predicting Unseen Triphones with Senones, 2TQE +''' +%#552 II /KPPGCRQNKU /0 #RTKN =? ( ,GNKPGM Continuous Speech Recognition by Statistical Methods, 2TQE QH VJG +''' 64 #RTKN =? ( ,GNKPGM Statistical Methods for Speech Recognition, %CODTKFIG /+6 2TGUU =? ( ,GNKPGM $ /GTKCNFQ 5 4QWMQU CPF / 5VTCWUU A Dynamic Language Model for Speech Recognition, 2TQE *# 5RGGEJ 0CVWTCN .CPIWCIG 9QTMUJQR 2CEKſE )TQXG %# (GDTWCT[ =? ( FG,QPI ,. )CWXCKP , FGD *CTVQI CPF - 0GVVGT 1 .+8': Speech Based Video Retrieval, 2TQE %$/+ŏ 6QWNQWUG (TCPEG 1EVQDGT =? ,WCPI $* Maximum-Likelihood Estimation for Mixture Multivariate Stochastic Observations of Markov Chains #66 6GEJPKECN ,QWTPCN 64 =? 5/ -CV\ Estimation of Probabilities from Sparse Data for the Language Model Component of a Speech Recognizer, +''' 6TCPU #EQWUVKEU 5RGGEJ 5KIPCN 2TQEGUUKPI ASSP-35 /CTEJ =? 6 -GOR CPF # 9CKDGN Unsupervised Training of a Speech Recognizer: Recent Experiments, 2TQE '5%# 'WTQURGGEJŏ 6 $WFCRGUV *WP ICT[ 5GRVGODGT =? & -GTUJCY #, 4QDKPUQP CPF 5, 4GPCNU The 1995 Abbot hybrid connectionist-HMM large-vocabulary recognition system, 2TQE #42# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR *CTTKOCP 0; (GDTWCT[
E\&5&3UHVV//&
=? 4 -PGUGT CPF * 0G[ Improved Clustering Techniques for Class-Based Statistical Language Modelling, 2TQE 'WTQURGGEJŏ $GTNKP 5GRVGODGT =? 4 -PGUGT CPF * 0G[ Improved backing-off for n-gram language modeling, 2TQE +''' +%#552 1 &GVTQKV /+ /C[ =? ( -WDCNC Design of the 1994 CSR Benchmark Tests, 2TQE #42# 5RQMGP .CPIWCIG 5[UVGOU 6GEJPQNQI[ 9QTMUJQR #WUVKP 6: ,CPWCT[ =? ( -WDCNC 6 #PCUVCUCMQU * ,KP , /CMJQWN . 0IW[GP 4 5EJYCTV\ CPF 0 ;WCP Toward Automatic Recognition of Broadcast News, 2TQE *# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR *CTTKOCP 0; (GDTWCT[ =? 0 -WOCT CPF #) #PFTGQW Heteroscedastic discriminant analysis and reduced rank HMMs for improved speech recognition, 5RGGEJ %QOOWPKECVKQP 26 &GEGODGT =? 4 -WJP 2 0IW[GP ,% ,WPSWC . )QNFYCUUGT 0 0KGF\KGNUMK 5 (KPEMG CPF - (KGNF / %QPVQNKPK Eigenvoices for Speaker Adaptation, 2TQE +% 5.2ŏ 5[FPG[ 0QXGODGT =? .( .COGN CPF ) #FFC On Designing Pronunciation Lexicons for Large Vocabulary, Continuous Speech Recognition, 2TQE +%5.2ŏ 1 2JKNCFGN RJKC 2# 1EVQDGT =? .( .COGN CPF 4 &G/QTK Speech Recognition of European Languages, 2TQE +''' #WVQOCVKE 5RGGEJ 4GEQIPKVKQP 9QTMUJQR 5PQYDKTF 7VCJ &G EGODGT =? .( .COGN CPF ,. )CWXCKP Continuous Speech Recognition at LIMSI, 2TQE #42# 9QTMUJQR QP %QPVKPWQWU 5RGGEJ 4GEQIPKVKQP 5VCPHQTF %# 5GRVGODGT =? .( .COGN CPF ,. )CWXCKP A Phone-based Approach to Non-Linguistic Speech Feature Identification, %QORWVGT 5RGGEJ .CPIWCIG 9 ,CPWCT[ =? . .COGN ,. )CWXCKP CPF ) #FFC Lightly Supervised and Unsupervised Acoustic Model Training %QORWVGT 5RGGEJ .CPIWCIG 16 ,CP WCT[ =? .( .COGN 5 4QUUGV 5- $GPPCEGH * $QPPGCW/C[PCTF . &GXKNNGTU CPF ,. )CWXCKP Development of Spoken Language Corpora for Travel Information 2TQE '5%# 'WTQURGGEJŏ 3 /CFTKF 5RCKP 5GRVGODGT =? -( .GG Large-vocabulary speaker-independent continuous speech recognition: The SPHINX system, 2J& 6JGUKU %CTPGIKG /GNNQP 7PKXGTUKV[ =? . .GG CPF 4% 4QUG Speaker Normalization Using Efficient Frequency Warping Procedures 2TQE +''' +%#552 1 #VNCPVC )# /C[
E\&5&3UHVV//&
=? %, .GIIGVVGT CPF 2% 9QQFNCPF Maximum Likelihood Linear Regression for Speaker Adaptation of Continuous Density Hidden Markov Models, %QORWVGT 5RGGEJ .CPIWCIG 9 #RTKN =? .KRQTCEG .4 Maximum Likelihood Estimation for Multivariate Observations of Markov Sources +''' 6TCPUCEVKQPU QP +PHQTOCVKQP 6JGQT[ IT28 =? 42 .KRROCPP Speech recognition by machines and humans, 5RGGEJ %QO OWPKECVKQP 22 ,WN[ =? & .KW CPF ( -WDCNC Fast Speaker Change Detection for Broadcast News Transcription and Indexing 2TQE '5%# 'WTQ5RGGEJŏ 3 $W FCRGUV *WPICT[ 5GRVGODGT =? /CFEQY Multi-site Data Collection for a Spoken Language Corpus, 2TQE *# 5RGGEJ 0CVWTCN .CPIWCIG 9QTMUJQR *CTTKOCP 0; (GDTWCT[ =? . /CPIW ' $TKNN # 5VQNEMG Finding Consensus in Speech Recognition: Word Error Minimization and Other Applications of Confusion Networks, %QORWVGT 5RGGEJ CPF .CPIWCIG 1EVQDGT =? $ /CM CPF ' $QEEJKGTK Subspace distribution clustering for continuous observation density hidden Markov models, 2TQE 'WTQURGGEJŏ 4JQFGU )TGGEG 5GRVGODGT =? ,, /CTKCPK Spoken Language Processing and Human-Machine Communication in the European Union Programs, KP ) 8CTKNG GF 'WTQURGGEJŏ '7 5RGGEJ 2TQLGEVU &C[ TGRQTV 4JQFGU )TGGEG 5GRVGODGT =? ,, /CTKCPK CPF .( .COGN An overview of EU programs related to conversational/interactive systems, 2TQE *# $TQCFECUV 0GYU 6TCPUETKRVKQP 7PFGTUVCPFKPI 9QTMUJQR .CPFUFQYPG 8# (GDTWCT[ =? 5 /CTVKP , .KGTOCPP CPF * 0G[ Algorithms for Bigram and Trigram Clustering, 2TQE 'WTQURGGEJŏ /CFTKF 5RCKP 5GRVGODGT =? / /C[DWT[ GF News on Demand, 5RGEKCN 5GEVKQP KP VJG %QOOWPKECVKQPU QH VJG #%/ 43 (GDTWCT[ =? & /KNNGT 4 5EJYCTV\ 4 9GKUEJGFGN CPF 4 5VQPG Named Entity Extraction from Broadcast News, 2TQE *# $TQCFECUV 0GYU 9QTMUJQR *GTPFQP 8# (GDTWCT[ =? / /QJTK / 4KNG[ & *KPFNG # .LQNKG CPF ( 2GTGKTC Full Expansion of Context-Dependent Networks in Large Vocabulary Speech Recognition, 2TQE +''' +%#552 5GCVVNG 9# /C[ =? * /WTXGKV , $WV\DGTIGT 8 &KICNCMKU CPF / 9GKPVTCWD Large-Vocabulary Dictation using SRI’s Decipher Speech Recognition System: Progressive
E\&5&3UHVV//&
Search Techniques, 2TQE +''' +%#552 II /KPPGCRQNKU /0 #RTKN =? * 0G[ The Use of a One-Stage Dynamic Programming Algorithm for Connected Word Recognition, +''' 6TCPU #EQWUVKEU 5RGGEJ CPF 5KIPCN 2TQEGUU KPI ASSP-32 #RTKN =? * 0G[ 4 *CGD7ODCEJ $* 6TCP CPF / 1GTFGT Improvements in Beam Search for 10000-Word Continuous Speech Recognition, 2TQE +''' +%#552 I 5CP (TCPEKUEQ %# /CTEJ =? . 0IW[GP CPF 4 5EJYCTV\ Single-Tree Method for Grammar-Directed Search, 2TQE +''' +%#552 2 2JQGPKZ #< /CTEJ =? ,, 1FGNN The Use of Decision Trees with Context Sensitive Phoneme Modelling, /2JKN 6JGUKU %CODTKFIG 7PKXGTUKV[ 'PIKPGGTKPI &GRV =? ,, 1FGNN 8 8CNVEJGX 2% 9QQFNCPF CPF 5, ;QWPI A One Pass Decoder Design for Large Vocabulary Recognition, 2TQE #42# *WOCP .CPIWCIG 6GEJPQNQI[ 9QTMUJQR 2TKPEGVQP 0, /CTEJ =? - 1JVUWMK 5 (WTWK 0 5CMWTCK # +YCUCMK CPF <2
E\&5&3UHVV//&
Program, 2TQE #42# 5RQMGP .CPIWCIG 5[UVGOU 6GEJPQNQI[ 9QTMUJQR #WUVKP 6: ,CPWCT[ =? &5 2CNNGVV ,) (KUEWU 9/ (KUJGT ,5 )CTQHQNQ #( /CTVKP CPF /# 2T\[DQEMK 1995 Hub-3 Multiple Microphone Corpus Benchmark Tests, 2TQE #42# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR *CTTKOCP 0; (GDTWCT[ =? &5 2CNNGVV ,) (KUEWU ,5 )CTQHQNQ #( /CTVKP CPF /# 2T\[DQEMK 1998 Broadcast News Benchmark Test Results: English and Non-English Word Error Rate Performance Measures, 2TQE *# $TQCFECUV 0GYU 9QTMUJQR *GTPFQP 8# (GDTWCT[ =? &$ 2CWN An efficient A stack decoder algorithm for continuous speech recognition with a stochastic language model, 2TQE +''' +%#552 5CP (TCPEKUEQ %# /CTEJ =? & 2QXG[ CPF 2 9QQFNCPF Improved Discriminative Training Techniques For Large Vocabulary Continuous Speech Recognition 2TQE +''' +%#552 5CNV .CMG %KV[ /C[ =? 2 2TKEG Evaluation of Spoken Language Systems: The ATIS Domain, 2TQE *# 5RGGEJ CPF 0CVWTCN .CPIWCIG 9QTMUJQR *KFFGP 8CNNG[ 2# ,WPG =? .4 4CDKPGT CPF $* ,WCPI An Introduction to Hidden Markov Models +''' #EQWUVKEU 5RGGEJ CPF 5KIPCN 2TQEGUUKPI /CIC\KPG ASSP-3 ,CPWCT[ =? /- 4CXKUJCPMCT Efficient Algorithms for Speech Recognition, 2J& 6JGUKU %CTPGIKG /GNNQP 7PKXGTUKV[ =? /& 4KNG[ 9 $[TPG / (KPMG 5 -JWFCPRW # .LQLNG , /E&QPQWIJ * 0QEM / 5CTCENCT % 9QQVGTU CPF )
E\&5&3UHVV//&
=? / 5EJWUVGT Memory-efficient LVCSR search using a one-pass stack decoder, %QORWVGT 5RGGEJ .CPIWCIG 14 ,CPWCT[ =? 4 5EJYCTV\ 5 #WUVKP ( -WDCNC CPF , /CMJQWN New uses for N-Best Sentence Hypothesis, within the BYBLOS Speech Recognition System, 2TQE +''' +%#552 I 5CP (TCPEKUEQ %# /CTEJ =? 4 5EJYCTV\ ; %JQY 5 4QWEQU / -TCUPGT CPF , /CMJQWN Improved Hidden Markov Modeling of Phonemes for Continuous Speech Recognition, 2TQE +''' +%#552 3 5CP &KGIQ %# /CTEJ =? 5 5GMKPG CPF 4 )TKUJOCP NYU Language Modeling Experiments for the 1995 CSR Evaluation, 2TQE #42# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR *CTTKOCP 0; (GDTWCT[ =? $ 5JCJUJCJCPK A Markov Random Field Approach to Bayesian Speaker Adaptation, 2TQE +''' +%#552 &GVTQKV /+ /C[ =? 4 5EJYCTV\ * ,KP ( -WDCNC CPF 5 /CVUQWMCU Modeling Those FConditions – Or Not, 2TQE *# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR %JCPVKNN[ 8# (GDTWCT[ =? - 5G[OQTG CPF 4 4QUGPHGNF Scalable backoff language models 2TQE +% 5.2ŏ 1 2JKNCFGNRJKC 2# 1EVQDGT =? / 5KGINGT 7 ,CKP $ 4CL CPF 4 5VGTP Automatic Segmentation, Classification and Clustering of Broadcast News Audio, 2TQE *# 5RGGEJ 4GEQIPK VKQP 9QTMUJQR %JCPVKNN[ 8# (GDTWCT[ =? / 5KW CPF * )KUJ Evaluation of word confidence for speech recognition systems %QORWVGT 5RGGEJ .CPIWCIG 13 1EVQDGT =? # 5VQNEMG Entropy-based Pruning of Backoff Language Models 2TQE *# $TQCFECUV 0GYU 6TCPUETKRVKQP 7PFGTUVCPFKPI 9QTMUJQR .CPFU FQYPG 8# (GDTWCT[ =? 5 6CMCJCUJK CPF 5 5CIC[COC Four-level Tied Structure for Efficient Representation of Acoustic Modeling, 2TQE +''' +%#552 &GVTQKV /+ /C[ =? .( 7GDGN CPF 2% 9QQFNCPF An Investigation into Vocal Tract Length Normalization, 2TQE '5%# 'WTQURGGEJŏ $WFCRGUV *WPICT[ 5GRVGODGT =? &# XCP .GGWYGP .) XCP FGP $GTI CPF *,/ 5VGGPGMGP Human Benchmarks for Speaker Independent Large Vocabulary Recognition Performance, 2TQE '5%# 'WTQURGGEJŏ /CFTKF 5RCKP 5GRVGODGT =? 6- 8KPVU[WM Speech discrimination by dynamic programming, -KDPGTPGVKMC 4
E\&5&3UHVV//&
=? 6- 8KPVU[WM Elements-wise recognition of continuous speech composed of words from a specified dictionary, %[DGTPGVKEU 7 /CTEJ#RTKN =? 9 9CJNUVGT Verbmobil: Translation of Face-to-Face Dialogs, 2TQE '5%# 'WTQURGGEJŏ $GTNKP )GTOCP[ Plenary 5GRVGODGT =? # 9CKDGN 2 )GWVPGT . /C[ſGNF 6QOQMK[Q 6 5EJWNV\ CPF / 9QU\E\[PC Multilinguality in Speech and Spoken Language Systems 2TQEGGFKPIU QH VJG +''' 5RGEKCN +UUWG QP 5RQMGP .CPIWCIG 2TQEGUUKPI 88 #W IWUV =? ( 9CNNU * ,KP 5 5KUVC CPF 4 5EJYCTV\ Probabilistic Models for Topic Detection and Tracking, 2TQE +''' +%#552 1 2JQGPKZ #< /CTEJ =? 5 9GIOCPP ( 5ECVVQPG + %CTR . )KNNKEM 44QVJ CPF , ;COTQP Dragon Systems’ 1997 Broadcast News Transcription System, 2TQE *# $TQCF ECUV 0GYU 6TCPUETKRVKQP 7PFGTUVCPFKPI 9QTMUJQR .CPFUFQYPG 8# (GDTWCT[ =? 5 9GIOCPP 2 <JCP CPF . )KNNKEM Progress in Broadcast News Transcription at Dragon Systems, 2TQE +''' +%#552 2JQGPKZ #< /CTEJ =? / 9GKPVTCWD ( $GCWHC[U < 4KXNKP ; -QPKI CPF # 5VQNEMG NeuralNetwork based Measures of Confidence for Word Recognition, 2TQE +''' +%#552 /WPKEJ )GTOCP[ #RTKN =? ( 9GUUGN - /CEJGTG[ CPF 4 5EJNiWVGT Using word probabilities as confidence measures, 2TQE +''' +%#552 5GCVVNG 9# /C[ =? ( 9GUUGN CPF * 0G[ Unsupervised training of acoustic models for large vocabulary continuous speech recognition 2TQE +''' #547ŏ /CFQPPC FK %CORKINKQ +VCN[ &GEGODGT =? +* 9KVVGP CPF 6% $GNN The Zero Frequency problem: Estimating the problems of Novel Events in Adaptive tex Compression 2TQE +''' 6TCPU QP +P HQTOCVKQP 6JGQT[ 37 ,WN[ =? 2% 9QQFNCPF CPF & 2QXG[ Large scale discriminative training of hidden Markov models for speech recognition, %QORWVGT 5RGGEJ CPF .CPIWCIG 16 ,CPWCT[ =? 2% 9QQFNCPF %, .GIIGVVGT ,, 1FGNN 8 8CNVEJGX CPF 5, ;QWPI The development of the 1994 HTK large vocabulary speech recognition system, 2TQE #42# 5RQMGP .CPIWCIG 5[UVGOU 6GEJPQNQI[ 9QTMUJQR #WUVKP 6: ,CPWCT[ =? 2% 9QQFNCPF /,( )CNGU & 2[G CPF 8 8CNVEJGX The HTK large vocabulary recognition system for the 1995 ARPA H3 task, 2TQE #42# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR *CTTKOCP 0; (GDTWCT[
E\&5&3UHVV//&
=? ,2 ;COTQP + %CTR . )KNNKEM 5 .QYG CPF 2 XCP /WNDTGIV A Hidden Markov Approach to Text Segmentation and Event Tracking 2TQE +''' +%#552 1 5GCVVNG 9# /C[ =? 5, ;QWPI A Review of Large-Vocabulary Continuous Speech Recognition, +''' 5KIPCN 2TQEGUUKPI /CIC\KPG 13 5GRVGODGT =? 5, ;QWPI / #FFC&GEMGT : #WDGTV % &WICUV ,. )CWXCKP &, -GT UJCY . .COGN &# .GGWYGP & 2[G *,/ 5VGGPGMGP #, 4QDKPUQP CPF 2% 9QQFNCPF Multilingual large vocabulary speech recognition: the European SQALE project, %QORWVGT 5RGGEJ .CPIWCIG 11 ,CPWCT[ =? 5, ;QWPI CPF . %JCUG Speech recognition evaluation: a review of the U.S. CSR and LVCSR programmes, %QORWVGT 5RGGEJ .CPIWCIG 12 1EVQDGT =? 5, ;QWPI ,, 1FGNN CPF 2% 9QQFNCPF Tree-Based State Tying for High Accuracy Acoustic Modeling, 2TQE #42# *WOCP .CPIWCIG 6GEJPQNQI[ 9QTM UJQR 2TKPEGVQP 0, /CTEJ =? 5, ;QWPI CPF 2% 9QQFNCPF The Use of State Tying in Continuous Speech Recognition, 2TQE '5%# 'WTQURGGEJŏ 3 $GTNKP )GTOCP[ 5GRVGODGT =? )
E\&5&3UHVV//&
6 Toward Spontaneous Speech Recognition and Understanding Sadaoki Furui Tokyo Institute of Technology
CONTENTS
+PVTQFWEVKQP (QWT %CVGIQTKGU QH 5RGGEJ 4GEQIPKVKQP 6CUMU 5RQPVCPGQWU 5RGGEJ 4GEQIPKVKQP CPF 7PFGTUVCPFKPI 4GXKGY ,CRCPGUG 0CVKQPCN 2TQLGEV QP 5RQPVCPGQWU 5RGGEJ %QTRWU CPF 2TQEGUUKPI 6GEJPQNQI[ #WVQOCVKE 6TCPUETKRVKQP QH 5RQPVCPGQWU 2TGUGPVCVKQP #WVQOCVKE 5RGGEJ 5WOOCTK\CVKQP CPF 'XCNWCVKQP 5RQPVCPGQWU 5RGGEJ 4GEQIPKVKQP CPF 7PFGTUVCPFKPI 4GUGCTEJ +UUWGU %QPENWUKQP 4GHGTGPEGU
6.1 Introduction 5RGGEJ TGEQIPKVKQP U[UVGOU CTG GZRGEVGF VQ RNC[ KORQTVCPV TQNGU KP CP CFXCPEGF OWNVKOGFKC UQEKGV[ YKVJ WUGTHTKGPFN[ JWOCPOCEJKPG KPVGTHCEGU =? 6JG ſGNF QH CWVQOCVKE URGGEJ TGEQIPKVKQP JCU YKVPGUUGF C PWODGT QH UKIPKſECPV CFXCPEGU KP VJG RCUV [GCTU URWTTGF QP D[ CFXCPEGU KP UKIPCN RTQEGUUKPI CNIQTKVJOU EQORWVC VKQPCN CTEJKVGEVWTGU CPF JCTFYCTG 6JGUG CFXCPEGU KPENWFG VJG YKFGURTGCF CFQRVKQP QH C UVCVKUVKECN RCVVGTP TGEQIPKVKQP RCTCFKIO C FCVCFTKXGP CRRTQCEJ YJKEJ OCMGU WUG QH C TKEJ UGV QH URGGEJ WVVGTCPEGU HTQO C NCTIG RQRWNCVKQP QH URGCMGTU VJG WUG QH UVQEJCUVKE CEQWUVKE CPF NCPIWCIG OQFGNKPI CPF VJG WUG QH F[PCOKE RTQITCOOKPI DCUGF UGCTEJ OGVJQFU = ? 6JG UVCVGQHVJGCTV KP CWVQOCVKE URGGEJ TGEQIPKVKQP ECP DG CFFTGUUGF KP UGXGTCN YC[U (KIWTG KNNWUVTCVGU VJG RTQITGUU QH URGGEJ TGEQIPKVKQP CPF WPFGTUVCPFKPI VGEJPQNQI[ CEEQTFKPI VQ IGPGTKE CRRNKECVKQP CTGCU TCPIKPI HTQO KUQNCVGF YQTF QT EQOOCPF TGEQIPKVKQP VQ PCVWTCN EQPXGTUCVKQP DGVYGGP JWOCP CPF OCEJKPG 6JG EQORNGZKV[ QH VJGUG IGPGTKE CRRNKECVKQP CTGCU KU EJCTCEVGTK\GF CNQPI VYQ FKOGPUKQPU VJG UK\G QH VJG XQECDWNCT[ CPF VJG URGCMKPI UV[NG +V UJQWNF DG QDXKQWU VJCV VJG NCTIGT VJG XQECDWNCT[ VJG OQTG FKHſEWNV VJG CRRNKECVKQP VCUM 5KOKNCTN[ VJG FGITGG QH EQP UVTCKPVU KP VJG URGCMKPI UV[NG JCU C XGT[ FKTGEV KPƀWGPEG QP VJG EQORNGZKV[ QH VJG
E\&5&3UHVV//&
6SHDNLQJVW\OH
6SRQWDQHRXV VSHHFK ZRUG VSRWWLQJ
)OXHQW VSHHFK
GLJLW VWULQJV
5HDG VSHHFK &RQQHFWHG VSHHFK
QDWXUDO FRQYHUVDWLRQ ZD\ GLDORJXH QHWZRUN WUDQVFULSWLRQ DJHQW V\VWHPGULYHQ LQWHOOLJHQW GLDORJXH PHVVDJLQJ QDPH GLDOLQJ RIILFH IRUPILOO GLFWDWLRQ E\YRLFH
YRLFH FRPPDQGV
,VRODWHG ZRUGV
GLUHFWRU\ DVVLVWDQFH
8QUHVWULFWHG 9RFDEXODU\VL]HQXPEHURIZRUGV
FIGURE 6.1 Progress of spoken language technology along the dimensions of vocabulary size and speaking styles.
CRRNKECVKQP C HTGG EQPXGTUCVKQP HWNN QH UNWTTKPI CPF GZVTCPGQWU UQWPFU UWEJ CU őWJŒ őWOŒ CPF RCTVKCN YQTFU KU HCT OQTG FKHſEWNV VJCP YQTFU URQMGP KP C TKIKFN[ FKUETGVG OCPPGT 6JWU VJG FKHſEWNV[ QH CP CRRNKECVKQP ITQYU HTQO VJG NQYGT NGHV EQTPGT VQ VJG WRRGT TKIJV EQTPGT KP VJG ſIWTG 6JG VJTGG DCTU KP VJG ſIWTG FGOCTECVG VJG CRRNKEC VKQPU VJCV ECP CPF ECPPQV DG UWRRQTVGF D[ VJG VGEJPQNQI[ HQT XKCDNG FGRNQ[OGPV KP VJG EQTTGURQPFKPI VKOG HTCOG +V UJQWNF DG PQVGF VJCV VJGUG VJTGG DCTU CTG PQV RCTCNNGN YJKEJ OGCPU VJCV VJG RTQITGUU QH URQPVCPGQWU URGGEJ TGEQIPKVKQP CPF WPFGTUVCPFKPI KU OWEJ UNQYGT VJCP VJCV QH OQTG TKIKFN[ URQMGP WVVGTCPEGU %QOOQP HGCVWTGU QH UVCVGQHVJGCTV URGGEJ TGEQIPKVKQP U[UVGOU GZKUV KP WUKPI EGR UVTCN RCTCOGVGTU CPF VJGKT TGITGUUKQP EQGHſEKGPVU CU URGGEJ HGCVWTGU VTKRJQPG *//U CU CEQWUVKE OQFGNU XQECDWNCTKGU QH UGXGTCN VJQWUCPF QT UGXGTCN VGP VJQWUCPF GPVTKGU CPF UVCVKUVKECN NCPIWCIG OQFGNU UWEJ CU DKITCOU CPF VTKITCOU 5WEJ OGVJQFU JCXG DGGP CRRNKGF PQV QPN[ VQ 'PINKUJ DWV CNUQ VQ (TGPEJ )GTOCP +VCNKCP CPF ,CRCPGUG CPF CNVJQWIJ VJGTG CTG UGXGTCN NCPIWCIGURGEKſ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
E\&5&3UHVV//&
NKPIWKUVKE OQFGNU WUGF JCXG IGPGTCNN[ DGGP DWKNV WUKPI YTKVVGP NCPIWCIG QT URGGEJ HTQO YTKVVGP NCPIWCIG 7PHQTVWPCVGN[ URQPVCPGQWU URGGEJ CPF URGGEJ HTQO YTKVVGP NCPIWCIG CTG XGT[ FKHHGTGPV DQVJ CEQWUVKECNN[ CPF NKPIWKUVKECNN[ $TQCFGPKPI VJG CR RNKECVKQP QH URGGEJ TGEQIPKVKQP VJWU ETWEKCNN[ FGRGPFU QP TCKUKPI VJG TGEQIPKVKQP RGT HQTOCPEG HQT URQPVCPGQWU URGGEJ +P QTFGT VQ KPETGCUG VJG TGEQIPKVKQP RGTHQTOCPEG HQT URQPVCPGQWU URGGEJ KV KU ETWEKCN VQ DWKNF CEQWUVKE CPF NCPIWCIG OQFGNU HQT URQP VCPGQWU URGGEJ /GVJQFU CRRN[KPI UVCVKUVKECN NCPIWCIG OQFGNKPI UWEJ CU DKITCOU CPF VTKITCOU QH YQTFU QT OQTRJGOGU VQ URQPVCPGQWU URGGEJ EQTRWU OC[ PQV DG CFG SWCVG 1WT MPQYNGFIG QH VJG UVTWEVWTG QH URQPVCPGQWU URGGEJ KU EWTTGPVN[ KPCFGSWCVG VQ CEJKGXG VJG PGEGUUCT[ DTGCMVJTQWIJU #NVJQWIJ URQPVCPGQWU URGGEJ GHHGEVU CTG SWKVG EQOOQP KP JWOCP EQOOWPKECVKQP CPF OC[ DG GZRGEVGF VQ KPETGCUG KP JWOCP OCEJKPG FKUEQWTUG CU RGQRNG DGEQOG OQTG EQOHQTVCDNG EQPXGTUKPI YKVJ OCEJKPGU OQFGNKPI QH URGGEJ FKUƀWGPEKGU KU QPN[ LWUV DGIKPPKPI 4GEQIPKVKQP QH URQPVCPGQWU URGGEJ YKNN TGSWKTG C RCTCFKIO UJKHV HTQO URGGEJ TGEQIPKVKQP VQ WPFGTUVCPFKPI YJGTG WPFGTN[KPI OGUUCIGU QH VJG URGCMGT CTG GZVTCEVGF KPUVGCF QH VTCPUETKDKPI CNN VJG URQ MGP YQTFU =? /WEJ QH QWT VJKPMKPI CDQWV URGGEJ TGEQIPKVKQP JCU DGGP HQEWUGF QP KVU WUG CU CP KPVGTHCEG KP JWOCPOCEJKPG KPVGTCEVKQPU OQUVN[ HQT KPHQTOCVKQP CEEGUU CPF GZVTCE VKQP 9KVJ KPETGCUGU KP EGNNWNCT RJQPG WUG CPF FGRGPFGPEG QP PGVYQTMGF KPHQTOCVKQP TGUQWTEGU CPF CU TCRKF CEEGUU VQ KPHQTOCVKQP DGEQOGU CP KPETGCUKPIN[ KORQTVCPV GEQ PQOKE HCEVQT VGNGRJQPG CEEGUU VQ FCVC CPF VGNGRJQPG VTCPUCEVKQPU YKNN PQ FQWDV TKUG FTCOCVKECNN[ 6JGTG KU C ITQYKPI KPVGTGUV JQYGXGT KP XKGYKPI URGGEJ PQV LWUV CU C OGCPU VQ CEEGUU KPHQTOCVKQP DWV CU KVUGNH C UQWTEG QH KPHQTOCVKQP +ORQTVCPV CV VTKDWVGU VJCV YQWNF OCMG URGGEJ OQTG WUGHWN KP VJKU TGURGEV KPENWFG TCPFQO CEEGUU UQTVKPI GI D[ URGCMGT D[ VQRKE D[ WTIGPE[ UECPPKPI CPF GFKVKPI *QY EQWNF QWT NKXGU DG EJCPIGF D[ UWEJ VQQNU! 'PCDNKPI UWEJ C XKUKQP EJCNNGPIGU QWT U[UVGOU UVKNN HWTVJGT KP PQKUG TQDWUVPGUU CPF KP URQPVCPGQWU URGGEJ GHHGEVU 9G ECP GPXKUKQP C ITGCV KPHQTOCVKQP TGXQNWVKQP QP RCT YKVJ VJG FGXGNQROGPV QH YTKV KPI U[UVGOU KH YG ECP UWEEGUUHWNN[ OGGV VJG EJCNNGPIGU QH URGGEJ DQVJ CU C OGFKWO HQT KPHQTOCVKQP CEEGUU CPF CU KVUGNH C UQWTEG QH KPHQTOCVKQP 5RGGEJ KU UVKNN VJG OGCPU QH EQOOWPKECVKQP WUGF ſTUV CPF HQTGOQUV D[ JWOCPU CPF QPN[ C UOCNN RGTEGPVCIG QH JWOCP EQOOWPKECVKQP KU YTKVVGP #WVQOCVKE URGGEJ WPFGTUVCPFKPI ECP CFF OCP[ QH VJG CFXCPVCIGU PQTOCNN[ CUUQEKCVGF QPN[ YKVJ VGZV TCPFQO CEEGUU UQTVKPI CPF CEEGUU CV FKHHGTGPV VKOGU CPF RNCEGU VQ VJG OCP[ DGPGſVU QH URGGEJ /CMKPI VJKU XKUKQP C TGCNKV[ YKNN TGSWKTG UKIPKſECPV CFXCPEGU
6.2 Four Categories of Speech Recognition Tasks 5RGGEJ TGEQIPKVKQP VCUMU ECP DG ENCUUKſGF KPVQ HQWT ECVGIQTKGU CU UJQYP KP 6CDNG CEEQTFKPI VQ VYQ ETKVGTKC YJGVJGT KV KU VCTIGVKPI WVVGTCPEGU HTQO JWOCP VQ JWOCP QT JWOCP VQ EQORWVGT CPF YJGVJGT VJG WVVGTCPEGU JCXG C FKCNQIWG QT OQPQNQIWG UV[NG
E\&5&3UHVV//&
TABLE 6.1
%CVGIQTK\CVKQP QH URGGEJ TGEQIPKVKQP VCUMU
*WOCP VQ JWOCP
*WOCP VQ OCEJKPG
&KCNQIWG
%CVGIQT[ + 5YKVEJDQCTF %CNN *QOG *WD OGGVKPI VCUM
%CVGIQT[ +++ #6+5 %QOOWPKECVQT KPHQTOCVKQP TGVTKGXCN TGUGTXCVKQP
/QPQNQIWG
%CVGIQT[ ++ $TQCFECUVU PGYU *WD NGEVWTG RTGUGPVCVKQP XQKEG OCKN
%CVGIQT[ +8 &KEVCVKQP
6JG VCDNG NKUVU V[RKECN VCUMU HQT GCEJ ECVGIQT[ /QUV QH VJG RTCEVKECN CRRNKECVKQP U[UVGOU YKFGN[ WUGF PQY CTG ENCUUKſGF CU %CVGIQT[ +++ TGEQIPK\KPI VJG WVVGTCPEGU KP JWOCPEQORWVGT FKCNQIWGU UWEJ CU KP VJG CKTNKPG KPHQTOCVKQP UGTXKEGU VCUM *#URQPUQTGF RTQLGEVU KPENWFKPI #6+5 CPF %QO OWPKECVQT CTG NC[KPI VJG HQWPFCVKQPU QH VJGUG U[UVGOU 7PNKMG QVJGT ECVGIQTKGU VJG U[UVGOU KP VJG %CVGIQT[ +++ CTG WUWCNN[ FGUKIPGF CPF FGXGNQRGF CHVGT ENGCTN[ FGſPKPI VJG CRRNKECVKQPVCUM 6JG OCEJKPG VJCV YG JCXG CVVGORVGF VQ FGUKIP UQ HCT KU CNOQUV YKVJQWV GZEGRVKQP NKOKVGF VQ VJG UKORNG VCUM QH EQPXGTVKPI C URGGEJ UKIPCN KPVQ C YQTF UGSWGPEG CPF VJGP FGVGTOKPKPI HTQO VJG YQTF UGSWGPEG VJG OGCPKPI VJCV KU őWPFGTUVCPFCDNGŒ *GTG VJG UGV QH WPFGTUVCPFCDNG OGUUCIGU KU ſ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
E\&5&3UHVV//&
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ſTUV DTKGƀ[ TGXKGYU %CVGIQTKGU + ++ CPF +++ CPF VJGP KV HQEWUGU QP %CVGIQT[ ++ # NCTIGUECNG PCVKQPCN RTQLGEV VQ KPXGUVKICVG VJG KUUWGU QH URQPVCPGQWU URGGEJ TGEQIPKVKQP KU KPVTQFWEGF
6.3 Spontaneous Speech Recognition and Understanding - Review 6.3.1 Category I (human-to-human dialogue) 5YKVEJDQCTF =? KU C *#URQPUQTGF NCTIG OWNVKURGCMGT EQTRWU QH URQPVCPGQWU EQPXGTUCVKQPCN VGNGRJQPG DCPFYKFVJ URGGEJ CPF VGZV HQT TGUGCTEJ QP NCTIG XQECDWNCT[ URGGEJ TGEQIPKVKQP CPF URGCMGT CWVJGPVKECVKQP #DQWV EQPXGTUCVKQPU D[ URGCMGTU HTQO CTQWPF VJG 75 YGTG EQNNGEVGF CWVQOCVKECNN[ QXGT 6 NKPGU +P GCEJ EQPXGTUCVKQP VYQ URGCMGTU YGTG CUMGF VQ FKUEWUU QPG QH FKHHGTGPV VQRKEU UWEJ CU RGVU ETKOG QT CKT RQNNWVKQP 6JGUG EQPXGTUCVKQPU CTG QH FWTCVKQP VJTGG VQ VGP OKPWVGU ſXG OKPWVGU KP CXGTCIG CPF URQMGP D[ RCKF XQNWPVGGTU QH DQVJ UGZGU KP GXGT[ OCLQT FKCNGEV QH #OGTKECP 'PINKUJ 6JKU COQWPVU VQ QXGT JQWTU QH URGGEJ CPF PGCTN[ VJTGG OKNNKQP YQTFU QH VGZV 4GEQIPKVKQP QH WVVGTCPEGU KP VJG 5YKVEJDQCTF EQTRWU KU C XGT[ EJCNNGPIKPI VCUM 1TCN EQOOWPKECVKQP KU VTCPUKGPV DWV OCP[ KORQTVCPV FGEKUKQPU UQEKCN EQPVTCEVU CPF HCEV ſPFKPIU CTG ſTUV ECTTKGF QWV QTCNN[ FQEWOGPVGF KP YTKVVGP HQTO CPF NCVGT TG VTKGXGF *WOCPU URGPF C NQV QH VKOG VTCPUHQTOKPI QTCN EQOOWPKECVKQPU KPVQ YTKVVGP FQEWOGPVU 4GUGCTEJ HQEWUKPI QP CWVQOCVKE OGGVKPI TGEQTF ETGCVKQP CPF CEEGUU JCU DGGP EQPFWEVGF =? 6JG TGUGCTEJ CKOU CV C TGCNKUVKE OGGVKPI UEGPCTKQ VJG EQTTG URQPFKPI URGGEJ TGEQIPKVKQP RTQDNGOU VJG CPCN[UKU QH TGVTKGXCN RGTHQTOCPEG VJG IGPGTCVKQP QH TGCFCDNG UWOOCTKGU CPF C RTCEVKECN WUGT KPVGTHCEG /GGVKPI TGEQIPK VKQP KU C XGT[ EJCNNGPIKPI .8%54 VCUM RCTCNNGN VQ VJCV QH *WD 5YKVEJDQCTF CPF *WD $TQCFECUV 0GYU 6JG FKHſ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
E\&5&3UHVV//&
VCNMKPI JGCFUGV 6JKU EQOGU CV VJG EQUV QH UKIPKſECPV EJCPPGN ETQUUVCNM
6.3.2 Category II (human-to-human monologue) 6JG *#URQPUQTGF *WD RTQLGEV JCU DGGP C FTKXKPI HQTEG DGJKPF TGUGCTEJ QP JWOCPVQJWOCP OQPQNQIWG URGGEJ TGEQIPKVKQP UKPEG =? +P VJKU RTQLGEV VGNG XKUKQP CPF TCFKQ PGYU DTQCFECUVU CTG TGEQTFGF CPF CPPQVCVGF 6JG OCVGTKCNU EQP UKUV QH YJCV JCU DGGP VGTOGF őHQWPF URGGEJŒ őHQWPFŒ KP PGYU DTQCFECUVU KP EQP VTCUV YKVJ VJG URGEKCNN[ TGEQTFGF őTGCFŒ URGGEJ KPXGUVKICVGF KP VJG HQTOGT *# ő0QTVJ #OGTKEC $WUKPGUU 0#$ PGYUŒ RTQLGEV +V RTQXGF VQ QHHGT C TKEJ CUUQTV OGPV QH VGEJPKECN EJCNNGPIGU VQ VJG EQOOWPKV[ KPENWFKPI XCTKGF URGCMKPI UV[NGU HQTGKIPCEEGPVGF 'PINKUJ VJG RTGUGPEG QH DCEMITQWPF OWUKE CPF DQVJ HWNN CU YGNN CU TGFWEGFDCPFYKFVJ EJCPPGN GHHGEVU 6JG NQYGUV YQTF GTTQT TCVGU KP VJG DGPEJ OCTM VGUV TGUWNVU HQT VJG NQYPQKUG DCUGNKPG ( CPF URQPVCPGQWU ( EQPFKVKQPU YGTG CPF TGURGEVKXGN[ 6JG FCVCDCUG YCU NCVGT GZVGPFGF VQ KPENWFG /CPFCTKP %JKPGUG CPF 5RCPKUJ +P C PGY VCUM őURQMGŒ YCU CFFGF VQ *WD VQ GZCOKPG VJG GHHGEVKXGPGUU QH DTQCFECUV PGYU TGEQIPKVKQP VGEJPQNQI[ KP IGPGTCVKPI KPHQTOCVKQP TKEJ GPVKVKGU CPF VQ DGIKP VQ OQXG VJG TGUGCTEJ HQEWU HTQO UKORNG VTCPUETKRVKQP VQYCTF URQMGP KPHQT OCVKQP WPFGTUVCPFKPI 6JG VCUM KPXQNXGF VJG TGEQIPKVKQP CPF KFGPVKſ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ſPKVKQP QH YQTFU VJG DTQCFECUV PGYU OCPWUETKRVU YGTG UGIOGPVGF KPVQ YQTFU OQTRJGOGU WUKPI C OQTRJQNQIKECN CPCN[\GT VQ ECNEWNCVG YQTF PITCO NCPIWCIG OQFGNU /CP[ ,CRCPGUG YQTFU JCXG OWNVKRNG TGCFKPIU CPF VJG EQTTGEV QPG ECP QPN[ DG FGEKFGF CEEQTFKPI VQ VJG EQP VGZV 6JGTGHQTG NCPIWCIG OQFGNU KP YJKEJ C YQTF YKVJ OWNVKRNG TGCFKPIU KU URNKV KPVQ FKHHGTGPV NCPIWCIG OQFGN GPVTKGU CEEQTFKPI VQ VJQUG TGCFKPIU JCXG DGGP EQP UVTWEVGF 5KPEG XCTKQWU EJCTCEVGTU UWEJ CU %JKPGUG EJCTCEVGTU OWNVKRNG V[RGU QH ,CRCPGUG EJCTCEVGTU PWODGTU CPF CNRJCDGVU CTG WUGF KP ,CRCPGUG VGZV KV KU JCTF VQ V[RG ,CRCPGUG VGZV KP TGCN VKOG 6JGTGHQTG EQORWVGTDCUGF U[UVGOU CTG KPFKURGPU CDNG HQT QPNKPG ,CRCPGUG ENQUGF ECRVKQPKPI 0*- UVCTVGF VJG ENQUGF ECRVKQPKPI WUKPI C TGCNVKOG URGGEJ TGEQIPK\GT HQNNQYGF D[ OCPWCN EQTTGEVKQP QH TGEQIPKVKQP GTTQTU KP /CTEJ 5KPEG TGEQIPKVKQP CEEWTCE[ HQT URQPVCPGQWU URGGEJ KU PQV [GV UCVKU HCEVQT[ ENQUGF ECRVKQPKPI KU RTQXKFGF QPN[ HQT VJG URGGEJ WVVGTGF D[ CPEJQTU KP VJG UVWFKQ 9KVJ VJG KPETGCUKPI PWODGT QH FKHHGTGPV OGFKC UQWTEGU HQT KPHQTOCVKQP FKUUGOKPC VKQP VJGTG KU C TCRKFN[ ITQYKPI PGGF HQT HCUV CWVQOCVKE RTQEGUUKPI QH CWFKQ FCVC UVTGCO #WVQOCVKQP QH CWFKQ UGIOGPVCVKQP VTCPUETKRVKQP CPF KPFGZCVKQP KU KPFKU RGPUCDNG # URQMGP FQEWOGPV KPFGZKPI CPF TGVTKGXCN U[UVGO EQODKPKPI C UVCVGQH
E\&5&3UHVV//&
$XGL[ VHUYHU
323VHUYHU
$65VHUYHU ,(VHUYHU
&DOOHU,'VHUYHU 6FDQ0DLO +XE'% (PDLOVHUYHU
,5VHUYHU &OLHQW
FIGURE 6.2 The SCANMail architecture [12]. VJGCTV URGGEJ TGEQIPK\GT YKVJ C VGZVDCUGF KPHQTOCVKQP TGVTKGXCN +4 U[UVGO JCU DGGP KPXGUVKICVGF =? 9KVJ SWGT[ GZRCPUKQP WUKPI EQOOGTEKCN VTCPUETKRVU EQO RCTCDNG OGCP RTGEKUKQPU JCXG DGGP QDVCKPGF QP OCPWCN TGHGTGPEG VTCPUETKRVKQPU CPF CWVQOCVKE VTCPUETKRVKQPU YKVJ C YQTF GTTQT TCVG QH 8QKEGOCKN URGGEJ TGEQIPKVKQP RTGUGPVU C EJCNNGPIKPI RTQDNGO UKPEG KV KU EJCTCEVGT K\GF D[ C XCTKGV[ QH URGCMKPI TCVGU CEEGPVU VCUMU CPF CEQWUVKE EQPFKVKQPU #FFK VKQPCNN[ RJGPQOGPC UWEJ CU FKUƀWGPEKGU TGUVCTVU TGRGVKVKQPU CPF DTQMGP YQTFU CTG EQOOQP +P EQPVTCUV VQ PCVWTCN FKCNQIWG XQKEGOCKN URGGEJ KU OQPQNQIWG VJCV KU C őQPGYC[Œ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ſGF KP VJG WUGTŏU RTQſNG (KIWTG UJQYU VJG CTEJKVGEVWTG QH VJG U[UVGO 6JG NCPIWCIG OQFGN HQT #54 KU C -CV\UV[NG DCEMQHH VTKITCO VTCKPGF QP YQTFU HTQO VJG VTCPUETKRVKQPU QH VJG JQWT VTCKPKPI UGV #P KORQTVCPV KUUWG TGNCVGF VQ VJG FGXGNQROGPV QH KPVGITCVGF XQKEGFCVC EQOOWPK ECVKQPU KU VJCV QH URGGEJ UWOOCTK\CVKQP IKXGP C URQMGP RCUUCIG RTQFWEG C UJQTV VGZVWCN RTGEKU QH KVU EQPVGPV # U[UVGO VJCV VTCPUOKVU VGZV UWOOCTKGU QH C WUGTŏU KP
E\&5&3UHVV//&
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ſGT QT HGCVWTG UGV # ,CRCPGUG PCVKQPCN RTQLGEV QP URQPVCPGQWU URGGEJ EQTRWU CPF RTQEGUUKPI VGEJPQN QI[ YCU KPKVKCVGF KP 6JKU RTQLGEV CKOU VQ DWKNF C NCTIGUECNG OQPQNQIWG URQPVC PGQWU URGGEJ EQTRWU CPF ETGCVG URQPVCPGQWU URGGEJ TGEQIPKVKQP CPF UWOOCTK\CVKQP VGEJPQNQI[ &GVCKNU YKNN DG GZRNCKPGF KP 5GEVKQP
6.3.3 Category III (human-to-machine dialogue) 6JGTG KU C ITQYKPI KPVGTGUV KP OQDKNG EQOOWPKECVKQP U[UVGOU VJCV CNNQY WUGTU VQ WUG VJGKT XQKEGU VQ FQ OQTG VJCP URGCMKPI VQ QVJGT RGQRNG GZCORNGU KPENWFG CEEGUUKPI KPHQTOCVKQP UGTXKEGU CPF KPVGTCEVKQP YKVJ DQQMKPI UGTXKEGU 2TQXKFKPI XQKEG KPVGTCEVKQP ECRCDKNKV[ CU C RCTV QH OWNVKOGFKC WUGT GZRGTKGPEG KU DG NKGXGF VQ CFF PCVWTCNPGUU CPF GHſEKGPE[ VQ JWOCPEQORWVGT KPVGTCEVKQPU 0WOGTQWU EQOOGTEKCN URQMGP FKCNQI U[UVGOU CTG EWTTGPVN[ DGKPI FGRNQ[GF RTKOCTKN[ HQT CE EGUU VQ KPHQTOCVKQP QXGT VJG VGNGRJQPG 6JGTG CTG JQYGXGT OCLQT QRGP TGUGCTEJ KUUWGU VJCV EJCNNGPIG VJG FGRNQ[OGPV QH EQORNGVGN[ PCVWTCN CPF WPEQPUVTCKPGF XQKEG KPVGTCEVKQPU GXGP HQT NKOKVGF VCUM FQOCKPU 6JGUG RTKOCTKN[ CTKUG DGECWUG VJG UVCVG QHVJGCTV KP CWVQOCVKE URGGEJ TGEQIPKVKQP CPF WPFGTUVCPFKPI KU HCT HTQO RGTHGEV 1PG QH VJG UKORNG URGGEJ WPFGTUVCPFKPI VCUMU VJCV JCU DGGP CVVGORVGF KU *#ŏU #KT 6TCXGN +PHQTOCVKQP 5[UVGO #6+5 +P VJKU VCUM VJG WUGT VCNMU VQ VJG OCEJKPG VQ QDVCKP ƀKIJV KPHQTOCVKQP WUKPI PCVWTCN URGGEJ UWEJ CU ő+ YQWNF NKMG VQ NGCXG 5CP (TCPEKUEQ HQT 0GY ;QTM QP &GEGODGT ſTUV RNGCUG NKUV VJG CXCKNCDNG ƀKIJVUŒ ő*QY OWEJ FQGU VJG ƀKIJV EQUV HTQO &GPXGT VQ &CNNCU!Œ 6JG *# %QOOWPKECVQT RTQLGEV KU C OWNVK[GCT OWNVKUKVG RTQLGEV NCWPEJGF KP =? 6JG CKO QH VJG RTQLGEV KU VQ EQPUVTWEV C EQORWVGT U[UVGO VJCV RNC[U VJG TQNG QH C VTCXGN CIGPV URGCMKPI D[ VGNGRJQPG YKVJ C EWUVQOGT +FGCNN[ VJKU U[UVGO YKNN RGTHQTO LWUV CU C JWOCP YQWNF EQPXGTUKPI YKVJ VJG WUGT VQ FGVGTOKPG VJG QWV NKPG QH VJG FGUKTGF KVKPGTCT[ SWGT[KPI CKTNKPG FCVCDCUGU VQ GUVCDNKUJ ƀKIJV CXCKNCDKNKV[ TGRQTVKPI UWKVCDNG ƀKIJVU VQ VJG WUGT CPUYGTKPI SWGUVKQPU VQ TGUQNXG WPEGTVCKPVKGU QT OKUWPFGTUVCPFKPIU CPF ſPCNN[ DQQMKPI VJG VTKR 6JG *# %QOOWPKECVQT FKCNQI CTEJKVGEVWTG KU JWD EGPVTKE CU UJQYP KP (KIWTG =? 6JG JWD KU C RTQITCOOCDNG VTCHſE TQWVGT VJCV KU TGURQPUKDNG HQT KPXQMKPI VJG FKHHGTGPV UGTXGTU KP VJG U[UVGO CPF TQWVKPI OGUUCIGU DGVYGGP VJGO 6JG JWD CTEJKVGEVWTG FQGU PQV FGſPG VJG HWPEVKQPCNKV[ DWV KPUVGCF RTQXKFGU UVCPFCTF #2+U 6JGTGHQTG VJG UGTXGTU FGRKEVGF KP VJG ſIWTG TGRTGUGPV C RCTVKEWNCT KPUVCPVKCVKQP QH VJG %QOOWPKECVQT CTEJKVGEVWTG 6JG UGTXGTU QRGTCVG VJTQWIJ ECNNDCEM HWPEVKQPU VJCV CTG KPXQMGF D[ VJG JWD 6JG JWD KVUGNH KU GXGPV FTKXGP WRQP TGEGKXKPI C PGY HTCOG
E\&5&3UHVV//&
*HQHUDWLRQ
776 5HQGHULQJ
2XWSXW FRQVWUXFWLRQ 7DVN PDQDJHU
'HYLFH ,)
+8% 6WDWH
$65 3RLQWLQJ
)UDPH FRQVWUXFWRU
&RQWH[W WUDFNLQJ
%DFN HQG 'LDORJXH PDQDJHU
FIGURE 6.3 AT&T Communicator architecture [15]. OGUUCIG KV ſPFU CPF KPXQMGU VJG CRRTQRTKCVG ECNNDCEM HWPEVKQPU CPF RCUUGU VJG HTCOG VQ VJG FGUVKPCVKQP UGTXGTU 6JG #66 %QOOWPKECVQT U[UVGO KU HQEWUKPI QP KUUWGU TGNCVGF VQ VJG FGUKIP QH OKZGFKPKVKCVKXG U[UVGOU 6JG KFGC QH OKZGFKPKVKCVKXG U[UVGOU KU VQ EQODKPG VJG ƀGZKDKNKV[ QH C WUGTKPKVKCVKXG U[UVGO YKVJ VJG EQPUVTCKPGF RTQDNGOUQNXKPI PCVWTG QH C U[UVGOKPKVKCVKXG U[UVGO (QT GZCORNG C TGCUQPCDNG TGURQPUG VQ VJG SWGT[ ő5JQY OG VJG ƀKIJVUŒ EQWNF DG ő2NGCUG VGNN OG YJGTG [QW YQWNF NKMG VQ ƀ[Œ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ſE VGTOU URQMGP KUQNCVKQP QT RQUUKDN[ GODGFFGF KP C PCVWTCN WVVGTCPEG 9KVJ VJG CUUWORVKQP VJCV CEVKQPU CTG NKMGN[ VQ DG GZRTGUUGF KP MG[YQTFU VJG U[UVGO ECP LWUV GORNQ[ MG[YQTFURQVVKPI VGEJPKSWGU VQ RGTHQTO VJG VCUM 6JKU MKPF QH U[UVGO KU UKORNG VQ KORNGOGPV
E\&5&3UHVV//&
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
EQPVKPWQWUN[ XCNWGF UKIPKſECPEG NGXGN DGVYGGP VJG CRRGCTCPEG QH C YQTF CPF VJG KPVGPFGF CEVKQP
6.4 Japanese National Project on Spontaneous Speech Corpus and Processing Technology 6.4.1 Project Overview (QT DWKNFKPI NCPIWCIG OQFGNU HQT URQPVCPGQWU URGGEJ NCTIG URQPVCPGQWU URGGEJ EQT RQTC CTG KPFKURGPUCDNG +P VJKU EQPVGZV C 5EKGPEG CPF 6GEJPQNQI[ #IGPE[ 2TKQTKV[ 2TQITCO GPVKVNGF ő5RQPVCPGQWU 5RGGEJ %QTRWU CPF 2TQEGUUKPI 6GEJPQNQI[Œ UVCTVGF KP ,CRCP KP =? 6JG RTQLGEV YKNN DG EQPFWEVGF QXGT C ſXG[GCT RGTKQF WPFGT VJG HQNNQYKPI VJTGG OCLQT VJGOGU CU UJQYP KP (KIWTG $WKNFKPI C NCTIGUECNG URQPVCPGQWU URGGEJ EQTRWU %QTRWU QH 5VCPFCTF ,CRCPGUG
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
E\&5&3UHVV//&
/DUJHVFDOH VSRQWDQHRXV VSHHFKFRUSXV
6SRQWDQHRXV VSHHFK
:RUOGNQRZOHGJH /LQJXLVWLFLQIRUPDWLRQ 3DUDOLQJXLVWLFLQIRUPDWLRQ 'LVFRXUVHLQIRUPDWLRQ
6SHHFK UHFRJQLWLRQ
7UDQ VFULSWLRQ
8QGHUVWDQGLQJ ,QIRUPDWLRQ H[WUDFWLRQ VXPPDUL]DWLRQ
6XPPDUL]HG WH[W .H\ZRUGV 6\QWKHVL]HG YRLFH
FIGURE 6.4 Overview of the Japanese national project on spontaneous speech corpus and processing technology.
6.4.2 Corpus %5, KU VJG EQTRWU QH URQPVCPGQWU OQPQNQIWG QH UVCPFCTF ,CRCPGUG =? /QTG RTG EKUGN[ %5, EQPVCKPU URGGEJ WVVGTGF VQ OWNVKRNG NKUVGPGTU KP C OQTG QT NGUU HQTOCN UQEKCN UGVVKPI 6JG VYQ OCKP UQWTEGU QH URQPVCPGQWU OQPQNQIWG HQT %5, CTG NKXG TGEQTFKPI QH XCTKQWU CECFGOKE EQPHGTGPEGUOGGVKPIU TGHGTTGF VQ CU #ECFGOKE 2TG UGPVCVKQP #2 JGTGCHVGT UWEJ CU VJG #EQWUVKECN 5QEKGV[ QH ,CRCP #5, OGGVKPIU CPF UVWFKQ TGEQTFKPI QH KPHQTOCN HTGG RWDNKE URGGEJ OCFG D[ RCKF XQNWPVCT[ UWDLGEVU 5KOWNCVGF 2WDNKE 5RGGEJ 525 6JG 525 KPENWFGU C YKFG XCTKGV[ QH VQRKEU KPENWFKPI VJG UWDLGEVUŏ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ſGNFURGEKſ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
E\&5&3UHVV//&
)RUWUDLQLQJ DPRUSKRORJLFDO DQDO\VLVDQG326 WDJJLQJSURJUDP
)RUVSHHFK UHFRJQLWLRQ
&66SRQWDQHRXV PRQRORJXH &RUH
0DQXDOO\WDJJHG ZLWKVHJPHQWDO DQGSURVRGLF LQIRUPDWLRQ
'LJLWL]HG VSHHFK WUDQVFULSWLRQ 326DQG VSHDNHU LQIRUPDWLRQ
NZRUGV
0ZRUGV
FIGURE 6.5 Overall design of the Corpus of Spontaneous Japanese.
6.5 Automatic Transcription of Spontaneous Presentation 6.5.1 Recognition Task 7UKPI VJG %5, EQTRWU RTGNKOKPCT[ TGEQIPKVKQP GZRGTKOGPVU CTG DGKPI EQPFWEVGF CV 6QM[Q +PUVKVWVG QH 6GEJPQNQI[ CU YGNN CU CV UGXGTCN QVJGT WPKXGTUKVKGU RCTVKEKRCVKPI KP VJG RTQLGEV +P VJKU GZRGTKOGPV RTGUGPVCVKQP URGGEJ WVVGTGF D[ OCNG URGCMGTU YCU WUGF CU C VGUV UGV QH URGGEJ TGEQIPKVKQP =? 6CDNG UJQYU CP QWVNKPG QH VJG VGUV UGV
6.5.2 Language and Acoustic Modeling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eb EQTRWU: 6TCPUETKDGF RTGUGPVCVKQPU EQPUKUVKPI QH CRRTQZKOCVGN[ M UGPVGPEGU YKVJ / YQTFU JCXG DGGP EQNNGEVGF HTQO VJG 9QTNF 9KFG 9GD 5RQPVCPGQWU
E\&5&3UHVV//&
TABLE 6.2
4GEQIPKVKQP VGUV UGV QH RTGUGPVCVKQPU +& # # # 2 , - 0 5 ; ;
%QPHGTGPEG PCOG #EQWUV 5QE ,CR #EQWUV 5QE ,CR #EQWUV 5QE ,CR 2JQPGVKEU 5QE ,CR 5QE ,CR .KPIWKUVKEU 0CVKQPCN .CPI 4GU +PUV #UUQE 0CVWTCN .CPI 2TQE #UUQE 5QEKQNKPI 5EKGPEGU 5RQPV 5RGGEJ %QTRWU /GGVKPI 5RQPV 5RGGEJ %QTRWU /GGVKPI
.GPIVJOKP
URGGEJ WUWCNN[ KPENWFGU XCTKQWU ſNNGF RCWUGU DWV VJG[ CTG PQV KPENWFGF KP VJKU RTGUGPVCVKQP EQTRWU #P GHHQTV KU VJWU OCFG VQ CFF ſNNGF RCWUGU VQ VJG RTGUGP VCVKQP EQTRWU DCUGF QP VJG UVCVKUVKECN EJCTCEVGTKUVKEU QH VJG ſNNGF RCWUGU 6JG VQRKEU QH VJG RTGUGPVCVKQPU EQXGT YKFG FQOCKPU KPENWFKPI UQEKCN KUUWGU CPF OGOQKTU 6JG HQNNQYKPI VYQ NCPIWCIG OQFGNU FGPQVGF CU SpnL CPF WebL JCXG DGGP EQP UVTWEVGF 'CEJ OQFGN EQPUKUVU QH DKITCOU CPF TGXGTUG VTKITCOU YKVJ DCEMKPIQHH 6JGKT XQECDWNCT[ UK\GU CTG M YQTFU SpnL: /CFG WUKPI VJG RTGUGPVCVKQPU KP VJG %5, 6JG URGCMGTU JCXG PQ QXGTNCR YKVJ VJQUG QH VJG VGUV UGV 5KPEG VJGTG CTG PQ RWPEVWCVKQP OCTMU KP VJG VTCP UETKRVKQP EQOOCU CTG KPUGTVGF YJGP C UKNGPEG RGTKQF QH OU QT NQPIGT KU GPEQWPVGTGF WebL: /CFG WUKPI VJG VGZV QH QWT 9GD EQTRWU 6JG HQNNQYKPI VYQ VKGFUVCVG VTKRJQPG *//U JCXG DGGP OCFG DQVJ JCXKPI M UVCVGU CPF )CWUUKCP OKZVWTGU KP GCEJ UVCVG SpnA: 7UKPI RTGUGPVCVKQPU KP VJG %5, WVVGTGF D[ OCNG URGCMGTU CRRTQZKOCVGN[ JQWTU 6JG URGCMGTU JCXG PQ QXGTNCR YKVJ VJQUG KP VJG VGUV UGV RdA: 7UKPI CRRTQZKOCVGN[ JQWTU QH TGCF URGGEJ WVVGTGF D[ OCP[ URGCMGTU
6.5.3 Recognition Results (KIWTG RTGUGPVU VJG VGUVUGV RGTRNGZKV[ QH VTKITCOU CPF VJG QWVQHXQECDWNCT[
118 TCVG HQT GCEJ RTGUGPVCVKQP EQORCTKPI VJG VYQ NCPIWCIG OQFGNU 6JG RGT RNGZKV[ CPF 118 QH SpnL OCFG HTQO VJG %5, CTG ENGCTN[ DGVVGT VJCP VJCV QH VJG YGDDCUGF OQFGN WebL WebL UJQYU JKIJ RGTRNGZKV[ CPF 118 TCVG UKPEG KV YCU GFKVGF CU C VGZV CPF VJGKT VQRKEU CTG OWEJ OQTG FKXGTUKſGF VJCP VJQUG QH VJG VGUV UGV
E\&5&3UHVV//&
FIGURE 6.6 Test-set perplexity and OOV rate for the two language models. (KIWTG UJQYU TGEQIPKVKQP TGUWNVU HQT VJG EQODKPCVKQPU QH VJG VYQ NCPIWCIG OQF GNU SpnL CPF WebL CPF VJG VYQ CEQWUVKE OQFGNU SpnA CPF RdA (KNNGTU CTG EQWPVGF CU YQTFU CPF KPENWFGF KP ECNEWNCVKPI VJG CEEWTCE[ +V KU ENGCTN[ UJQYP VJCV SpnL CEJKGXGU OWEJ DGVVGT TGUWNVU VJCP WebL CPF SpnA IKXGU OWEJ DGVVGT TGUWNVU VJCP RdA 6JGUG TGUWNVU KPFKECVG VJCV KV KU ETWEKCN VQ OCMG NCPIWCIG OQFGNU HTQO C URQPVCPGQWU URGGEJ EQTRWU VQ CFGSWCVGN[ TGEQIPK\G URQPVCPGQWU URGGEJ +V KU CNUQ UWIIGUVGF VJCV CEQWUVKE OQFGNU OCFG HTQO %5, JCXG DGVVGT EQXGTCIG QH VTKRJQPGU CPF DGVVGT OCVEJKPI QH CEQWUVKE EJCTCEVGTKUVKEU EQTTGURQPFKPI VQ VJG URGCMKPI UV[NG CPF CNUQ JCXG DGVVGT OCVEJKPI QH TGEQTFKPI EQPFKVKQPU YKVJ VJG VGUV UGV 6JG OGCP CEEWTCE[ HQT VJG EQODKPCVKQP QH SpnL CPF Spn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ſFGPEG OGCUWTG KU CRRNKGF 6JG HQNNQYKPI UVGRU CTG RGTHQTOGF
E\&5&3UHVV//&
FIGURE 6.7 Word accuracy for each combination of models. /CMKPI C TGITGUUKQP ENCUU VTGG JCXKPI NGCH PQFGU HQT VJG SpnA RJQPG OQFGN 4GEQIPK\KPI VJG VGUVUGV WVVGTCPEGU WUKPI VJG SpnA CU C URGCMGT KPFGRGPFGPV OQFGN #RRN[KPI VJG /..4 CFCRVCVKQP DCUGF QP VJG TGEQIPKVKQP TGUWNV HQT GCEJ WVVGT CPEG VQ OCMG C URGCMGT CFCRVKXG OQFGN 4GTGEQIPK\KPI VJG VGUVUGV WVVGTCPEGU WUKPI VJG URGCMGT CFCRVKXG OQFGN +VGTCVKPI VJG CFCRVCVKQP RTQEGUU WUKPI VJG TGUWNVKPI VTCPUETKRVKQP (KIWTG RTGUGPVU VJG GHHGEV QH VJG CFCRVCVKQP YJGP SpnL KU WUGF CU VJG NCPIWCIG OQFGN ő5+Œ KPFKECVGU VJG DCUGNKPG EQPFKVKQP WUKPI VJG URGCMGT KPFGRGPFGPV RJQPG OQFGN SpnA ő5#Œ KPFKECVGU VJG TGUWNV CHVGT KVGTCVKQPU QH VJG /..4 CFCRVCVKQP 6JG UKPING UVGR QH /..4 KORTQXGU YQTF CEEWTCE[ D[ CP CDUQNWVG VQ CPF VJG UGEQPF CFCRVCVKQP UVGR HWTVJGT KORTQXGU CEEWTCE[ D[ QP CXGTCIG 6JG KO RTQXGOGPV CNOQUV UCVWTCVGU CV VJG VJKTF KVGTCVKQP CPF VJG OGCP YQTF CEEWTCE[ CHVGT VJG VJKTF KVGTCVKQP KU $[ CRRN[KPI VYQ QT VJTGG UVGRU QH /..4 CFCRVCVKQP VJG GTTQT TCVG KU TGFWEGF D[ TGNCVKXG VQ VJG URGCMGT KPFGRGPFGPV ECUG
6.5.4 Analysis on Individual Differences +PFKXKFWCN FKHHGTGPEGU KP URQPVCPGQWU RTGUGPVCVKQP URGGEJ TGEQIPKVKQP RGTHQTOCPEGU JCXG DGGP CPCN[\GF WUKPI OKPWVGU HTQO GCEJ RTGUGPVCVKQP IKXGP D[ OCNG URGCM
E\&5&3UHVV//&
FIGURE 6.8 Results of unsupervised adaptation.
GTU HQT C VQVCN QH OKPWVGU 6JG URGCMGTU JCXG PQ QXGTNCR YKVJ VJQUG KP VJG VTCKP KPI UGV 6JG OGCP YQTF CEEWTCE[ HQT VJG URGCMGTU KU CPF HQT VJG 5+
URGCMGTKPFGRGPFGPV CPF 5# URGCMGTCFCRVKXG EQPFKVKQPU TGURGEVKXGN[ 6JG UVCP FCTF FGXKCVKQP KU HQT VJG 5+ CPF HQT VJG 5# EQPFKVKQP #U UJQYP D[ VJG UVCPFCTF FGXKCVKQP TGEQIPKVKQP CEEWTCE[ NCTIGN[ XCTKGU HTQO URGCMGT VQ URGCMGT %QT TGNCVKQP CPF TGITGUUKQP CPCN[UGU JCXG DGGP CRRNKGF VQ VJG YQTF TGEQIPKVKQP CEEWTCE[ CPF XCTKQWU URGCMGT CVVTKDWVGU 6.5.4.1 Speaker Attributes 5GXGP MKPFU QH URGCMGT CVVTKDWVGU JCXG DGGP EQPUKFGTGF KP VJG CPCN[UKU 6JG[ CTG YQTF CEEWTCE[ #EE CXGTCIGF CEQWUVKE HTCOG NKMGNKJQQF #. URGCMKPI TCVG 54 YQTF RGTRNGZKV[ 22 QWV QH XQECDWNCT[ TCVG 14 ſNNGF RCWUG TCVG (4 CPF TGRCKT TCVG 44 6JG URGCMKPI TCVG FGſPGF CU VJG PWODGT QH RJQPGOGU RGT UGEQPF CPF VJG CXGTCIGF CEQWUVKE HTCOG NKMGNKJQQF CTG ECNEWNCVGF WUKPI VJG TGUWNV QH HQTEGF CNKIPOGPV QH VJG TGHGTGPEG VTKRJQPG NCDGNU CHVGT TGOQXKPI RCWUG RGTKQFU 6JG YQTF RGTRNGZKV[ KU ECNEWNCVGF WUKPI VTKITCOU KP YJKEJ RTGFKEVKQP QH QWV QH XQECDWNCT[ YQTFU KU PQV KPENWFGF 6JG ſNNGF RCWUG TCVG CPF VJG TGRCKT TCVG CTG VJG PWODGT QH ſNNGF RCWUGU CPF TGRCKTU FKXKFGF D[ VJG PWODGT QH YQTFU TGURGEVKXGN[ 6CI KPHQTOCVKQP KPENWFGF KP VJG %5, VTCPUETKRVKQP KU WUGF VQ FGVGTOKPG YJGVJGT C YQTF KU C ſNNGF RCWUGTGRCKT QT PQV +P VJG %5, TGRCKTU CTG FGſPGF QPN[ HQT YQTF HTCIOGPVU CPF C TGRJTCUGF YJQNG YQTF KU PQV OCTMGF CU C TGRCKT 6JG ECNEWNCVKQPU QH YQTF CEEWTCE[ QWV QH XQECDWNCT[ TCVG CPF YQTF RGTRNGZKV[ CTG DCUGF QP VJG TGHGTGPEG VGZV CHVGT GZENWFKPI TGRCKTU
E\&5&3UHVV//&
TABLE 6.3
%QTTGNCVKQP EQGHſEKGPV OCVTKZ VJG NQYGT VTKCPIWNCT OCVTKZ UJQYU VJG EQTTGNCVKQP EQ GHſEKGPVU CPF VJG WRRGT VTKCPIWNCT OCVTKZ UJQYU VJG XCNWG VJCV KU VJG UKIPKſECPEG NGXGN $QNF HCEG KPFKECVGU C UKIPKſECPV XCNWG YKVJ VJG UKIPKſECPEG NGXGN QH #EE 5+ #EE 5+ #EE 5# #. 5+ #. 5# 54 22 14 (4 44
#EE 5#
#. 5+
Ō
Ō
Ō 0.28 Ō -0.42 -0.40 -0.54 0.38 -0.30
Ō 0.32 -0.47 -0.33 -0.51 0.38 -0.31
Ō -0.54
#. 5#
54
22
14
(4
6.5.4.2 Correlation Analysis 6CDNG UJQYU VJG EQTTGNCVKQP OCVTKZ QH URGCMGT CVVTKDWVGU +P VJG VCDNG VJG NQYGT VTKCPIWNCT OCVTKZ UJQYU VJG EQTTGNCVKQP EQGHſEKGPVU CPF VJG WRRGT VTKCPIWNCT OCVTKZ UJQYU VJG QDUGTXGF UKIPKſECPEG NGXGNU XCNWGU 6JG EQTTGNCVKQP EQGHſEKGPVU YTKV VGP KP DQNF HCEG KPFKECVG UKIPKſECPV XCNWGU CV UKIPKſECPEG NGXGN XCNWGU %QTTGNCVKQP DGVYGGP CEQWUVKE NKMGNKJQQF CPF URGCMKPI TCVG 6JG EQTTGNCVKQP EQGHſEKGPVU DGVYGGP CEQWUVKE NKMGNKJQQF CPF URGCMKPI TCVG CTG CPF HQT VJG 5+ CPF 5# CEQWUVKE OQFGN TGURGEVKXGN[ 6JGTG KU C VGPFGPE[ VJCV VJG JKIJGT VJG URGCMKPI TCVG KU VJG NQYGT VJG CEQWUVKE NKMGNKJQQF DGEQOGU 6JG #MCKMG +PHQTOCVKQP %TKVGTKQP #+% =? KPFKECVGU VJCV VJG ſTUV QTFGT TGITGUUKQP OQFGN KU DGVVGT VJCP VJG UGEQPF QTFGT OQFGN HQT TGITGUUKPI VJG CEQWUVKE NKMGNKJQQF QP VJG URGCMKPI TCVG 6JKU KPFKECVGU VJCV VJGTG KU C NKPGCT TGNCVKQPUJKR DGVYGGP VJG URGCMKPI TCVG CPF VJG CEQWUVKE NKMGNKJQQF CXGTCIGF QXGT RTGUGPVCVKQPU # UVTQPIGT CTVKEWNCVKQP GHHGEV KP HCUVGT URGCMGTU KU RTQDCDN[ C ECWUG QH VJG FGETGCUG QH NKMGNKJQQF 6JG WPUWRGTXKUGF CFCRVCVKQP KPETGCUGU VJG CEQWUVKE NKMGNKJQQF DWV RTGUGTXGU VJG TGNCVKQPUJKR DGVYGGP VJG URGCMKPI TCVG CPF VJG CEQWUVKE NKMGNKJQQF YKVJ C UNKIJV KPETGCUG KP VJG EQTTGNCVKQP EQGHſEKGPV %QTTGNCVKQP DGVYGGP YQTF RGTRNGZKV[ CPF UGXGTCN NKPIWKUVKE CVVTKDWVGU 6JGTG GZKUVU UKIPKſECPV EQTTGNCVKQP DGVYGGP VJG YQTF RGTRNGZKV[ CPF VJG QWV QH XQECDWNCT[ TCVG YKVJ C EQTTGNCVKQP EQGHſEKGPV QH 6JGTG KU C VGPFGPE[ VJCV RTGUGPVCVKQPU JCXKPI C JKIJGT QWV QH XQECDWNCT[ TCVG UJQY C JKIJGT RGTRNGZKV[ 6JG EQTTGNCVKQP EQGHſEKGPV QH VJG ſNNGF RCWUG HTGSWGPE[ CPF VJG RGTRNGZKV[ KU KPFKECVKPI VJCV VJG[ CTG TCVJGT WPEQTTGNCVGF 6JG TGRCKT HTGSWGPE[ CPF VJG RGTRNGZKV[ JCXG C EQTTGNCVKQP EQGHſEKGPV QH 5KPEG VJG RGTRNGZKV[ YCU ECNEWNCVGF CHVGT TGOQXKPI TGRCKTU VJKU TGUWNV UJQYU VJCV VJG NKPIWKUVKE FKHſEWNV[ GZENWFKPI TGRCKTU JCU CNOQUV PQ EQTTGNCVKQP YKVJ VJG TGRCKT TCVG
E\&5&3UHVV//&
44
Ō Ō -0.62 0.33 0.52 -0.50 -0.41
%QTTGNCVKQP DGVYGGP YQTF CEEWTCE[ CPF UGXGTCN CVVTKDWVGU 6JG EQTTGNCVKQP EQGHſEKGPV DGVYGGP VJG YQTF CEEWTCE[ 5+ CPF VJG URGCMKPI TCVG KU (KIWTG UJQYU VJG TGNCVKQPUJKR DGVYGGP VJG YQTF CEEWTCE[ CPF VJG URGCMKPI TCVG 6JG TGNCVKQPUJKR UGGOU OQPQVQPKE CPF GXGP XGT[ UNQY URGCMKPI TCVG FQGU PQV FGETGCUG VJG CEEWTCE[ YJKEJ KU UKOKNCT VQ VJG TGUWNV HQT VJG CEQWUVKE NKMGNKJQQF 6JG #+% CNUQ KPFKECVGU VJCV VJG ſTUV QTFGT OQFGN KU UWRGTKQT VQ VJG UGEQPF QTFGT OQFGN HQT TGITGUUKPI VJG YQTF CEEWTCE[ QP VJG URGCMKPI TCVG 6JG EQTTGNCVKQP DGVYGGP VJG YQTF CEEWTCE[ 5+ CPF VJG CEQWUVKE NKMGNKJQQF KU +P QTFGT VQ CPCN[\G VJG TGCN EQTTGNCVKQP RCTVKCN EQTTGNCVKQP KU ECNEW NCVGF 6JG TGUWNVCPV EQTTGNCVKQP EQGHſEKGPV CFLWUVGF HQT VJG URGCMKPI TCVG KU YJKEJ OGCPU VJCV VJG EQTTGNCVKQP KU PQV UVCVKUVKECNN[ UKIPKſECPV +P QVJGT YQTFU VJG EQTTGNCVKQP DGVYGGP VJG YQTF CEEWTCE[ CPF VJG CEQWUVKE NKMGNKJQQF KU URWTKQWU 1P VJG QVJGT JCPF RCTVKCN EQTTGNCVKQP EQGHſEKGPV DGVYGGP VJG YQTF CEEWTCE[ CPF VJG URGCMKPI TCVG CFLWUVGF HQT VJG CEQWUVKE NKMGNKJQQF KU YJKEJ KU UKIPKſECPV CV C UKIPKſECPEG NGXGN CPF RCTVKCN EQTTGNCVKQP EQGHſ EKGPV DGVYGGP VJG CEQWUVKE NKMGNKJQQF CPF VJG URGCMKPI TCVG CFLWUVGF HQT VJG YQTF CEEWTCE[ KU YJKEJ KU UKIPKſECPV CV C UKIPKſECPEG NGXGN 5KOKNCT TGUWNVU CTG QDVCKPGF HQT VJG 5# EQPFKVKQPU 6JG EQTTGNCVKQP EQGHſEKGPV DGVYGGP VJG YQTF CEEWTCE[ CPF VJG TGRCKT HTGSWGPE[ KU 6JGTG KU C YGCM RQUKVKXG EQTTGNCVKQP QH DGVYGGP VJG YQTF CEEW TCE[ CPF VJG ſNNGF RCWUG HTGSWGPE[ DWV VJKU KU CNUQ C URWTKQWU EQTTGNCVKQP UKPEG RCTVKCN EQTTGNCVKQP EQGHſEKGPV CFLWUVGF HQT VJG URGCMKPI TCVG KU 6JG EQTTGNCVKQP EQGHſEKGPV DGVYGGP VJG YQTF CEEWTCE[ CPF VJG QWV QH XQECD WNCT[ TCVG KU 6JGTG KU C PGICVKXG EQTTGNCVKQP QH DGVYGGP VJG YQTF CEEWTCE[ 5+ CPF VJG RGTRNGZKV[ DWV VJKU KU CNUQ URWTKQWU VJG RCTVKCN EQTTGNCVKQP DGVYGGP VJG YQTF CEEWTCE[ CPF VJG RGTRNGZKV[ CFLWUVGF HQT VJG QWV QH XQECDW NCT[ TCVG KU (KIWTG UJQYU VJG UWOOCT[ QH EQTTGNCVKQP DGVYGGP CNN VJG CPCN[\GF CVVTKDWVGU 6.5.4.3 Regression Analysis 6JG HQNNQYKPI GSWCVKQPU CPF UJQY NKPGCT TGITGUUKQP OQFGNU QH VJG YQTF CEEWTCE[ YKVJ VJG UKZ RTGUGPVCVKQP CVVTKDWVGU YJGP 5+ CPF 5# CEQWUVKE OQFGNU CTG TGURGEVKXGN[ WUGF HQT URGGEJ TGEQIPKVKQP
+P VJG GSWCVKQP TGITGUUKQP EQGHſEKGPV HQT VJG TGRCKT TCVG KU CPF VJG EQGHſ EKGPV HQT VJG QWV QH XQECDWNCT[ TCVG KU 6JKU OGCPU VJCV KPETGCUG QH VJG TGRCKT TCVG QT VJG QWV QH XQECDWNCT[ TCVG TGURGEVKXGN[ EQTTGURQPFU VQ QT FGETGCUG
E\&5&3UHVV//&
FIGURE 6.9 Speaking rate vs. word accuracy.
QH VJG YQTF CEEWTCE[ 6JKU KU RTQDCDN[ DGECWUG C UKPING TGEQIPKVKQP GTTQT ECWUGF D[ C TGRCKT QT CP QWV QH XQECDWNCT[ YQTF VTKIIGTU UGEQPFCT[ GTTQTU FWG VQ VJG NKPIWKUVKE EQPUVTCKPVU 6JG FGVGTOKPCVKQP EQGHſEKGPVU QH VJG OWNVKRNG NKPGCT TGITGUUKQPU CPF CTG CPF TGURGEVKXGN[ DQVJ QH YJKEJ CTG UKIPKſECPV CV NGXGN 6JKU OGCPU VJCV CDQWV JCNH QH VJG XCTKCPEG QH VJG YQTF CEEWTCE[ ECP DG GZRNCKPGF D[ VJG OQFGN 6CDNG UJQYU PQTOCNK\GF TGRTGUGPVCVKQP QH VJG TGITGUUKQP CPCN[UKU YKVJ VJG GSWC VKQPU CPF KP YJKEJ VJG XCTKCDNGU CTG PQTOCNK\GF KP VGTOU QH VJG OGCP CPF XCTKCPEG DGHQTG VJG CPCN[UKU KP QTFGT VQ UJQY VJG GHHGEVU QH GZRNCKPKPI XCTKCDNGU QP VJG YQTF CEEWTCE[ 6JG VCDNG UJQYU VJG PQTOCNK\GF TGITGUUKQP EQGHſEKGPV VJG XCNWG CPF VJG EQPſFGPEG KPVGTXCN 6JG PQTOCNK\GF TGITGUUKQP EQGHſEKGPVU QH VJG URGCMKPI TCVG VJG QWV QH XQECDWNCT[ TCVG CPF VJG TGRCKT TCVG CTG TGNCVKXGN[ NCTIG GURGEKCNN[ YJGP 5# CEQWUVKE OQFGN KU WUGF 6JG CEQWUVKE NKMGNKJQQF JCU TGNCVKXGN[ C UOCNN EQGHſEKGPV KP DQVJ VJG 5+ CPF 5# TGITGUUKQP OQFGNU 6JKU OGCPU VJCV CN VJQWIJ VJG CEQWUVKE NKMGNKJQQF JCU UKIPKſECPV EQTTGNCVKQP YKVJ VJG YQTF CEEWTCE[ KV KU URWTKQWU CU KPFKECVGF KP VJG RTGXKQWU UWDUGEVKQP 6.5.4.4 Selection of Major Attributes #U C UWRRNGOGPVCT[ GZRGTKOGPV C DCEMYCTF GNKOKPCVKQP RTQEGFWTG JCU DGGP GO RNQ[GF VQ KFGPVKH[ TGNCVKXGN[ KORQTVCPV RTGFKEVQTU QH VJG YQTF CEEWTCE[ # DCEMYCTF GNKOKPCVKQP RTQEGUU DGIKPU YKVJ CNN QH VJG UKZ RTGFKEVQTU KP VJG OQFGN CPF VJG OQFGN KU TGſVVGF VQ VJG FCVC CHVGT TGOQXKPI C XCTKCDNG YKVJ VJG NCTIGUV XCNWG 6JG TGſV VKPI RTQEGUU KU KVGTCVGF TGOQXKPI VJG NGCUV UKIPKſECPV XCTKCDNG KP VJG OQFGN WPVKN CNN TGOCKPKPI XCTKCDNGU JCXG XCNWGU UOCNNGT VJCP 6JG KORQTVCPV RTGFKEVQTU KFGP VKſGF CTG VJG URGCMKPI TCVG VJG QWV QH XQECDWNCT[ TCVG CPF VJG TGRCKT TCVG YJKEJ
E\&5&3UHVV//&
14 44
22 #EE
(4
#. 54 %QTTGNCVKQP 5RWTKQWUEQTTGNCVKQP
FIGURE 6.10 Summary of correlation between various attributes. TABLE 6.4 4GUWNVU QH UVCPFCTFK\GF TGITGUUKQP CPCN[UKU HQT YQTF CEEWTCE[ UJQYKPI UVCPFCTFK\GF TGITGU UKQP EQGHſEKGPV %QGHH XCNWG CPF EQPſFGPEG KPVGTXCN %+ #. 5+ 54 5+ 22 14 (4 44
%QGHH 5+
2
%+
#. 5# 54 5+ 22 14 (4 44
%QGHH 5#
2
%+
EQTTGURQPF VQ VJG CVVTKDWVGU UJQYKPI TGNCVKXGN[ NCTIG EQGHſEKGPVU KP 6CDNG 6JG FGVGTOKPCVKQP EQGHſEKGPVU QH VJG TGITGUUKQP OQFGNU QP VJGUG VJTGG CVVTKDWVGU CTG HQT DQVJ URGCMGT KPFGRGPFGPV CPF CFCRVKXG ECUGU 6JKU XCNWG KU CNOQUV VJG UCOG CU VJCV QH VJG OQFGNU QP CNN CVVTKDWVGU +V ECP DG EQPENWFGF VJCV VJG OCKP HCEVQTU QH KPFK XKFWCN FKHHGTGPEGU QH VJG YQTF CEEWTCE[ CTG VJG URGCMKPI TCVG VJG QWV QH XQECDWNCT[ TCVG CPF VJG TGRCKT TCVG
6.5.5 Discussion 2TGNKOKPCT[ TGEQIPKVKQP GZRGTKOGPVU JCXG DGGP RGTHQTOGF WUKPI VGP URGCMGTUŏ RTG UGPVCVKQP WVVGTCPEGU QH CRRTQZKOCVGN[ JQWTU .CPIWCIG OQFGNU DCUGF QP C URQP VCPGQWU URGGEJ EQTRWU CPF 9GD EQTRWU YGTG EQORCTGF KP VGTOU QH VGUVUGV RGTRNGZ KV[ 118 TCVG CPF YQTF OQTRJGOG CEEWTCE[ 6YQ CEQWUVKE OQFGNU OCFG D[ WUKPI URQPVCPGQWU URGGEJ CPF TGCF URGGEJ YGTG CNUQ EQORCTGF $QVJ EQORCTKUQPU UJQYGF VJCV CEQWUVKE CPF NCPIWCIG OQFGNKPI DCUGF QP CP CEVWCN URQPVCPGQWU URGGEJ EQTRWU KU HCT OQTG GHHGEVKXG VJCP EQPXGPVKQPCN OQFGNKPI DCUGF QP TGCF URGGEJ +V YCU EQP
E\&5&3UHVV//&
ſ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ſECPV VQ [KGNF KPFKXKFWCN FKHHGT GPEGU KP VJG YQTF CEEWTCE[ 6JG CXGTCIGF CEQWUVKE NKMGNKJQQF QH TGHGTGPEG RJQPGOG UGSWGPEGU CPF VJG VGUV UGV RGTRNGZKV[ YGTG HQWPF VQ DG TGNCVKXGN[ OKPQT HCEVQTU QH KP FKXKFWCN FKHHGTGPEGU KP VJG YQTF CEEWTCE[ 7PUWRGTXKUGF /..4 URGCMGT CFCRVCVKQP FQGU PQV EJCPIG VJG UVTWEVWTG QH VJG KPFKXKFWCN FKHHGTGPEGU #RRTQZKOCVGN[ JCNH QH VJG XCTKCPEG KP VJG YQTF CEEWTCE[ YCU GZRNCKPGF D[ C TGITGUUKQP OQFGN WUKPI VJQUG VJTGG OCLQT CVVTKDWVGU (WVWTG TGUGCTEJ KPENWFGU VJG KPXGUVKICVKQP QH GHſEKGPV OGVJQFU HQT TGFWEKPI VJG GHHGEVU QH VJG OCLQT CVVTKDWVGU QP VJG TGEQIPKVKQP CEEWTCE[ 6Q EQRG YKVJ VJG URGCMKPI TCVG RTQDNGO C OGVJQF WUKPI UGRCTCVG CEQWUVKE OQFGNU HQT GCEJ URGCMKPI TCVG =? CPF CPQVJGT OGVJQF YJKEJ VCMGU KPVQ CEEQWPV VJG URGCMKPI TCVG KP VJG VTGGDCUGF *// UVCVG ENWUVGTKPI JCXG DGGP RTQRQUGF =? 5KPEG VJG TGEQIPKVKQP CEEWTCE[ HQT URQPVCPGQWU URGGEJ KU UVKNN TCVJGT NQY KV KU KORGT CVKXG VQ EQPVKPWG VJG EQNNGEVKQP QH C NCTIG EQTRWU QH URQPVCPGQWU URGGEJ CPF WUG KV HQT DWKNFKPI NCPIWCIG CPF CEQWUVKE OQFGNU (WVWTG TGUGCTEJ KUUWGU KPENWFG C JQY VQ VTCPUETKDG CPF CPPQVCVG URQPVCPGQWU URGGEJ D JQY VQ CRRN[ OQTRJQNQIKECN CPCN[UKU VQ VJG VTCPUETKDGF URQPVCPGQWU URGGEJ E JQY VQ DWKNF RTGEKUG CPF [GV IGPGTCN ſNNGF RCWUG OQFGNU F JQY VQ KPEQTRQTCVG TGRCKTU JGUKVCVKQPU TGRGVKVKQPU RCTVKCN YQTFU CPF FKUƀWGPEKGU G JQY VQ CFCRV VJG NCPIWCIG OQFGNU VQ GCEJ VCUM H JQY VQ CFCRV VQ URGCMKPI UV[NGU CPF VQRKEU QH RTGUGPVCVKQPU CPF I JQY VQ DWKNF CEQWUVKE OQFGNU VJCV ſV URQPVCPGQWU URGGEJ 5GIOGPVCVKQP QH URQPVCPGQWU WVVGTCPEGU KPVQ UGPVGPEGU KU QPG QH VJG KORQTVCPV KU UWGU 6JG 8KVGTDK FGEQFKPI CNIQTKVJO WUWCNN[ WUGF KP URGGEJ TGEQIPKVKQP FGVGTOKPGU C TGEQIPKVKQP J[RQVJGUKU QPN[ CHVGT FGVGEVKPI VJG GPF QH VJG KPRWV WVVGTCPEG +P CF FKVKQP VJG OWNVKRNGRCUU UGCTEJ CNIQTKVJO YKFGN[ WUGF KP .8%54 CNYC[U PGGFU VQ KPVGTTWRV VJG KPRWV CV UQOG TGCUQPCDNG RQUKVKQPU *QYGXGT KP URQPVCPGQWU URGGEJ WVVGTCPEGU CTG PQV UGRCTCVGF UGPVGPEG D[ UGPVGPEG +PUVGCF NQPI RCWUGU CTG UQOG VKOGU KPUGTVGF KP C UGPVGPEG 1P VJG QVJGT JCPF OWNVKRNG UGPVGPEGU CTG UQOGVKOGU WVVGTGF EQPVKPWQWUN[ YKVJQWV KPUGTVKPI ENGCT RCWUGU 6JGTGHQTG KV KU PGEGUUCT[ VQ UWEEGUUKXGN[ FGVGTOKPG TGEQIPKVKQP TGUWNVU DGHQTG FGVGEVKPI UGPVGPEG DQWPFCTKGU =? QT KPJGTKV C YQTF JKUVQT[ HQT NKPIWKUVKE NKMGNKJQQF ECNEWNCVKQP VQ VJG PGZV UGPVGPEG J[ RQVJGUKU =? +P VJG UWRRQTVKPI U[UVGOU HQT OCMKPI RTGUGPVCVKQP TGEQTFU KV KU ETWEKCN VQ QDVCKP VJG 0DGUV J[RQVJGUGU GHſEKGPVN[ UKPEG OWNVKRNG J[RQVJGUGU CTG PGEGUUCT[ HQT GTTQT EQTTGEVKQP KP VJG RQUV RTQEGUUKPI (QT VJKU TGCUQP C PGY FGEQFGT YJKEJ ECP RTQEGUU URGGEJ EQPVKPWQWUN[ YKVJQWV TGN[KPI QP UGPVGPEG DQWPFCT[ KPHQTOCVKQP JCU DGGP RTQRQUGF =?
E\&5&3UHVV//&
6.6 Automatic Speech Summarization and Evaluation 6.6.1 Summarization of Each Sentence Utterance %WTTGPVN[ XCTKQWU PGY CRRNKECVKQPU QH .8%54 U[UVGOU UWEJ CU CWVQOCVKE ENQUGF ECRVKQPKPI = ? OCMKPI OKPWVGU QH OGGVKPIU CPF EQPHGTGPEGU = ? CPF UWO OCTK\KPI CPF KPFGZKPI QH URGGEJ FQEWOGPVU HQT KPHQTOCVKQP TGVTKGXCN = ? CTG CEVKXGN[ DGKPI KPXGUVKICVGF 6TCPUETKDGF URGGEJ WUWCNN[ KPENWFGU PQV QPN[ TGFWP FCPV KPHQTOCVKQP UWEJ CU FKUƀWGPEKGU ſNNGF RCWUGU TGRGVKVKQPU TGRCKTU CPF YQTF HTCIOGPVU DWV CNUQ KTTGNGXCPV KPHQTOCVKQP ECWUGF D[ TGEQIPKVKQP GTTQTU 6JGTGHQTG GURGEKCNN[ HQT URQPVCPGQWU URGGEJ RTCEVKECN CRRNKECVKQPU WUKPI URGGEJ TGEQIPK\GT TG SWKTG C RTQEGUU QH URGGEJ UWOOCTK\CVKQP YJKEJ TGOQXGU TGFWPFCPV CPF KTTGNGXCPV KPHQTOCVKQP CPF GZVTCEVU TGNCVKXGN[ KORQTVCPV KPHQTOCVKQP EQTTGURQPFKPI VQ WUGTUŏ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ſECPEG UEQTG CU YGNN CU C EQPſ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
YJGTG CPF CTG YGKIJVKPI HCEVQTU HQT DCNCPEKPI COQPI CPF 6.6.1.1 Word Significance Score 6JG YQTF UKIPKſECPEG UEQTG KPFKECVGU VJG TGNCVKXG UKIPKſECPEG QH GCEJ YQTF KP VJG QTKIKPCN UGPVGPEG 6JG COQWPV QH KPHQTOCVKQP DCUGF QP VJG HTGSWGPE[ QH GCEJ
E\&5&3UHVV//&
6JGDGCWVKHWNEJGTT[DNQUUQOUDNQQOKPURTKPI
FIGURE 6.11 An example of dependency structure. YQTF KU WUGF CU VJG YQTF UKIPKſECPEG UEQTG HQT GCEJ VQRKE YQTF 9G EJQQUG PQWPU CPF XGTDU CU VQRKE YQTFU # ƀCV UEQTG KU IKXGP VQ YQTFU QVJGT VJCP VQRKE YQTFU 6Q TGFWEG VJG TGRGVKVKQP QH YQTFU KP VJG UWOOCTK\GF UGPVGPEG C ƀCV UEQTG KU CNUQ IKXGP VQ GCEJ TGCRRGCTKPI PQWP CPF XGTD 6.6.1.2 Linguistic Score 6JG NKPIWKUVKE UEQTG ½ OGCUWTGF D[ C DKITCO RTQDCDKNKV[ KPFKECVGU VJG CRRTQRTKCVGPGUU QH YQTF UVTKPIU KP C UWOOCTK\GF UGPVGPEG
½
6.6.1.3 Word Confidence Score 6JG EQPſFGPEG UEQTG KU KPEQTRQTCVGF VQ YGKIJ CEQWUVKECNN[ CU YGNN CU NKP IWKUVKECNN[ TGNKCDNG TGEQIPKVKQP TGUWNVU 5RGEKſECNN[ C RQUVGTKQT RTQDCDKNKV[ QH GCEJ VTCPUETKDGF YQTF VJCV KU VJG TCVKQ QH C YQTF J[RQVJGUKU RTQDCDKNKV[ VQ VJCV QH CNN QVJGT J[RQVJGUGU KU ECNEWNCVGF WUKPI C YQTF ITCRJ QDVCKPGF D[ C FGEQFGT CPF WUGF CU C EQPſFGPEG OGCUWTG 6.6.1.4 Word Concatenation Score 5WRRQUG őVJG DGCWVKHWN EJGTT[ DNQUUQOU DNQQO KP URTKPIŒ KU UWOOCTK\GF CU őVJG DGCWVKHWN URTKPIŒ 6JG NCVVGT RJTCUG KU C ITCOOCVKECNN[ EQTTGEV DWV UGOCPVKECNN[ KP EQTTGEV UWOOCTK\CVKQP 5KPEG VJG CDQXG NKPIWKUVKE UEQTG KU PQV RQYGTHWN GPQWIJ VQ CXQKF UWEJ C RTQDNGO VJG YQTF EQPECVGPCVKQP UEQTG ½ KU KPEQTRQTCVGF VQ IKXG C RGPCNV[ HQT C EQPECVGPCVKQP DGVYGGP YQTFU YKVJ PQ FGRGPFGPE[ KP VJG QTKIKPCN UGPVGPEG 6JG YQTF EQPECVGPCVKQP KP C UWOOCTK\GF UGPVGPEG KU TGUVTKEVGF D[ VJG FGRGPFGPE[ UVTWEVWTG KP VJG QTKIKPCN UGPVGPEG CU GZGORNKſGF KP (KIWTG 6JG YQTF CV VJG DGIKPPKPI QH CP CTTQY KU PCOGF őOQFKſGTŒ CPF VJG YQTF CV VJG GPF QH VJG CTTQY KU PCOGF őJGCFŒ TGURGEVKXGN[ 6JG 'PINKUJ FGRGPFGPE[ ITCOOCT EQPUKUVU QH DQVJ őTKIJVJGCFGFŒ FGRGPFGPE[ KPFKECVGF D[ TKIJV CTTQYU CPF őNGHVJGCFGFŒ FGRGPFGPE[ KPFKECVGF D[ NGHV CTTQYU CU UJQYP KP (KIWTG 6JG FGRGPFGPEKGU ECP DG YTKVVGP
E\&5&3UHVV//&
5 D E E
D D
w wi wi wm wk wk wn wl wj wj wL
FIGURE 6.12 A phrase structure tree based on a dependency structure.
CU RJTCUG UVTWEVWTG ITCOOCT &%() FGRGPFGPE[ EQPVGZV HTGG ITCOOCT
(right-headed) (left-headed)
YJGTG CTG PQPVGTOKPCN U[ODQNU CPF KU C VGTOKPCN U[ODQN YQTF 5KPEG VJG FGRGPFGPEKGU DGVYGGP YQTFU CTG WUWCNN[ CODKIWQWU YJGVJGT FGRGPFGP EKGU GZKUV QT PQV DGVYGGP YQTFU KU IKXGP D[ RTQDCDKNKVKGU VJCV QPG YQTF KU OQFKſGF D[ QVJGTU DCUGF QP VJG 5&%() 6JG YQTF FGRGPFGPE[ RTQDCDKNKV[ KU C RQUVGTKQT RTQDC DKNKV[ GUVKOCVGF D[ VJG +PUKFG1WVUKFG RTQDCDKNKVKGU QDVCKPGF WUKPI C OCPWCNN[ RCTUGF EQTRWU (KIWTG KNNWUVTCVGU CP GZCORNG QH C RJTCUG UVTWEVWTG VTGG DCUGF QP C FGRGPFGPE[ UVTWEVWTG HQT C UGPVGPEG EQPUKUVKPI QH YQTFU ½ 6JG RTQDCDKNKV[ VJCV CPF JCU C FGRGPFGPE[ UVTWEVWTG KU ECNEWNCVGF CU C RTQFWEV QH VJG RTQDCDKNKVKGU QH VJG HQNNQYKPI UGSWGPEG YJGP C UGPVGPEG KU FGTKXGF HTQO VJG KPKVKCN U[ODQN VJG KU CRRNKGF KU FGTKXGF HTQO KU FGTKXGF HTQO TWNG QH ·½ KU FGTKXGF HTQO CPF KU FGTKXGF HTQO 6JG RTQDCDKNKV[ QH CRRN[KPI VJG TWNG QH KU CNUQ CFFGF
+P IGPGTCN CU UJQYP KP (KIWTG C OQFKſGT FGTKXGF HTQO ECP DG FKTGEVN[ EQP PGEVGF YKVJ C JGCF FGTKXGF HTQO KP C UWOOCTK\GF UGPVGPEG +P CFFKVKQP VJG OQFK ſGT ECP DG CNUQ EQPPGEVGF YKVJ GCEJ YQTF YJKEJ OQFKſGU VJG JGCF 6JG YQTF EQP ECVGPCVKQP RTQDCDKNKV[ DGVYGGP CPF KU FGſPGF CU C UWO QH VJG FGRGPFGPE[ RTQDCDKNKVKGU DGVYGGP CPF CPF DGVYGGP CPF GCEJ QH ·½ 7U KPI VJG FGRGPFGPE[ RTQDCDKNKVKGU VJG YQTF EQPECVGPCVKQP UEQTG KU
E\&5&3UHVV//&
ECNEWNCVGF D[
+P VJG 5&%() QPN[ VJG PWODGT QH PQPVGTOKPCN U[ODQNU KU FGVGTOKPGF CPF CNN EQO DKPCVKQPU QH TWNGU CTG CRRNKGF TGEWTUKXGN[ 6JG PQPVGTOKPCN U[ODQN JCU PQ URGEKſE HWPEVKQP UWEJ CU C PQWP RJTCUG 'XGP KH VTCPUETKRVKQP TGUWNVU D[ C URGGEJ TGEQIPK\GT CTG KNNHQTOGF VJG FGRGPFGPE[ UVTWEVWTG ECP DG TQDWUVN[ GUVKOCVGF D[ VJG 5&%() +P VJG ECUG QH ,CRCPGUG WVVGTCPEG UWOOCTK\CVKQP VJG YQTF EQPECVGPCVKQP UEQTG KU OQTG EQORCEV VJCP 'PINKUJ UKPEG ,CRCPGUG UGPVGPEGU JCXG QPN[ őTKIJVJGCFGFŒ FG RGPFGPEKGU +P CFFKVKQP VJG YQTF FGRGPFGPE[ UVTWEVWTG KP GCEJ RJTCUG KU FGVGTOKP KUVKE CPF ECP DG TGRTGUGPVGF D[ VJG TGIWNCT ITCOOCT
6.6.2 Summarization of Multiple Utterances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ſ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
6.6.3 Evaluation 6.6.3.1 Word Network of Manual Summarization Results for Evaluation 6Q CWVQOCVKECNN[ GXCNWCVG UWOOCTK\GF UGPVGPEGU EQTTGEVN[ VTCPUETKDGF URGGEJ KU OCPWCNN[ UWOOCTK\GF D[ JWOCP UWDLGEVU CPF WUGF CU EQTTGEV VCTIGVU 6JG OCP WCN UWOOCTK\CVKQP TGUWNVU CTG OGTIGF KPVQ C YQTF PGVYQTM YJKEJ CRRTQZKOCVGN[ GZ
E\&5&3UHVV//&
RTGUUGU CNN RQUUKDNG EQTTGEV UWOOCTK\CVKQP KPENWFKPI UWDLGEVKXG XCTKCVKQPU # őUWO OCTK\CVKQP CEEWTCE[Œ QH CWVQOCVKE UWOOCTK\CVKQP KU ECNEWNCVGF WUKPI VJG YQTF PGV YQTM # YQTF UVTKPI VJCV KU VJG OQUV UKOKNCT VQ VJG CWVQOCVKE UWOOCTK\CVKQP TGUWNV GZVTCEVGF HTQO VJG YQTF PGVYQTM KU EQPUKFGTGF CU C EQTTGEV VCTIGV HQT VJG CWVQOCVKE UWOOCTK\CVKQP 6JG CEEWTCE[ EQORCTKPI VJG UWOOCTK\GF UGPVGPEG YKVJ VJG VCTIGV YQTF UVTKPI KU WUGF CU C OGCUWTG QH VJG NKPIWKUVKE EQTTGEVPGUU CPF OCKPVGPCPEG QH QTKIKPCN OGCPKPIU QH VJG WVVGTCPEG 6.6.3.2 Evaluation Data (KTUV ,CRCPGUG 68 DTQCFECUV PGYU WVVGTCPEGU TGEQTFGF KP YGTG WUGF VQ GXCNWCVG VJG RTQRQUGF OGVJQF (KHV[ WVVGTCPEGU YKVJ YQTF TGEQIPKVKQP CEEWTCE[ CDQXG YJKEJ YCU VJG CXGTCIG TCVG QXGT VJG WVVGTCPEGU YGTG UGNGEVGF CPF WUGF HQT VJG GXCNWCVKQP +P CFFKVKQP ſXG PGYU CTVKENGU EQPUKUVKPI QH ſXG UGPVGPEGU GCEJ YGTG UWOOCTK\GF WUKPI VJG UWOOCTK\CVKQP VGEJPKSWG HQT OWNVKRNG WVVGTCPEGU 0GZV 'PINKUJ 68 DTQCFECUV PGYU WVVGTCPEGU %00 PGYU TGEQTFGF KP RTQXKFGF D[ 0+56 CU C VGUV UGV QH VQRKE FGVGEVKQP CPF VTCEMKPI 6&6 YGTG VCIIGF D[ VJG $TKNN VCIIGT CPF WUGF VQ GXCNWCVG VJG RTQRQUGF OGVJQF (KXG PGYU CTVKENGU EQPUKUVKPI QH WVVGTCPEGU KP CXGTCIG YGTG VTCPUETKDGF D[ VJG ,#075 =? URGGEJ TGEQIPKVKQP U[U VGO 6JG OWNVKRNG WVVGTCPEG UWOOCTK\CVKQP YCU RGTHQTOGF HQT GCEJ QH VJG ſXG PGYU CTVKENGU (KHV[ WVVGTCPEGU CTDKVTCTKN[ EJQUGP HTQO VJG ſXG PGYU CTVKENGU YGTG WUGF HQT VJG UGPVGPEG D[ UGPVGPEG UWOOCTK\CVKQP /GCP YQTF TGEQIPKVKQP CEEWTCEKGU QH VJG WVVGTCPEGU WUGF HQT VJG OWNVKRNG WVVGTCPEG UWOOCTK\CVKQP CPF VJQUG HQT UGPVGPEG D[ UGPVGPEG UWOOCTK\CVKQP YGTG CPF TGURGEVKXGN[ 6.6.3.3 Training Data for Summarization Models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ſTOGF VJCV VJG CWVQOCVKECNN[ UWOOCTK\GF UGPVGPEGU WUKPI YQTF DKITCO DCUGF QP PGYURCRGT VGZV YGTG OWEJ DGVVGT VJCP VJQUG DCUGF QP DTQCFECUV PGYU OCPWUETKRVU =? 5&%() HQT YQTF EQPECVGPCVKQP UEQTG YCU DWKNV WUKPI VGZV HTQO VJG OCPWCNN[ RCTUGF EQTRWU QH VJG /CKPKEJK PGYURCRGT RWDNKUJGF HTQO VQ EQORTKUKPI CRRTQZKOCVGN[ / UGPVGPEGU YKVJ / YQTFU 6JG PWODGT QH PQPVGTOKPCN U[ODQNU YCU +P VJG 'PINKUJ URGGEJ ECUG C YQTF UKIPKſECPEG OQFGN C DKITCO NCPIWCIG OQFGN CPF 5&%() YGTG EQPUVTWEVGF WUKPI TQWIJN[ / YQTFU HTQO QXGT M UGPVGPEGU QH VJG 9CNN 5VTGGV ,QWTPCN EQTRWU CPF VJG $TQYP EQTRWU KP 2GPP 6TGGDCPM
E\&5&3UHVV//&
REC
TRS
REC
,CRCPGUG
I_L_T SUB
I I_L
RDM
I
I_L_C I_L_T I_L_C_T
I_L
I_L_T SUB
I I_L
RDM
I_L_C I_L_T I_L_C_T
I_L
I
RDM
5WOOCTK\CVKQKPCEEWTCE[=?
RDM
5WOOCTK\CVKQKPCEEWTCE[=?
TRS
'PINKUJ
FIGURE 6.13 Each utterance summarizations at 70% summarization ratio. 6.6.3.4 Evaluation Results /CPWCN VTCPUETKRVKQP 645 CPF CWVQOCVKE VTCPUETKRVKQP 4'% YGTG DQVJ UWOOC TK\GF +P VJG UWOOCTK\CVKQP QH 4'% VJG HQNNQYKPI UEQTG EQPFKVKQPU YGTG EQORCTGF
¯ 5KIPKſECPEG UEQTG ¯ 5KIPKſECPEG CPF NKPIWKUVKE UEQTGU ¯ 5KIPKſECPEG NKPIWKUVKE CPF EQPſFGPEG UEQTGU ¯ 5KIPKſECPEG NKPIWKUVKE CPF EQPECVGPCVKQP UEQTGU ¯ #NN UEQTGU +P VJG UWOOCTK\CVKQP QH 645 UKPEG VJGTG KU PQ TGEQIPKVKQP GTTQT VJG EQPFKVKQPU KP ENWFKPI VJG EQPſ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ſECPVN[ OQTG GHHGEVKXG VJCP 4&/ 6JG DGVVGT TGUWNVU QDVCKPGF D[ KPEQTRQTCVKPI GCEJ UEQTG KPFKECVG VJCV CNN QH VJG UEQTGU CTG GHHGEVKXG VQ KORTQXG VJG UWOOCTK\CVKQP CEEWTCE[ &GVCKNGF
E\&5&3UHVV//&
TRS
I I_L
I_L_C_T
I_L_T
SUB
,CRCPGUG
I_L_C I_L_T
I
I_L
I_L_T SUB
I_L I
RDM
REC
RDM
I
I_L
I_L_C I_L_T I_L_C_T
RDM
5WOOCTK\CVKQKPCEEWTCE[=?
RDM
5WOOCTK\CVKQKPCEEWTCE[=?
REC
TRS
'PINKUJ
FIGURE 6.14 Article summarizations at 30% summarization ratio. KPXGUVKICVKQP TGXGCNU VJCV VJG OGVJQF WUKPI VJG YQTF EQPECVGPCVKQP UEQTG TGFWEGU OGCPKPI CNVGTCVKQP
6.6.4 Discussion 'CEJ WVVGTCPEG CPF C YJQNG PGYU CTVKENG EQPUKUVKPI QH OWNVKRNG WVVGTCPEGU QH ,CRCPGUG CPF 'PINKUJ DTQCFECUV PGYU URGGEJ JCXG DGGP UWOOCTK\GF D[ VJG CWVQOCVKE URGGEJ UWOOCTK\CVKQP OGVJQF DCUGF QP VJG YQTF UKIPKſECPEG NKPIWKUVKE YQTF EQPſ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ſFGPEG UEQTG YJKEJ JCU DGGP KPEQTRQTCVGF KPVQ VJG UWOOCTK\CVKQP UEQTG VQ GZENWFG YQTF TGEQIPKVKQP GTTQTU VJG NKPIWKUVKE UEQTG KU GHHGEVKXG VQ TGFWEG QWVQHEQPVGZV YQTF GZVTCEVKQP DQVJ HTQO TGEQIPKVKQP GTTQTU CPF JWOCP FKUƀWGPEKGU +P UWOOCTK\KPI ,CRCPGUG PGYU URGGEJ VJG EQPſFGPEG OGCUWTG EQWNF KORTQXG VJG UWOOCTK\KPI RGTHQTOCPEG D[ GZENWFKPI KPEQPVGZV YQTF GTTQTU +P VJG 'PINKUJ ECUG VJG EQPſFGPEG OGCUWTG ECP PQV QPN[ GZENWFG YQTF GTTQTU DWV CNUQ JGNR GZVTCEVKPI ENGCTN[ RTQPQWPEGF KORQTVCPV YQTFU %QPUGSWGPVN[ VJG WUG QH VJG EQPſ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
E\&5&3UHVV//&
VKQPKPI QH DTQCFECUV PGYU CPF RTGUGPVKPI KPHQTOCVKQP KP PGYUQPFGOCPF U[UVGOU
6.7 Spontaneous Speech Recognition and Understanding Research Issues 6.7.1 Language Models and Corpora 1PG QH VJG OQUV KORQTVCPV RTGUGPV KUUWGU HQT URQPVCPGQWU URGGEJ TGEQIPKVKQP KU JQY VQ ETGCVG NCPIWCIG OQFGNU TWNGU 9JGP TGEQIPK\KPI URQPVCPGQWU URGGEJ KV KU PGE GUUCT[ VQ FGCN YKVJ XCTKCVKQPU VJCV CTG PQV GPEQWPVGTGF YJGP TGEQIPK\KPI URGGEJ VJCV KU TGCF HTQO VGZVU 6JGUG XCTKCVKQPU KPENWFG GZVTCPGQWU YQTFU QWVQHXQECDWNCT[ YQTFU WPITCOOCVKECN UGPVGPEGU FKUƀWGPE[ RCTVKCN YQTFU TGRCKTU JGUKVCVKQPU CPF TGRGVKVKQPU 5VQEJCUVKE NCPIWCIG OQFGNKPI UWEJ CU DKITCOU CPF VTKITCOU JCU DGGP C XGT[ RQYGTHWN VQQN UQ KV YQWNF DG XGT[ GHHGEVKXG VQ GZVGPF KVU WVKNKV[ D[ KPEQTRQTCV KPI UGOCPVKE MPQYNGFIG +V YQWNF CNUQ DG WUGHWN VQ KPVGITCVG WPKſECVKQP ITCOOCTU CPF EQPVGZVHTGG ITCOOCTU HQT GHſEKGPV YQTF RTGFKEVKQP +V KU ETWEKCN VQ FGXGNQR TQ DWUV CPF ƀGZKDNG FGEQFKPI CNIQTKVJOU VJCV OCVEJ VJG EJCTCEVGTKUVKEU QH URQPVCPGQWU URGGEJ # RCTCFKIO UJKHV HTQO VJG RTGUGPV VTCPUETKRVKQPDCUGF CRRTQCEJ VQ C FGVGEVKQPDCUGF CRRTQCEJ YKNN DG KORQTVCPV VQ UQNXG VJG URQPVCPGQWUURGGEJ URGEKſE RTQDNGOU =? *QY VQ GZVTCEV EQPVGZVWCN KPHQTOCVKQP RTGFKEV WUGTUŏ TGURQPUGU CPF HQEWU QP MG[ YQTFU CTG XGT[ KORQTVCPV KUUWGU 5V[NG UJKHVKPI KU CNUQ CP KORQTVCPV RTQDNGO KP URQP VCPGQWU URGGEJ TGEQIPKVKQP +P V[RKECN NCDQTCVQT[ GZRGTKOGPVU URGCMGTU CTG TGCFKPI NKUVU QH UGPVGPEGU TCVJGT VJCP VT[KPI VQ CEEQORNKUJ C TGCN VCUM 7UGTU CEVWCNN[ VT[KPI VQ CEEQORNKUJ C VCUM JQYGXGT WUG C FKHHGTGPV NKPIWKUVKE UV[NG #FCRVCVKQP QH NKPIWKUVKE OQFGNU CEEQTFKPI VQ VCUMU CPF VQRKEU KU CNUQ C XGT[ KORQTVCPV KUUWG UKPEG EQNNGEVKPI C NCTIG NKPIWKUVKE FCVCDCUG HQT GXGT[ PGY VCUM KU FKHſEWNV CPF EQUVN[ 6JG CRRGVKVGU QH VQFC[ŏU UVCVKUVKECN URGGEJ RTQEGUUKPI VGEJPKSWGU HQT VTCKPKPI OCVGTKCN CTG YGNN FGUETKDGF D[ VJG CRJQTKUO ő6JGTGŏU PQ FCVC NKMG OQTG FCVCŒ .CTIG UVTWE VWTGF EQNNGEVKQPU QH URGGEJ CPF VGZV CTG GUUGPVKCN VQ RTQITGUU KP URGGEJ TGEQIPKVKQP TGUGCTEJ 7PNKMG VJG VTCFKVKQPCN CRRTQCEJ KP YJKEJ MPQYNGFIG QH VJG URGGEJ DGJCX KQT KU őFKUEQXGTGFŒ CPF őFQEWOGPVGFŒ D[ JWOCP GZRGTVU UVCVKUVKECN OGVJQFU RTQXKFG CP CWVQOCVKE RTQEGFWTG VQ őNGCTPŒ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
E\&5&3UHVV//&
FKXGTUKV[ RCTVKEWNCTN[ KP UKVWCVKQPU YJGTG VJG TGCN OCEJKPG OC[ HCKN # JWOCP YK\CTF ECPPQV KPVGPVKQPCNN[ UKOWNCVG CNN V[RGU QH OCEJKPG GTTQT CPF VJWU VJG TGEQTFGF FCVC OC[ HCKN VQ RTQXKFG EQORNGVG KPHQTOCVKQP QH TGCN JWOCPOCEJKPG KPVGTCEVKQP 6JG TGEQTFGF FCVC PGGFU VQ DG XGTKſGF NCDGNGF CPF CPPQVCVGF D[ RGQRNG YJQUG MPQYNGFIG YKNN DG KPVTQFWEGF KPVQ VJG FGUKIP QH VJG U[UVGO VJTQWIJ KVU NGCTPKPI RTQ EGUU KG XKC UWRGTXKUGF VTCKPKPI QH VJG U[UVGO CHVGT VJG FCVC JCU DGGP NCDGNGF .C DGNKPI CPF CPPQVCVKQP HQT URQPVCPGQWU URGGEJ ECP GCUKN[ DGEQOG WPOCPCIGCDNG (QT GZCORNG JQY FQ YG CPPQVCVG URGGEJ TGRCKTU CPF RCTVKCN YQTFU JQY FQ VJG RJQPGVKE VTCPUETKDGTU TGCEJ C EQPUGPUWU KP CEQWUVKERJQPGVKE NCDGNU YJGP VJGTG KU CODKIWKV[ CPF JQY FQ YG TGRTGUGPV C UGOCPVKE PQVKQP! 'TTQTU KP NCDGNKPI CPF CPPQVCVKQP YKNN TGUWNV KP U[UVGO RGTHQTOCPEG FGITCFCVKQP *QY VQ GPUWTG VJG SWCNKV[ QH VJG CPPQVCVGF TGUWNVU KU VJWU QH C OCLQT EQPEGTP 4GUGCTEJ KP CWVQOCVKPI QT ETGCVKPI VQQNU VQ CUUKUV VJG XGTKſECVKQP RTQEGFWTG KU D[ KVUGNH CP KPVGTGUVKPI UWDLGEV #PQVJGT CTGC QH TGUGCTEJ VJCV JCU ICKPGF KPVGTGUV KU C OQFGNKPI OGVJQFQNQI[ CPF VJG CUUQEKCVGF FCVC EQNNGEVKQP UEJGOG VJCV ECP TGFWEG VJG VCUM FGRGPFGPE[ 6Q OCZKOK\G VJG RGTHQTOCPEG QPG UJQWNF CNYC[U UVTKXG HQT FCVC VJCV VTWN[ TGƀGEVU VJG QRGTCVKPI EQPFKVKQP +V VJWU ECNNU HQT C FCVCDCUG EQNNGEVKQP RNCP VJCV KU EQPUKUVGPV YKVJ VJG VCUM 6JKU FCVC EQNNGEVKQP GHHQTV YQWNF UQQP DGEQOG WPOCPCIGCDNG KH VJG U[UVGO FGUKIPGT JCU VQ TGFQ FCVC EQNNGEVKQP HQT GCEJ CPF GXGT[ CRRNKECVKQP VJCV KU DGKPI FGXGNQRGF +V KU VJGTGHQTG FGUKTCDNG VQ FGUKIP C VCUMKPFGRGPFGPV FCVC UGV CPF C OQFGNKPI OGVJQF VJCV FGNKXGTU C TGCUQPCDNG RGTHQTOCPEG WRQP ſTUV WUG CPF ECP SWKEMN[ CNNQY KPſGNF VTKCNU HQT HWTVJGT TGXKUKQP CU UQQP CU VCUMFGRGPFGPV FCVC DGEQOG CXCKNCDNG 4GUGCTEJ TGUWNV KP VJKU CTGC ECP QHHGT VJG DGPGſV QH C TGFWEGF CRRNKECVKQP FGXGNQROGPV EQUV
6.7.2 Message-driven Speech Recognition and Understanding 5VCVGQHVJGCTV CWVQOCVKE URGGEJ TGEQIPKVKQP U[UVGOU GORNQ[ VJG ETKVGTKQP QH OCZK OK\KPI YJGTG ½ KU C YQTF UGSWGPEG CPF ½ KU CP CEQWUVKE QDUGTXCVKQP UGSWGPEG 6JKU ETKVGTKQP KU TGCUQPCDNG HQT FKEVCVKPI TGCF URGGEJ *QYGXGT VJG WNVKOCVG IQCN QH CWVQOCVKE URGGEJ TGEQIPKVKQP KU VQ GZVTCEV VJG WPFGTN[KPI OGUUCIGU QH VJG URGCMGT HTQO VJG URGGEJ UKIPCNU *GPEG YG PGGF VQ OQFGN VJG RTQEGUU QH URGGEJ IGPGTCVKQP CPF TGEQIPKVKQP CU UJQYP KP (KIWTG =? YJGTG KU VJG OGUUCIG EQPVGPV VJCV C URGCMGT KPVGPFGF VQ EQPXG[ 6JG OGUUCIG KU TGCNK\GF CU C YQTF UGSWGPEG VJTQWIJ C NKPIWKUVKE EJCPPGN URGEKſGF D[ C RTQD CDKNKV[ OGCUWTG 6JG NKPIWKUVKE EJCPPGN KU RTQDCDKNKUVKE CU VJGTG CTG OCP[ YC[U VQ GZRTGUU VJG UCOG OGUUCIG UQOG OQTG NKMGN[ VJCP QVJGTU 6JG YQTF UGSWGPEG VJGP IGVU TGCNK\GF VJTQWIJ VJG CEQWUVKE EJCPPGN CU C UGSWGPEG QH CEQWUVKE UKIPCNU 6JG CEQWUVKE EJCPPGN KPVTQFWEGU XCTKCDKNKV[ FWG VQ XCTKQWU TGCUQPU KP ENWFKPI URGCMGTU CPF CEQWUVKE GPXKTQPOGPVU 0Q QPG URGCMGT ECP TGRGCV GZCEVN[ VJG UCOG YCXGHQTO GXGP WVVGTKPI VJG UCOG YQTF CPF PQ VYQ URGCMGTU CTG CNKMG KP VGTOU QH VJG EQPſIWTCVKQP QH VJGKT CTVKEWNCVQT[ CRRCTCVWU 6JG UGSWGPEG QH UQWPFU TCFK CVGF HTQO VJG OQWVJ QH VJG URGCMGT RTQRCICVGU KP CEQWUVKE YCXGU VJTQWIJ VJG TQQO 6JG CEQWUVKE YCXG EQPXQNXGF YKVJ VJG TQQO CEQWUVKE TGURQPUG CPF OKZGF YKVJ VJG CEQWUVKE CODKGPV TGCEJGU VJG OKETQRJQPG CPF KU ſPCNN[ EQPXGTVGF KPVQ CP GNGEVTKE UKIPCN 6JG GNGEVTKE UKIPCN RTQRCICVGU VJTQWIJ C VTCPUOKUUKQP TQWVG ECDNGU YKTGU QT
E\&5&3UHVV//&
3 :_0
3;_:
/LQJXLVWLF : FKDQQHO
$FRXVWLF FKDQQHO
3 0
0HVVDJH VRXUFH
0
/DQJXDJH 9RFDEXODU\ *UDPPDU 6HPDQWLFV &RQWH[W +DELWV
;
6SHHFK UHFRJQL]HU
6SHDNHU 5HYHUEHUDWLRQ 1RLVH 7UDQVPLVVLRQ FKDUDFWHULVWLFV 0LFURSKRQH
FIGURE 6.15 A communication-theoretic view of speech generation and recognition.
VJG VGNGRJQPG PGVYQTM CPF DGEQOGU YJGP KV KU TGEGKXGF D[ VJG TGEQIPKVKQP CPF WPFGTUVCPFKPI U[UVGO %JCTCEVGTKUVKEU QH CNN VJGUG RTQEGUUGU XCT[ UWDUVCPVKCNN[ #EEQTFKPI VQ VJKU OQFGN VJG URGGEJ TGEQIPKVKQP CPF WPFGTUVCPFKPI RTQEGUU KU VQ TGXGTUG VJG IGPGTCVKQP RTQEGUU VQ TGEQXGT YJKEJ ECP DG TGRTGUGPVGF CU VJG OCZK OK\CVKQP QH VJG HQNNQYKPI C RQUVGTKQTK RTQDCDKNKV[ =?
7UKPI $C[GUŏ TWNG 'S ECP DG GZRTGUUGF CU
(QT UKORNKEKV[ YG ECP CRRTQZKOCVG VJG GSWCVKQP CU
KU ECNEWNCVGF WUKPI JKFFGP /CTMQX OQFGNU KP VJG UCOG YC[ CU KP WUWCN TGEQIPKVKQP RTQEGUUGU 6JKU PGY HQTOWNCVKQP QH URGGEJ TGEQIPKVKQP YCU CRRNKGF VQ VJG ,CRCPGUG DTQCFECUV PGYU VTCPUETKRVKQP CPF KV YCU HQWPF VJCV YQTF GTTQT TCVGU YGTG UNKIJVN[ TGFWEGF D[ VJKU OGVJQF 6JGTG KU CNUQ C RQUUKDKNKV[ VQ IKXG HGGFDCEM HTQO VJG őWPFGTUVCPFKPI OQFWNGŒ VQ VJG URGGEJ TGEQIPKVKQP OQFWNG UWEJ VJCV FGEQFKPI J[RQVJGUGU ECP DG RTQRGTN[ CFLWUVGF CPF JQRGHWNN[ EQPXGTIG VQ VJG OQUV EQTTGEV YQTF UGSWGPEG CU YGNN CU VJG OQUV EQT TGEV WPFGTUVCPFKPI QH VJG WVVGTCPEG
E\&5&3UHVV//&
6.7.3 Statistical Approaches and Speech Science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ſEKCNN[ 5RGGEJ VGEJPQNQI[ KU TG NCVGF VQ OCP[ UEKGPVKſE CPF GPIKPGGTKPI ſGNFU KPENWFKPI RJ[UKQNQI[ CPF RU[EJQNQI[ QH URGGEJ RTQFWEVKQP CPF RGTEGRVKQP CEQWUVKEU RJ[UKEU UKIPCN RTQEGUUKPI EQOOW PKECVKQP CPF KPHQTOCVKQP VJGQT[ EQORWVGT UEKGPEG RCVVGTP TGEQIPKVKQP CPF NKPIWKU VKEU KV JCU CP KPVGTFKUEKRNKPCT[ PCVWTG +V ECP CNUQ DG UCKF VJCV URGGEJ TGUGCTEJ GZKUVU CV VJG DQWPFCT[ DGVYGGP PCVWTCN UEKGPEG CPF GPIKPGGTKPI -PQYNGFIG CPF VGEJPQNQI[ HTQO C YKFG TCPIG QH CTGCU KPENWFKPI VJG WUG QH CTVKE WNCVQT[ CPF RGTEGRVWCN EQPUVTCKPVU YKNN DG PGEGUUCT[ VQ FGXGNQR URGGEJ VGEJPQNQI[ (QT GZCORNG YJGP UGXGTCN RJQPGOGU QT U[NNCDNGU CTG EQPVKPWQWUN[ URQMGP CU KP VJG ECUG QH WUWCN UGPVGPEG URGGEJ VJG VQPIWG LCY NKRU GVE OQXG CU[PEJTQPQWUN[ KP RCTCNNGN CPF [GV YKVJ EQWRNGF TGNCVKQPUJKRU %WTTGPV URGGEJ CPCN[UKU VGEJPKSWGU JQYGXGT TGRTGUGPV URGGEJ CU C UKORNG VKOG UGTKGU QH URGEVTC +V YKNN DGEQOG PGE GUUCT[ VQ CPCN[\G URGGEJ D[ FGEQORQUKPI KV KPVQ UGXGTCN JKFFGP HCEVQTU DCUGF QP URGGEJ RTQFWEVKQP OGEJCPKUOU 6JKU CRRTQCEJ UGGOU VQ DG GUUGPVKCN HQT UQNXKPI VJG EQCTVKEWNCVKQP RTQDNGO QPG QH VJG OQUV KORQTVCPV RTQDNGOU KP URGGEJ TGEQIPKVKQP 6JG JWOCP JGCTKPI U[UVGO KU HCT OQTG TQDWUV VJCP OCEJKPG U[UVGOU - OQTG TQDWUV PQV QPN[ CICKPUV VJG FKTGEV KPƀWGPEG QH CFFKVKXG PQKUG DWV CNUQ CICKPUV URGGEJ XCTKCVKQPU
VJCV KU VJG KPFKTGEV KPƀWGPEG QH PQKUG GXGP KH VJG PQKUG KU XGT[ KPEQPUKUVGPV 5RGGEJ TGEQIPK\GTU CTG VJGTGHQTG GZRGEVGF VQ DGEQOG OQTG TQDWUV YJGP VJG HTQPV GPF WVKNK\GU OQFGNU QH JWOCP JGCTKPI 6JKU ECP DG FQPG D[ KOKVCVKPI VJG RJ[UKQNQIKECN QTICPU QT D[ TGRTQFWEKPI RU[EJQCEQWUVKE EJCTCEVGTKUVKEU #NVJQWIJ KV KU PQV CNYC[U PGEGUUCT[ QT GHſEKGPV HQT URGGEJ TGEQIPKVKQP U[UVGOU VQ FKTGEVN[ KOKVCVG JWOCP URGGEJ RTQFWE VKQP CPF RGTEGRVKQP OGEJCPKUOU KV YKNN DGEQOG OQTG KORQTVCPV KP VJG PGCT HWVWTG VQ DWKNF OCVJGOCVKECN OQFGNU DCUGF QP VJGUG OGEJCPKUOU KP QTFGT VQ KORTQXG VJG RGTHQTOCPEG QH URQPVCPGQWU URGGEJ TGEQIPKVKQP =?
6.7.4 Research on the Human Brain 7R VQ VJG RTGUGPV VJG ſGNFU QH URGGEJ RGTEGRVKQP CPF CWVQOCVKE URGGEJ TGEQIPKVKQP JCXG DGGP YKFGN[ UGRCTCVG *QYGXGT KP QTFGT VQ DWKNF URQPVCPGQWU URGGEJ WPFGT UVCPFKPI U[UVGOU KV KU ETWEKCN VQ CPCN[\G VJG HWPEVKQP YKVJKP VJG JWOCP DTCKP 6JG
E\&5&3UHVV//&
HWPEVKQP OWUV VJGP DG TGCNK\GF WUKPI GPIKPGGTKPI OQFGNU (QT VJGUG RWTRQUGU JWOCP URGGEJ RGTEGRVKQP TGUGCTEJ PGGFU VQ UJKHV HTQO VCTIGVKPI UJQTV HTCIOGPVU UWEJ CU RJQPGOGU CPF U[NNCDNGU VQ NCTIGT WPKVU UWEJ CU YQTFU RJTCUGU UGPVGPEGU CPF RCTC ITCRJU 4GUGCTEJ UJQWNF KPXGUVKICVG JQY OGCPKPIU EQPXG[GF D[ URGGEJ CTG WPFGT UVQQF +V KU KPFKURGPUCDNG VQ DWKNF C NCTIG EQTRWU QH URQPVCPGQWU URGGEJ CPF EQPFWEV EQTRWUDCUGF TGUGCTEJ QP DQVJ VJG OGEJCPKUOU QH JWOCP URGGEJ RGTEGRVKQP CPF VJG GPIKPGGTKPI URGGEJ WPFGTUVCPFKPI U[UVGOU YKVJ ENQUG EQPPGEVKQP CPF EQQRGTCVKQP 7NVKOCVGN[ KP QTFGT VQ OCMG URGGEJ TGEQIPKVKQP U[UVGOU TGCNN[ WUGHWN CPF EQOHQTV CDNG HQT WUGTU VJG[ UJQWNF OCVEJ QT GZEGGF JWOCP ECRCDKNKVKGU 6JCV KU VJG[ UJQWNF DG HCUVGT OQTG CEEWTCVG OQTG KPVGNNKIGPV OQTG MPQYNGFIGCDNG NGUU GZRGPUKXG CPF GCUKGT VQ EQOOWPKECVG YKVJ VJCP JWOCP UVCHH (QT VJKU RWTRQUG VJG WNVKOCVG U[UVGOU OWUV DG CDNG VQ JCPFNG EQPEGRVWCN KPHQTOCVKQP (KIWTG UJQYU C FKCITCO QH JW OCP URGGEJ IGPGTCVKQP CPF RGTEGRVKQP RTQEGUU #NVJQWIJ QDUGTXCVKQP CPF OQFGNKPI QH VJG OQXGOGPV QH XQECN U[UVGOU CNQPI YKVJ VJG RJ[UKQNQIKECN OQFGNKPI QH CWFKVQT[ RGTKRJGTCN U[UVGOU JCXG TGEGPVN[ OCFG ITGCV RTQITGUU VJG OGEJCPKUO QH URGGEJ KP HQTOCVKQP RTQEGUUKPI KP QWT JWOCP DTCKP JCU JCTFN[ DGGP KPXGUVKICVGF 2U[EJQNQIKECN GZRGTKOGPVU QP JWOCP OGOQT[ ENGCTN[ UJQY VJCV URGGEJ RNC[U C HCT OQTG KORQTVCPV CPF GUUGPVKCN TQNG VJCP XKUKQP KP VJG JWOCP OGOQT[ CPF VJKPMKPI RTQEGUUGU 9JGTGCU OQFGNU QH UGRCTCVKPI CEQWUVKE UQWTEGU JCXG DGGP TGUGCTEJGF KP őCWFKVQT[ UEGPG CPCN [UKUŒ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
6.7.5 Dynamic Spectral Features 2U[EJQNQIKECN CPF RJ[UKQNQIKECN TGUGCTEJ KPVQ JWOCP URGGEJ RGTEGRVKQP OGEJCPKUOU UJQYU VJCV VJG JWOCP JGCTKPI QTICPU CTG JKIJN[ UGPUKVKXG VQ EJCPIGU KP UQWPFU KG VQ VTCPUKVKQPCN F[PCOKE UQWPFU CPF VJCV VJG VTCPUKVKQPCN HGCVWTGU QH VJG URGGEJ URGE VTWO CPF VJG URGGEJ YCXG RNC[ ETWEKCN TQNGU KP RJQPGOG RGTEGRVKQP =? 6JG NGPIVJ QH VJG VKOG YKPFQYU KP YJKEJ UQWPF VTCPUKVKQPU CTG RGTEGKXGF JCXG C JKGTCTEJKECN UVTWEVWTG CPF TCPIG HTQO VJG QTFGT QH UGXGTCN OKNNKUGEQPFU VQ UGXGTCN UGEQPFU 6JG JK GTCTEJKECN NC[GTU EQTTGURQPF VQ XCTKQWU URGGEJ HGCVWTGU UWEJ CU RJQPGOGU U[NNCDNGU CPF RTQUQFKE HGCVWTGU +V JCU CNUQ DGGP TGRQTVGF VJCV VJG JWOCP JGCTKPI OGEJCPKUO RGTEGKXGU C VCTIGV XCNWG GUVKOCVGF HTQO VJG VTCPUKVKQPCN KPHQTOCVKQP GZVTCEVGF WUKPI F[PCOKE URGEVTCN HGCVWTGU 6JG TGRTGUGPVCVKQP QH VJG F[PCOKE EJCTCEVGTKUVKEU QH URGGEJ YCXGU CPF URGEVTC JCU DGGP UVWFKGF CPF UGXGTCN WUGHWN OGVJQFU JCXG DGGP RTQRQUGF = ? *QYGXGT VJG RGTHQTOCPEG QH VJGUG OGVJQFU KU PQV [GV UCVKUHCEVQT[ CPF OQUV QH VJG UWEEGUUHWN URGGEJ CPCN[UKU OGVJQFU FGXGNQRGF VJWU HCT CUUWOG C UVCVKQPCT[ UKIPCN CV NGCUV HQT
E\&5&3UHVV//&
6SHHFKJHQHUDWLRQ 7H[WJHQHUDWLRQ
6SHHFKSURGXFWLRQ
,QWHQVLRQ
3KRQHPHV SURVRG\
$UWLFXODWRU\ PRWLRQV
0HVVDJH IRUPXODWLRQ
/DQJXDJH FRGH
1HXURPXVFXODU FRQWUROV
'LVFUHWHVLJQDO ESV
&RQWLQXRXVVLJQDO
ESV
ESV
,QIRUPDWLRQUDWH 6HPDQWLFV
9RFDOWUDFW V\VWHP a ESV $FRXVWLF ZDYHIRUP
3KRQHPHV :RUGVV\QWD[
)HDWXUH ([WUDFWLRQ FRGLQJ
6SHFWUXP DQDO\VLV
/DQJXDJH WUDQVODWLRQ
1HXUDO WUDQVGXFWLRQ
%DVLODU 0HPEUDQH PRWLRQ
0HVVDJH XQGHUVWDQGLQJ
'LVFUHWHVLJQDO
/LQJXLVWLFGHFRGLQJ
&RQWLQXRXVVLJQDO
6SHHFKSHUFHSWLRQ
$FRXVWLFSURFHVVLQJ
FIGURE 6.16 Speech-generation and speech-perception processes. GCEJ DCUKE UJQTV RGTKQF +V KU UVKNN XGT[ FKHſEWNV VQ TGNCVG VKOG HWPEVKQPU QH RKVEJ CPF GPGTI[ VQ RGTEGRVWCN RTQUQFKE KPHQTOCVKQP +H IQQF OGVJQFU HQT TGRTGUGPVKPI VJG F[PCOKEU QH URGGEJ CUUQEKCVGF YKVJ XCTKQWU VKOG NGPIVJU CTG FKUEQXGTGF VJG[ UJQWNF JCXG C UWDUVCPVKCN KORCEV QP VJG EQWTUG QH URQPVCPGQWU URGGEJ TGUGCTEJ
6.8 Conclusion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
E\&5&3UHVV//&
URGGEJ TGEQIPKVKQP VQ WPFGTUVCPFKPI YJGTG VJG WPFGTN[KPI OGUUCIGU QH VJG URGCMGT KG OGCPKPIEQPVGPV VJCV VJG URGCMGT KPVGPFGF VQ EQPXG[ CTG GZVTCEVGF KPUVGCF QH UKORN[ VTCPUETKDKPI CNN VJG URQMGP YQTFU YKNN DG KPFKURGPUCDNG 6Q OGGV VJKU PGGF C ſXG[GCT PCVKQPCN RTQLGEV HQT TCKUKPI VJG VGEJPQNQIKECN NGXGN QH URGGEJ TGEQIPKVKQP CPF WPFGTUVCPFKPI EQOOGPEGF KP ,CRCP KP 6JG RTQLGEV HQEWUGU QP DWKNFKPI C NCTIGUECNG URQPVCPGQWU URGGEJ EQTRWU VQIGVJGT YKVJ CEQWU VKE CPF NKPIWKUVKE OQFGNKPI HQT URQPVCPGQWU URGGEJ TGEQIPKVKQP CPF UWOOCTK\CVKQP 'ZRGTKOGPVCN TGUWNVU UJQY VJCV CEQWUVKE CPF NCPIWCIG OQFGNKPI DCUGF QP VJG CE VWCN URQPVCPGQWU URGGEJ EQTRWU KU HCT OQTG GHHGEVKXG VJCP OQFGNKPI DCUGF QP TGCF URGGEJ +V KU CNUQ UJQYP VJCV VJG RTQRQUGF CWVQOCVKE URGGEJ UWOOCTK\CVKQP OGVJQF GHHGEVKXGN[ GZVTCEVU TGNCVKXGN[ KORQTVCPV KPHQTOCVKQP CPF TGOQXGU TGFWPFCPV CPF KT TGNGXCPV KPHQTOCVKQP +V YKNN DGEQOG KORQTVCPV VQ WUG CTVKEWNCVQT[ CPF RGTEGRVWCN EQPUVTCKPVU VQ UQNXG XCT KQWU HWPFCOGPVCN RTQDNGOU KP URQPVCPGQWU URGGEJ OQFGNKPI +V YKNN CNUQ DGEQOG ETWEKCN VQ CPCN[\G VJG HWPEVKQP YKVJKP VJG JWOCP DTCKP VJCV KU JQY JWOCPDGKPIU CTG WPFGTUVCPFKPI URGGEJ CPF VJG HWPEVKQP OWUV VJGP DG TGCNK\GF WUKPI GPIKPGGTKPI OQFGNU 4GUGCTEJ UJQWNF KPXGUVKICVG JQY OGCPKPIU EQPXG[GF D[ URGGEJ CTG WPFGT UVQQF
References =? $* ,WCPI CPF 5 (WTWK ő#WVQOCVKE TGEQIPKVKQP CPF WPFGTUVCPFKPI QH URQMGP NCPIWCIG # ſTUV UVGR VQYCTFU PCVWTCN JWOCPOCEJKPG EQOOWPKECVKQPŒ 2TQE +''' RR =? . 4 4CDKPGT CPF $ * ,WCPI (WPFCOGPVCNU QH 5RGGEJ 4GEQIPKVKQP 0GY ,GTUG[ 2TGPVKEG*CNN +PE =? 5 (WTWK &KIKVCN 5RGGEJ 2TQEGUUKPI 5[PVJGUKU CPF 4GEQIPKVKQP PF 'FKVKQP 0GY ;QTM /CTEGN &GMMGT =? * 0G[ ő%QTRWUDCUGF UVCVKUVKECN OGVJQFU KP URGGEJ CPF NCPIWCIG RTQEGUU KPIŒ KP %QTRWUDCUGF /GVJQFU KP .CPIWCIG CPF 5RGGEJ 2TQEGUUKPI ;QWPI 5 CPF $NQQVJQQHV ) 'F RR =? $ * ,WCPI ő(TQO URGGEJ TGEQIPKVKQP VQ WPFGTUVCPFKPI 5JKHVKPI RCTCFKIO VQ CEJKGXG PCVWTCN JWOCPOCEJKPG EQOOWPKECVKQPŒ 2TQE VJ +%# CPF VJ /GGVKPI #5# RR =? , , )QFHTG[ ' % *QNNKOCP CPF , /E&CPKGN ő5YKVEJDQCTF 6GNGRJQPG URGGEJ EQTRWU HQT TGUGCTEJ CPF FGXGNQROGPVŒ 2TQE +%#552 RR +
=? # 9CKDGN GV CN ő#FXCPEGU KP CWVQOCVKE OGGVKPI TGEQTF ETGCVKQP CPF CEEGUUŒ 2TQE +%#552 RR +
E\&5&3UHVV//&
=? # + 4WFPKEM[ ő*WD $WUKPGUU DTQCFECUV PGYUŒ 2TQE *# 5RGGEJ 4GEQIPKVKQP 9QTMUJQR RR =? - 1JVUWMK GV CN ő4GEGPV CFXCPEGU KP ,CRCPGUG DTQCFECUV PGYU VTCPUETKRVKQPŒ 2TQE 'WTQURGGEJ RR =? 6 +OCK GV CN ő2TQITGUUKXG RCUU FGEQFGT HQT TGCNVKOG DTQCFECUV PGYU ECR VKQPKPIŒ 2TQE +%#552 RR +++ =? ,. )CWXCKP . .COGN ; FG -CTECFKQ CPF ) #FFC ő6TCPUETKRVKQP CPF KPFGZCVKQP QH DTQCFECUV FCVCŒ 2TQE +%#552 RR +++ =? , *KTUEJDGTI / $CEEJKCPK & *KPFNG 2 +UGPJQWT # 4QUGPDGTI . 5VCTM . 5VGCF 5 9JKVVCMGT CPF )
E\&5&3UHVV//&
=? * #MCKMG ő+PHQTOCVKQP VJGQT[ CPF CP GZVGPUKQP QH VJG OCZKOWO NKMGNKJQQF RTKPEKRNGŒ 2TQE +5+6 $ 0 2GVTQX CPF ( %UCMK GFU #MCFGOKCK -KCFQ $W FCRGUV RR =? , <JGPI * (TCPEQ CPF ( 9GPI ő9QTFNGXGN TCVG QH URGGEJ OQFGNKPI WUKPI TCVGURGEKſE RJQPGU CPF RTQPWPEKCVKQPUŒ 2TQE +%#552 RR
=? % (WIGP CPF + 4QIKPC ő+PVGITCVKPI F[PCOKE URGGEJ OQFCNKVKGU KPVQ EQPVGZV FGEKUKQP VTGGUŒ 2TQE +%#552 RR =? 1 5GICYC 6 6CMGFC CPF ( +VCMWTC ő%QPVKPWQWU URGGEJ TGEQIPKVKQP YKVJQWV GPFRQKPV FGVGEVKQPŒ 2TQE +%#552 RR + =? 6 -CYCJCTC * 0CPLQ CPF 5 (WTWK ő#WVQOCVKE VTCPUETKRVKQP QH URQPVCPGQWU NGEVWTG URGGEJŒ 2TQE #547 =? < -NCWU ő#WVQOCVKE IGPGTCVKQP QH EQPEKUG UWOOCTKGU QH URQMGP FKCNQIWGU KP WPTGUVTKEVGF FQOCKPUŒ 2TQE 5+)+4 0GY 1TNGCPU =? 4 8CNGP\C GV CN ő5WOOCTK\CVKQP QH URQMGP CWFKQ VJTQWIJ KPHQTOCVKQP GZ VTCEVKQPŒ 2TQE '5%# 9QTMUJQR QP #EEGUUKPI +PHQTOCVKQP KP 5RQMGP #WFKQ %CODTKFIG RR =? % *QTK CPF 5 (WTWK ő#WVQOCVKE URGGEJ UWOOCTK\CVKQP DCUGF QP YQTF UKI PKſECPEG CPF NKPIWKUVKE NKMGNKJQQFŒ 2TQE +''' +PV %QPH #EQWUV 5RGGEJ 5KIPCN 2TQEGUU +UVCPDWN RR =? % *QTK CPF 5 (WTWK ő+ORTQXGOGPVU KP CWVQOCVKE URGGEJ UWOOCTK\CVKQP CPF GXCNWCVKQP OGVJQFUŒ 2TQE +PV %QPH 5RQMGP .CPIWCIG 2TQEGUUKPI $GKLKPI RR +8 =? % *QTK CPF 5 (WTWK ő#FXCPEGU KP CWVQOCVKE URGGEJ UWOOCTK\CVKQPŒ 2TQE 'WTQURGGEJ #CNDQTI RR +++ =? - 1JVUWMK 5 (WTWK # +YCUCMK CPF 0 5CMWTCK ő/GUUCIGFTKXGP URGGEJ TGEQIPKVKQP CPF VQRKEYQTF GZVTCEVKQPŒ 2TQE +''' +PV %QPH #EQWUV 5RGGEJ 5KIPCN 2TQEGUU 2JQGPKZ RR =? 5 (WTWK ő6QYCTFU VJG 7NVKOCVG 5[PVJGUKU4GEQIPKVKQP 5[UVGOŒ KP 8QKEG %QOOWPKECVKQP DGVYGGP *WOCPU CPF /CEJKPGU GFU D[ 4QG & $ CPF 9KNRQP , ) 9CUJKPIVQP & % 0CVKQPCN #ECFGO[ 2TGUU RR =? 5 (WTWK ő1P VJG TQNG QH URGEVTCN VTCPUKVKQP HQT URGGEJ RGTEGRVKQPŒ , #EQWUV 5QE #O RR =? 5 (WTWK CPF / #MCIK ő1P VJG TQNG QH URGEVTCN VTCPUKVKQP KP RJQPGOG RGTEGR VKQP CPF KVU OQFGNKPIŒ 2TQE VJ +%# 6QTQPVQ %CPCFC # =? 5 (WTWK ő5RGCMGTKPFGRGPFGPV KUQNCVGF YQTF TGEQIPKVKQP WUKPI F[PCOKE HGC VWTGU QH URGGEJ URGEVTWOŒ +''' 6TCPU #EQWUV 5RGGEJ 5KIPCN 2TQEGUU #552 RR
E\&5&3UHVV//&
7 Speaker Authentication Qi Li£ and Biing-Hwang Juang Ý £ Bell Labs; Ý Avaya Labs Research
CONTENTS
+PVTQFWEVKQP 2CVVGTP 4GEQIPKVKQP KP 5RGCMGT #WVJGPVKECVKQP 5RGCMGT 8GTKſECVKQP 5[UVGO 8GTDCN +PHQTOCVKQP 8GTKſECVKQP 5RGCMGT #WVJGPVKECVKQP D[ %QODKPKPI 58 CPF 8+8 5WOOCT[ 4GHGTGPEGU
#OQPI XCTKQWU WUGT CWVJGPVKECVKQP VGEJPKSWGU URGCMGT CWVJGPVKECVKQP EQPEGTPU YKVJ CWVJGPVKECVKPI C RGTUQPŏU KFGPVKV[ XKC XQKEG 6JGTG CTG VYQ CRRTQCEJGU VQ URGCMGT CWVJGPVKECVKQP URGCMGT XGTKſECVKQP 58 CPF XGTDCN KPHQTOCVKQP XGTKſECVKQP 8+8 6JG 58 CRRTQCEJ CVVGORVU VQ XGTKH[ C URGCMGTŏU KFGPVKV[ DCUGF QP JKUJGT XQKEG EJCT CEVGTKUVKEU YJKNG VJG 8+8 CRRTQCEJ XGTKſGU C URGCMGTŏU KFGPVKV[ VJTQWIJ XGTKſECVKQP QH VJG EQPVGPV QH JKUJGT WVVGTCPEG U +P VJKU EJCRVGT YG ſTUV KPVTQFWEG VJG TGNCVGF RCVVGTP TGEQIPKVKQP CPF XGTKſECVKQP VGEJPKSWGU CPF VJGP RTGUGPV CP 58 U[UVGO C 8+8 U[UVGO CPF C EQODKPGF U[UVGO YKVJ DQVJ 58 CPF 8+8 HQT EQPXGPKGPEG CPF RGTHQTOCPEG KORTQXGOGPV 6JGUG U[UVGOU CTG TGCF[ HQT TGCNYQTNF CRRNKECVKQPU
7.1 Introduction 6Q GPUWTG VJG UGEWTKV[ QH CPF RTQRGT CEEGUU VQ RTKXCVG KPHQTOCVKQP KORQTVCPV VTCPUCE VKQPU CPF VJG EQORWVGT CPF EQOOWPKECVKQP PGVYQTMU RCUUYQTFU QT RGTUQPCN KFGP VKſECVKQP PWODGTU 2+0 JCXG DGGP WUGF GZVGPUKXGN[ KP QWT FCKN[ NKHG 6Q HWTVJGT GPJCPEG VJG NGXGN QH UGEWTKV[ CU YGNN CU EQPXGPKGPEG DKQOGVTKE HGCVWTGU UWEJ CU UKIPCVWTG ſPIGTRTKPV JCPF UJCRG G[G KTKU CPF XQKEG JCXG CNUQ DGGP EQPUKFGTGF #OQPI CNN DKQOGVTKE HGCVWTGU C RGTUQPŏU XQKEG KU VJG OQUV EQPXGPKGPV QPG HQT RGT
E\&5&3UHVV//&
5RGCMGT#WVJGPVKECVKQP
5RGCMGT4GEQIPKVKQP
#WVJGPVKECVKQPD[ URGGEJEJCTCEVGTKUVKEU
5RGCMGT 8GTKHKECVKQP
8GTDCN+PHQTOCVKQP 8GTKHKECVKQP
#WVJGPVKECVKQPD[ XGTDCNEQPVGPV
5RGCMGT +FGPVKHKECVKQP
FIGURE 7.1 Speaker authentication approaches.
UQPCN KFGPVKſECVKQP RWTRQUGU DGECWUG KV KU GCU[ VQ RTQFWEG ECRVWTG QT VTCPUOKV QXGT VJG WDKSWKVQWU VGNGRJQPG PGVYQTM +V CNUQ ECP DG UWRRQTVGF YKVJ GZKUVKPI UGTXKEGU YKVJQWV TGSWKTKPI URGEKCN FGXKEGU Speaker authentication CU CP CRRNKECVKQP QH RCV VGTP TGEQIPKVKQP KU VJG RTQEGUU QH CWVJGPVKECVKPI C WUGT XKC JKUJGT URQMGP KPRWV *QY VQ CWVQOCVG URGGFKN[ VJG CWVJGPVKECVKQP RTQEGFWTG CPF CEJKGXG C JKIJ CEEWTCE[ RQUGU C UGTKQWU VGEJPKECN EJCNNGPIG VQ URGGEJ TGUGCTEJGTU #U UJQYP KP (KI VJG CRRTQCEJ VQ URGCMGT CWVJGPVKECVKQP ECP DG ECVGIQTK\GF KPVQ VYQ ITQWRU QPG WUGU C URGCMGTŏU XQKEG EJCTCEVGTKUVKEU YJKEJ NGCFU VQ URGCMGT TGEQI PKVKQP CPF VJG QVJGT HQEWUGU QP VJG XGTDCN EQPVGPV QH VJG URQMGP WVVGTCPEG YJKEJ NGCFU VQ XGTDCN KPHQTOCVKQP XGTKſECVKQP 6JGUG VYQ VGEJPKSWGU ECP CNUQ DG EQODKPGF VQ RTQXKFG CP GPJCPEGF U[UVGO CU KPFKECVGF D[ VJG FCUJGF NKPG
7.1.1 Speaker Recognition and Verification Speaker recognition ECP DG HQTOWNCVGF KP VYQ QRGTCVKPI OQFGU URGCMGT XGTKſECVKQP CPF URGCMGT KFGPVKſECVKQP Speaker verification 58 KU VJG RTQEGUU QH XGTKH[KPI CP WPMPQYP URGCMGT YJGVJGT UJG KU VJG RGTUQP CU ENCKOGF KG C [GUPQ hypothesis testing RTQDNGO Speaker identification 5+& KU VJG RTQEGUU QH CUUQEKCVKPI CP WPMPQYP URGCMGT YKVJ C OGODGT KP C RTGTGIKUVGTGF MPQYP RQRWNCVKQP KG C OWNVKRNGEJQKEG classification RTQDNGO +P VJKU EJCRVGT YG YKNN HQEWU QP VJG VCUM QH URGCMGT XGTKſEC VKQP 5RGCMGT TGEQIPKVKQP CU QPG QH VJG XQKEG CWVJGPVKECVKQP VGEJPKSWGU JCU DGGP UVWFKGF HQT UGXGTCN FGECFGU = ? # V[RKECN 58 U[UVGO KU UJQYP KP (KI YJKEJ JCU VYQ QRGTCVKPI UEGPCTKQU GPTQNNOGPV CPF VGUV UGUUKQPU # URGCMGT PGGFU VQ GPTQNN ſTUV DGHQTG UJG ECP WUG VJG U[UVGO +P VJG GPTQNNOGPV UGUUKQP VJG WUGTŏU KFGPVKV[ UWEJ CU CP CEEQWPV PWODGT VQIGVJGT YKVJ C RCUURJTCUG UWEJ CU C FKIKV UVTKPI QT C MG[ RJTCUG NKMG őQRGP UGUCOGŒ UJQYP KP VJG ſIWTG KU CUUKIPGF VQ VJG
E\&5&3UHVV//&
6TCKPKPI7VVGTCPEGU 1RGP5GUCOG 1RGP5GUCOG 1RGP5GUCOG
/QFGN 6TCKPKPI
5RGCMGT&GRGPFGPV/QFGN
&CVCDCUG 'PTQNNOGPV5GUUKQP 6GUV5GUUKQP
+FGPVKV[%NCKO 6GUV7VVGTCPEG 1RGP5GUCOG
5RGCMGT 8GTKHKGT
5EQTGU
FIGURE 7.2 A speaker verification system.
URGCMGT 6JG U[UVGO VJGP RTQORVU VJG URGCMGT VQ UC[ VJG RCUURJTCUG UGXGTCN VKOGU VQ CNNQY VTCKPKPI QT EQPUVTWEVKPI QH C URGCMGTFGRGPFGPV 5& OQFGN VJCV TGIKUVGTU VJG URGCMGTŏU URGGEJ EJCTCEVGTKUVKEU 6JG FKIKV UVTKPI ECP DG VJG UCOG CU VJG CEEQWPV PWODGT CPF VJG MG[ RJTCUG ECP DG UGNGEVGF D[ VJG WUGT UQ KV KU GCU[ VQ TGOGODGT #P GPTQNNGF URGCMGT ECP WUG VJG XGTKſECVKQP U[UVGO KP C HWVWTG VGUV 5KOKNCT RTQEGFWTG CRRNKGU KP VJG ECUG QH URGCMGT KFGPVKſECVKQP 6JGUG UEJGOGU CTG UQOGVKOGU TGHGTTGF VQ CU direct method CU VJG[ WUG VJG VCNMGTŏU URGGEJ EJCTCEVGTKUVKEU VQ KPHGT QT XGTKH[ VJG VCNMGTŏU KFGPVKV[ directly +P C VGUV UGUUKQP VJG WUGT ſTUV ENCKOU JKUJGT KFGPVKV[ D[ GPVGTKPI QT URGCMKPI VJG KFGPVKV[ KPHQTOCVKQP 6JG U[UVGO VJGP RTQORVU VJG URGCMGT VQ UC[ VJG RCUURJTCUG 6JG RCUURJTCUG WVVGTCPEG KU EQORCTGF CICKPUV VJG UVQTGF 5& OQFGN 6JG URGCMGT KU CEEGRVGF KH VJG XGTKſECVKQP UEQTG GZEGGFU C RTGUGV VJTGUJQNF QVJGTYKUG VJG URGCMGT KU TGLGEVGF 0QVG VJCV VJG RCUURJTCUG OC[ QT OC[ PQV DG MGRV KP UGETGV 9JGP VJG RCUURJTCUGU CTG VJG UCOG KP VTCKPKPI CPF VGUVKPI VJG U[UVGO KU ECNNGF C fixed pass-phrase system (TGSWGPVN[ C UJQTV RJTCUG QT C EQPPGEVGFFKIKV UGSWGPEG UWEJ CU C VGNGRJQPG QT CEEQWPV PWODGT KU EJQUGP CU VJG ſZGF RCUURJTCUG 7UKPI C FKIKV UVTKPI HQT C RCUURJTCUG JCU C FKUVKPEVKXG FKHHGTGPEG HTQO QVJGT PQPFKIKV EJQKEGU 6JG JKIJ RGTHQTOCPEG QH EWTTGPV EQPPGEVGF FKIKV URGGEJ TGEQIPKVKQP U[UVGOU CPF GO DGFFGF GTTQT EQTTGEVKPI RQUUKDKNKVKGU QH FKIKV UVTKPIU OCMG KV HGCUKDNG VJCV VJG KFGPVKV[ ENCKO ECP DG OCFG XKC URQMGP TCVJGT VJCP MG[KP KPRWV = ? +H UWEJ CP QRVKQP KU KPUVCNNGF VJG URQMGP FKIKV UVTKPI KU ſTUV TGEQIPK\GF D[ CP CWVQOCVKE URGGEJ TGE QIPK\GT #54 CPF VJG UVCPFCTF XGTKſECVKQP RTQEGFWTG VJGP HQNNQYU WUKPI VJG UCOG FKIKV UVTKPI 1DXKQWUN[ UWEEGUUHWN XGTKſECVKQP QH C URGCMGT TGNKGU WRQP C EQTTGEV TGEQIPKVKQP QH VJG KPRWV FKIKV UVTKPI # UGEWTKV[ EQPEGTP OC[ DG TCKUGF CDQWV WUKPI ſZGF RCUURJTCUGU UKPEG C URQMGP RCUU RJTCUG ECP DG VCRGTGEQTFGF D[ KORQUVQTU CPF WUGF KP NCVGT VTKCNU VQ IGV CEEGUU VQ VJG U[UVGO # VGZVRTQORVGF 58 U[UVGO JCU DGGP RTQRQUGF VQ EKTEWOXGPV UWEJ C RTQD NGO # text-prompted system WUGU C UGV QH URGCMGTFGRGPFGPV YQTF QT UWDYQTF OQF
E\&5&3UHVV//&
GNU RQUUKDN[ HQT C UOCNN XQECDWNCT[ UWEJ CU VJG FKIKVU 6JGUG OQFGNU CTG GORNQ[GF CU VJG DWKNFKPI DNQEMU HQT EQPUVTWEVKPI VJG OQFGNU HQT VJG RTQORVGF WVVGTCPEG YJKEJ OC[ QT OC[ PQV DG RCTV QH VJG VTCKPKPI OCVGTKCN 9JGP VJG WUGT VTKGU VQ CEEGUU VJG U[UVGO VJG U[UVGO RTQORVU VJG WUGT VQ WVVGT C TCPFQON[ RKEMGF UGSWGPEG QH YQTFU KP VJG XQECDWNCT[ 6JG YQTF UGSWGPEG KU CNKIPGF YKVJ VJG RTGVTCKPGF YQTF OQFGNU CPF C XGTKſECVKQP FGEKUKQP KU OCFG DCUGF WRQP VJG GXCNWCVGF NKMGNKJQQF UEQTG %QORCTGF VQ C ſZGFRJTCUG U[UVGO UWEJ C VGZVRTQORVGF U[UVGO PQTOCNN[ PGGFU NQPIGT GPTQNN OGPV VKOG KP QTFGT VQ EQNNGEV GPQWIJ FCVC VQ VTCKP VJG 5& YQTF QT UWDYQTF OQFGNU 6JG RGTHQTOCPEG QH C VGZVRTQORVGF U[UVGO KU KP IGPGTCN PQV CU JKIJ CU VJCV QH C ſZGFRJTCUG U[UVGO 6JKU KU FWG VQ VJG HCEV VJCV VJG RJTCUG OQFGN EQPUVTWEVGF HTQO EQPECVGPCVKPI GNGOGPVCT[ YQTF QT UWDYQTF OQFGNU KU WUWCNN[ PQV CU CEEWTCVG CU VJCV FKTGEVN[ VTCKPGF HTQO VJG RJTCUG WVVGTCPEG KP C ſZGFRJTCUG U[UVGO &GVCKNU QP C VGZVRTQORVGF U[UVGO CPF KVU RGTHQTOCPEG ECP DG HQWPF GI =? 6JG CDQXG U[UVGOU CTG ECNNGF VGZVFGRGPFGPV QT VGZVEQPUVTCKPGF 58 U[UVGOU DG ECWUG VJG KPRWV WVVGTCPEG KU EQPUVTCKPGF GKVJGT D[ C ſZGF RJTCUG QT D[ C ſZGF XQ ECDWNCT[ # XGTKſECVKQP U[UVGO ECP CNUQ DG VGZVKPFGRGPFGPV +P C text-independent SV system C URGCMGTŏU OQFGN KU VTCKPGF QP VJG IGPGTCN URGGEJ EJCTCEVGTKUVKEU QH VJG RGTUQPŏU XQKEG = ? 1PEG UWEJ C OQFGN KU VTCKPGF VJG URGCMGT ECP DG XGTKſGF TGICTFNGUU QH VJG WPFGTN[KPI VGZV QH VJG URQMGP KPRWV 5WEJ C U[UVGO JCU YKFG CR RNKECVKQPU KP OQPKVQTKPI CRRNKECVKQPU HQT XGTKH[KPI C URGCMGT QP C EQPVKPWQWU DCUKU +P QTFGT VQ EJCTCEVGTK\G C URGCMGTŏU IGPGTCN XQKEG RCVVGTP YKVJQWV C VGZV EQPUVTCKPV YG PQTOCNN[ PGGF C NCTIG COQWPV QH RJQPGVKECNN[ QT CEQWUVKECNN[ TKEJ VTCKPKPI FCVC KP VJG GPTQNNOGPV RTQEGFWTG #NUQ YKVJQWV VJG VGZV QT NGZKECN EQPUVTCKPV NQPIGT VGUV WVVGTCPEGU CTG WUWCNN[ PGGFGF VQ OCKPVCKP C UCVKUHCEVQT[ 58 RGTHQTOCPEG 9KVJQWV C NCTIG VTCKPKPI UGV CPF NQPI VGUV WVVGTCPEGU VJG RGTHQTOCPEG QH C VGZVKPFGRGPFGPV U[UVGO KU WUWCNN[ KPHGTKQT VQ VJCV QH C VGZVFGRGPFGPV U[UVGO +P GXCNWCVKPI CP 58 U[UVGO KH KV KU DQVJ VTCKPGF CPF VGUVGF D[ VJG UCOG UGV QH URGCMGTU KV KU ECNNGF C closed test QVJGTYKUG CP open test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
7.1.2 Verbal Information Verification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
E\&5&3UHVV//&
ŎŎ+PYJKEJ[GCTYGTG[QWDQTP!ŏŏ )GVCPFXGTKH[VJGCPUYGTWVVGTCPEG %QTTGEV
9TQPI
ŎŎ+PYJKEJEKV[UVCVGFKF[QWITQYWR!ŏŏ
4GLGEVKQP
)GVCPFXGTKH[VJGCPUYGTWVVGTCPEG %QTTGEV
9TQPI
ŎŎ/C[+JCXG[QWTVGNGRJQPGPWODGTRNGCUG!ŏŏ
4GLGEVKQP
)GVCPFXGTKH[VJGCPUYGTWVVGTCPEG %QTTGEV
#EEGRVCPEG QPWVVGTCPEGU
9TQPI
4GLGEVKQP
FIGURE 7.3 An example of verbal information verification by asking sequential questions. (Similar sequential tests can also be applied in speaker verification and other biometric or multi-modality verification.)
VGUVKPI XQKEG OC[ EQOG HTQO FKHHGTGPV VGNGRJQPG JCPFUGVU CPF PGVYQTMU CEQWUVKE OKUOCVEJ DGVYGGP VJG VTCKPKPI CPF VGUVKPI GPXKTQPOGPVU OC[ QEEWT 6JG 5& OQFGNU VTCKPGF QP VJG FCVC EQNNGEVGF KP CP GPTQNNOGPV UGUUKQP OC[ PQV RGTHQTO YGNN YJGP VJG VGUV UGUUKQP KU KP C FKHHGTGPV GPXKTQPOGPV QT XKC C FKHHGTGPV VTCPUOKUUKQP EJCPPGN 6JG OKUOCVEJ UKIPKſECPVN[ CHHGEVU VJG 58 RGTHQTOCPEG 6JKU KU C UKIPKſECPV FTCYDCEM QH VJG FKTGEV OGVJQF KP YJKEJ VJG TQDWUVPGUU KP EQORCTCVKXG GXCNWCVKQP KU FKHſEWNV VQ GPUWTG #NVGTPCVKXGN[ KP NKIJV QH VJG RTQITGUU KP OQFGNKPI HQT URGGEJ TGEQIPKVKQP VJG EQPEGRV CPF CNIQTKVJO QH 8+8 YCU RTQRQUGF =? VQ VCMG CFXCPVCIG QH VJG FKHHGTGPV EJCTCEVGTKUVKE HQEWU QP VJG URGGEJ UKIPCN PCOGN[ VJCV QH VJG URGCMGT XU VJCV QH VJG URGGEJ 6JG 8+8 OGVJQF KU VJG RTQEGUU QH XGTKH[KPI URQMGP WVVGTCPEGU CICKPUV VJG KPHQTOC VKQP UVQTGF KP C IKXGP RGTUQPCN FCVC RTQſNG # 8+8 U[UVGO OC[ WUG C FKCNQIWG RTQ EGFWTG VQ XGTKH[ C WUGT D[ CUMKPI SWGUVKQPU #P GZCORNG QH C 8+8 U[UVGO KU UJQYP KP (KI +V KU UKOKNCT VQ C V[RKECN VGNGDCPMKPI RTQEGFWTG CHVGT CP CEEQWPV PWODGT KU RTQXKFGF VJG QRGTCVQT XGTKſGU VJG WUGT D[ CUMKPI UQOG RGTUQPCN KPHQTOCVKQP UWEJ CU OQVJGTŏU OCKFGP PCOG DKTVJ FCVG CFFTGUU JQOG VGNGRJQPG PWODGT GVE 6JG WUGT OWUV RTQXKFG CPUYGTU VQ VJG SWGUVKQPU EQTTGEVN[ KP QTFGT VQ ICKP CEEGUU VQ JKUJGT CE EQWPV CPF UGTXKEGU +P VJKU OCPPGT C VCNMGTŏU KFGPVKV[ KU GODGFFGF KP VJG MPQYNGFIG UJG JCU VQYCTFU UQOG RCTVKEWNCT SWGUVKQPU CPF VJWU QPG QHVGP EQPUKFGTU 8+8 CP indirect method 6Q CWVQOCVG VJG YJQNG RTQEGFWTG VJG SWGUVKQPU ECP DG RTQORVGF D[ C VGZVVQURGGEJ U[UVGO 665 QT D[ RTGTGEQTFGF OGUUCIGU 6JG FKHHGTGPEG DGVYGGP URGCMGT TGEQIPKVKQP VJG FKTGEV OGVJQF CPF XGTDCN KPHQTOC
E\&5&3UHVV//&
VKQP XGTKſECVKQP VJG KPFKTGEV OGVJQF ECP DG HWTVJGT CFFTGUUGF KP VJG HQNNQYKPI VJTGG CURGEVU (KTUV KP C URGCMGT TGEQIPKVKQP U[UVGO GKVJGT HQT URGCMGT KFGPVKſECVKQP QT HQT URGCMGT XGTKſECVKQP YG PGGF VQ VTCKP URGCMGTFGRGPFGPV 5& OQFGNU YJKNG KP 8+8 YG WUWCNN[ WUG UVCVKUVKECN OQFGNU YKVJ CUUQEKCVGF CEQWUVKERJQPGVKE KFGPVKVKGU 5GEQPF C URGCMGT TGEQIPKVKQP U[UVGO PGGFU VQ GPTQNN C PGY WUGT CPF VQ VTCKP VJG 5& OQFGN YJKNG C 8+8 U[UVGO FQGU PQV TGSWKTG XQKEG GPTQNNOGPV +PUVGCF C WUGTŏU RGTUQPCN FCVC RTQſNG KU ETGCVGF YJGP VJG WUGTŏU CEEQWPV KU UGV WR (KPCNN[ KP URGCMGT TGEQIPKVKQP VJG U[UVGO JCU VJG CDKNKV[ VQ TGLGEV CP KORQUVGT GXGP YJGP VJG KPRWV WV VGTCPEG EQPVCKPU C NGIKVKOCVG RCUURJTCUG KH VJG WVVGTCPEG KPFGGF HCKNU VQ OCVEJ VJG RTGVTCKPGF 5& OQFGN +P 8+8 KV KU UQNGN[ VJG WUGTŏU TGURQPUKDKNKV[ VQ RTQVGEV JKU QT JGT RGTUQPCN KPHQTOCVKQP DGECWUG PQ URGCMGTURGEKſE XQKEG EJCTCEVGTKUVKEU CTG WUGF KP VJG XGTKſECVKQP RTQEGUU +P TGCN CRRNKECVKQPU VJGTG CTG UGXGTCN YC[U VQ EKTEWOXGPV VJG UKVWCVKQP KP YJKEJ CP KORQUVQT WUGU C URGCMGTŏU RGTUQPCN KPHQTOCVKQP QDVCKPGF HTQO GCXGUFTQRRKPI C RCTVKEWNCT UGUUKQP # 8+8 U[UVGO ECP CUM HQT KPHQTOCVKQP VJCV OC[ PQV DG C EQPUVCPV HTQO QPG UGUUKQP VQ CPQVJGT GI VJG COQWPV QT FCVG QH VJG NCUV FGRQUKV QT C UWDUGV QH VJG TGIKUVGTGF RGTUQPCN KPHQTOCVKQP KG C PWODGT QH TCPFQON[ UGNGEVGF KPHQTOCVKQP ſGNFU KP VJG RGTUQPCN FCVC RTQſNG (WTVJGTOQTG CU YG CTG IQKPI VQ RTGUGPV KP 5GEVKQP C 8+8 U[UVGO ECP DG OKITCVGF VQ CP 58 U[UVGO CU KPFKECVGF D[ VJG FCUJ NKPG KP (KI +P RCTVKEWNCT 8+8 ECP DG WUGF VQ HCEKNKVCVG CWVQOCVKE GPTQNNOGPV HQT 58
7.2 Pattern Recognition in Speaker Authentication +P VJKU UGEVKQP YG TGXKGY RCVVGTP TGEQIPKVKQP VGEJPKSWGU KP URGCMGT CWVJGPVKECVKQP 5VCTVKPI YKVJ VJG $C[GUKCP FGEKUKQP VJGQT[ YG KPVTQFWEG VJG UVCVKUVKECN OQFGNKPI CR RTQCEJ HQT UVCVKQPCT[ CPF PQPUVCVKQPCT[ RTQEGUUGU CNIQTKVJOU HQT URGGEJ UGIOGPVC VKQP CPF J[RQVJGUKU VGUVKPI
7.2.1 Bayesian Decision Theory
+P CP ENCUU TGEQIPKVKQP RTQDNGO YG CTG IKXGP CP QDUGTXCVKQP QT C HGCVWTG XGEVQT Ó KP C FKOGPUKQPCN 'WENKFGCP URCEG Ê CPF C UGV QH ENCUUGU FGUKIPCVGF CU ½ ¾ CPF CUMGF VQ OCMG C FGEKUKQP VQ ENCUUKH[ Ó KPVQ UC[ ENCUU YJGTG QPG ENCUU ECP DG QPG URGCMGT QT QPG CEQWUVKE WPKV 9G FGPQVG VJKU CU CP CEVKQP $[ $C[GU HQTOWNC VJG RTQDCDKNKV[ QH DGKPI ENCUU IKXGP Ó KU VJG RQUVGTKQT QT a posteriori RTQDCDKNKV[
Ó Ó Ó
E\&5&3UHVV//&
YJGTG KU VJG EQPFKVKQPCN RTQDCDKNKV[ KU RTKQT RTQDCDKNKV[ CPF
ECP DG XKGYGF CU C UECNG HCEVQT VJCV IWCTCPVGGU VJCV VJG RQUVGTKQT RTQDCDKNKVKGU UWO VQ QPG .GV DG VJG NQUU HWPEVKQP FGUETKDKPI VJG NQUU KPEWTTGF HQT VCMKPI CEVKQP YJGP VJG VTWG ENCUU KU 6JG GZRGEVGF NQUU QT TKUM CUUQEKCVGF YKVJ VCMKPI CEVKQP KU
6JKU NGCFU VQ VJG Bayes decision rule 6Q OKPKOK\G VJG QXGTCNN TKUM EQORWVG VJG CDQXG TKUM HQT CPF VJGP UGNGEV VJG CEVKQP UWEJ VJCV KU OKPK OWO (QT URGCMGT CWVJGPVKECVKQP YG CTG KPVGTGUVGF KP VJG \GTQQPG NQUU HWPEVKQP
+V CUUKIPU PQ NQUU VQ C EQTTGEV FGEKUKQP CPF C WPKV NQUU VQ CP GTTQT GSWKXCNGPV VQ EQWPVKPI VJG GTTQTU 6JG TKUM VQ VJKU URGEKſE NQUU HWPEVKQP KU
6JWU VQ OKPKOK\G VJG TKUM QT GTTQT TCVG YG VCMG CEVKQP VJCV OCZKOK\GU VJG RQUVG TKQT RTQDCDKNKV[ 6CMG CEVKQP YJGTG
5KPEG VJG GZRGEVGF XCNWG QH VJKU NQUU HWPEVKQP KU GSWKXCNGPV VQ GTTQT TCVG VJKU KU CNUQ ECNNGF OKPKOWOGTTQTTCVG ENCUUKſECVKQP =? 4GECNNKPI VJG $C[GU HQTOWNC KP 'S
YJGP VJG FGPUKV[ JCU DGGP GUVKOCVGF HQT CNN ENCUUGU CPF VJG RTKQT RTQD CDKNKVKGU CTG MPQYP YG ECP TGYTKVG VJG CDQXG FGEKUKQP TWNG CU 6CMG CEVKQP YJGTG
5Q HCT YG QPN[ EQPUKFGT VJG ECUG QH C UKPING QDUGTXCVKQP QT HGCVWTG XGEVQT +P URGCMGT CWVJGPVKECVKQP YG CNYC[U GPEQWPVGT QT GORNQ[ C UGSWGPEG QH QDUGTXCVKQPU YJGTG KU VJG VQVCN PWODGT QH QDUGTXCVKQPU #HVGT URGGEJ UGIOGP VCVKQP YJKEJ YKNN DG FKUEWUUGF NCVGT YG CUUWOG VJCV FWTKPI C UJQTV VKOG RGTKQF
E\&5&3UHVV//&
VJGUG UGSWGPVKCN QDUGTXCVKQPU CTG RTQFWEGF D[ VJG UCOG URGCMGT CPF VJG[ DGNQPI VQ VJG UCOG CEQWUVKE ENCUU QT WPKV UC[ (WTVJGTOQTG KH YG CUUWOG VJCV VJG QDUGTXC VKQPU CTG KPFGRGPFGPV CPF KFGPVKECNN[ FKUVTKDWVGF KKF VJG LQKPV RQUVGTKQT RTQDCDKN KV[ KU OGTGN[ VJG RTQFWEV QH VJG EQORQPGPV RTQDCDKNKVKGU
(TQO 'S VJG FGEKUKQP TWNG HQT VJG EQORQWPF FGEKUKQP RTQDNGO KU
+P RTCEVKEG VJG FGEKUKQP KU WUWCNN[ DCUGF QP VJG NQI NKMGNKJQQF UEQTG
7.2.2 Stochastic Models for Stationary Process #U FKUEWUUGF CDQXG VJG FGEKUKQP QP CWVJGPVKECVKQP KU OCFG D[ EQORWVKPI VJG NKMGNK JQQF DCUGF QP VJG RTQDCDKNKV[ FGPUKV[ HWPEVKQPU pdfU QH VJG HGCVWTG XGEVQT 2CTCO GVGTU VJCV FGſPG VJGUG pdfU JCXG VQ DG GUVKOCVGF C RTKQTK 6JGTG CTG OCP[ OQFGN UVTWEVWTGU IGPGTCN GPQWIJ VQ EJCTCEVGTK\G C URGGEJ pdf *GTG YG HQEWU QP VJG )CWUUKCP OKZVWTG OQFGN )// YJKEJ KU FGſPGF CU
YJGTG KU VJG )// HQT ENCUU KU C OKZVWTG YGKIJV YJKEJ OWUV UCVKUH[ VJG EQPUVTCKPV KU VJG VQVCN PWODGT QH OKZVWTG EQORQPGPVU CPF KU C )CWUUKCP FGPUKV[ HWPEVKQP
GZR
6
YJGTG CPF CTG VJG FKOGPUKQPCN OGCP XGEVQT CPF EQXCTKCPEG OCVTKZ QH VJG ŏVJ EQORQPGPV )KXGP C UGSWGPEG QH HGCVWTG XGEVQTU VJG )// RCTCOGVGTU ECP DG GUVKOCVGF KVGTC VKXGN[ WUKPI C JKNNENKODKPI CNIQTKVJO UWEJ CU VJG $CWO9GNEJ =? QT VJG GZRGEVCVKQP OCZKOK\CVKQP '/ CNIQTKVJO =? #U JCU DGGP RTQXGF VJG CNIQTKVJO GPUWTGU OQPQ VQPKE KPETGCUG KP VJG NQINKMGNKJQQF FWTKPI VJG KVGTCVKXG RTQEGFWTG WPVKN C ſZGFRQKPV UQNWVKQP KU TGCEJGF = ? +P OQUV CRRNKECVKQPU OQFGN RCTCOGVGT GUVKOCVKQP ECP
E\&5&3UHVV//&
DG CEEQORNKUJGF KP C HGY KVGTCVKQPU #V GCEJ UVGR QH VJG KVGTCVKQP VJG RCTCOGVGT GUVKOCVKQP HQTOWNCU HQT OKZVWTG CTG
YJGTG
6
1PG CRRNKECVKQP QH VJG CDQXG OQFGN KU EQPVGZVKPFGRGPFGPV URGCMGT KFGPVKſECVKQP YJGTG YG CUUWOG VJCV GCEJ URGCMGTŏU URGGEJ EJCTCEVGTKUVKEU OCPKHGUV QPN[ CEQWU VKECNN[ CPF KU TGRTGUGPVGF D[ QPG OQFGN ENCUU 9JGP C URQMGP WVVGTCPEG KU NQPI GPQWIJ KV KU TGCUQPCDNG VQ CUUWOG VJCV VJG CEQWUVKE EJCTCEVGTKUVKE KU KPFGRGPFGPV QH KVU EQPVGPV (QT C ITQWR QH URGCMGTU KP VJG GPTQNNOGPV RJCUG YG VTCKP )//ŏU WUKPI VJG TGGUVKOCVKQP CNIQTKVJO TGURGEVKXGN[ +P VJG VGUV RJCUG IKXGP CP QDUGTXCVKQP UGSWGPEG VJG QDLGEVKXG KU VQ ſPF KP VJG RTGUETKDGF URGCMGT RQRWNC VKQP VJG URGCMGT OQFGN VJCV CEJKGXGU VJG OCZKOWO RQUVGTKQT RTQDCDKNKV[ (TQO 'S
CUUWOG VJG RTKQT KU VJG UCOG HQT CNN URGCMGTU VJG FGEKUKQP TWNG KU
6CMG CEVKQP YJGTG
YJGTG KU VJG CEVKQP QH FGEKFKPI VJCV VJG QDUGTXCVKQP KU HTQO URGCMGT 6JG XGEVQT SWCPVK\CVKQP 83 OGVJQF =? KU CPQVJGT CRRTQCEJ VQ URGCMGT KFGPVKſEC VKQP +V WUGU C speaker-dependent EQFGDQQM VQ EJCTCEVGTK\G QH C URGCMGTŏU XQKEG 6JG EQFGDQQM KU IGPGTCVGF D[ C ENWUVGTKPI RTQEGFWTG DCUGF WRQP C RTGFGſPGF QDLGEVKXG FKUVQTVKQP OGCUWTG YJKEJ EQORWVGU VJG FKUUKOKNCTKV[ DGVYGGP CP[ VYQ IKXGP XGEVQTU =? 6JG EQFGDQQM ECP CNUQ DG EQPUKFGTGF CP KORNKEKV TGRTGUGPVCVKQP QH C OKZVWTG FKUVTKDWVKQP WUGF VQ FGUETKDG VJG UVCVKUVKECN RTQRGTVKGU QH VJG UQWTEG KG VJG RCTVKEW NCT VCNMGT +P VJG VGUV UGUUKQP KPRWV XGEVQTU HTQO VJG WPMPQYP VCNMGT CTG EQORCTGF YKVJ VJG PGCTGUV EQFGDQQM GPVT[ CPF VJG EQTTGURQPFKPI FKUVQTVKQPU CTG CEEWOWNCVGF VQ HQTO VJG DCUKU HQT C ENCUUKſECVKQP FGEKUKQP
7.2.3 Stochastic Models for Non-Stationary Process +P VJG CDQXG UGEVKQP UVCVKQPCTKV[ QH URGGEJ KU CUUWOGF CPF VJG OGVJQFU HQT VCNMGT KFGPVKſECVKQP FGUETKDGF VJGTGKP FQGU PQV OCMG WUG QH VJG VGORQTCN KPHQTOCVKQP QH URGGEJ +P OCP[ CRRNKECVKQPU VJG VGORQTCN KPHQTOCVKQP KU PGEGUUCT[ CPF KORQTVCPV
E\&5&3UHVV//&
C
C C
C00
C C
U
C
U
D
U
D
D
C00 U0 D0
FIGURE 7.4 Left-to-right hidden Markov model. KP OCMKPI C FGEKUKQP # OQTG RQYGTHWN OQFGN Ō JKFFGP /CTMQX OQFGN *// KU VJGP CRRNKGF VQ EJCTCEVGTK\G DQVJ VJG VGORQTCN UVTWEVWTG CPF VJG EQTTGURQPFKPI UVCVKUVKECN XCTKCVKQPU CNQPI VJG RCTCOGVGT VTCLGEVQT[ QH CP WVVGTCPEG +P URGGEJ CPF URGCMGT TGEQIPKVKQP CP *// KU VTCKPGF VQ TGRTGUGPV VJG CEQWUVKE RCV VGTP QH C UWDYQTF C YQTF QT C YJQNG RCUURJTCUG 6JGTG CTG OCP[ XCTKCPVU QH *//U 6JG UKORNGUV MKPF KU CP UVCVG NGHVVQTKIJV OQFGN YKVJQWV C UVCVGUMKR CU UJQYP KP (KIWTG 6JKU KU YKFGN[ WUGF KP URGCMGT CWVJGPVKECVKQP 6JG ſIWTG UJQYU C /CTMQX EJCKP YKVJ C UGSWGPEG QH UVCVGU TGRTGUGPVKPI VJG GXQNWVKQP QH URGGEJ UKI PCNU 9KVJKP GCEJ UVCVG C )CWUUKCP OKZVWTG OQFGN )// KU WUGF VQ EJCTCEVGTK\G VJG QDUGTXGF URGGEJ HGCVWTG XGEVQT CU C OWNVKXCTKCVG FKUVTKDWVKQP #P *// ECP DG EQORNGVGN[ EJCTCEVGTK\GF D[ VJTGG UGVU QH RCTCOGVGTU VJG UVCVG VTCPUKVKQP RTQDCDKNKVKGU VJG QDUGTXCVKQP FGPUKVKGU CPF VJG KPKVKCN UVCVG RTQDC DKNKVKGU CU UJQYP KP VJG HQNNQYKPI PQVCVKQP
YJGTG KU VJG VQVCN PWODGT QH UVCVGU )KXGP CP QDUGTXCVKQP UGSWGPEG VJG OQFGN RCTCOGVGTU QH ECP DG VTCKPGF D[ CP KVGTCVKXG OGVJQF VQ QRVK OK\G C RTGUETKDGF RGTHQTOCPEG ETKVGTKQP GI OCZKOWO NKMGNKJQQF +P RTCEVKEG VJG segmental K-mean CNIQTKVJO =? JCU DGGP YKFGN[ WUGF (QNNQYKPI OQFGN KPKVKCNK\C VKQP VJG QDUGTXCVKQP UGSWGPEG KU UGIOGPVGF KPVQ UVCVGU DCUGF QP VJG EWTTGPV OQFGN 6JGP YKVJKP GCEJ UVCVG C PGY )// KU VTCKPGF D[ VJG CDQXG '/ CNIQTKVJO VQ OCZKOK\G VJG NKMGNKJQQF 6JG PGY *// KU VJGP WUGF VQ TGUGIOGPV VJG QDUGTXC VKQP UGSWGPEG CPF TGGUVKOCVKQP QH OQFGN RCTCOGVGTU GPUWGU 6JG KVGTCVKXG RTQEGFWTG WUWCNN[ EQPXGTIGU KP C HGY KVGTCVKQPU 1VJGT VJCP VJG OCZKOWO NKMGNKJQQF ETKVGTKQP VJG OQFGN ECP CNUQ DG VTCKPGF D[ QR VKOK\KPI C FKUETKOKPCVKXG HWPEVKQP (QT GZCORNG VJG /KPKOWO %NCUUKſECVKQP 'TTQT
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
E\&5&3UHVV//&
ENCUUKſECVKQP GTTQT QT OCZKOK\KPI VJG OWVWCN KPHQTOCVKQP DGVYGGP VJG VCTIGV ENCUU CPF QVJGTU 6JG FKUETKOKPCVKXG VTCKPKPI CNIQTKVJOU JCXG DGGP CRRNKGF UWEEGUUHWNN[ VQ URGGEJ TGEQIPKVKQP 6JG /%')2& CNIQTKVJO JCU CNUQ DGGP CRRNKGF VQ URGCMGT TGEQIPKVKQP = ? )GPGTCNN[ URGCMKPI VJG OQFGNU VTCKPGF D[ FKUETKOKPC VKXG QDLGEVKXG HWPEVKQPU [KGNF DGVVGT TGEQIPKVKQP CPF XGTKſECVKQP RGTHQTOCPEG DWV VJG NQPI VTCKPKPI VKOG OCMGU KV NGUU CVVTCEVKXG VQ TGCN CRRNKECVKQPU
7.2.4 Speech Segmentation
)KXGP CP *// CPF C UGSWGPEG QH QDUGTXCVKQPU VJG QRVKOCN UVCVG UGIOGPVCVKQP ECP DG FGVGTOKPGF D[ GXCNWCVKPI VJG OCZKOWO LQKPV UVCVGQDUGTXCVKQP RTQDCDKNKV[ EQPXGPVKQPCNN[ ECNNGF OCZKOWO NKMGNKJQQF FGEQFKPI 1PG RQRWNCT CNIQTKVJO VJCV CEEQORNKUJ VJKU QDLGEVKXG GHſEKGPVN[ KU VJG 8KVGTDK CNIQ TKVJO = ? 9JGP HCUV FGEQFKPI CPF HQTEGF CNKIPOGPV KU FGUKTGF C PGY TGFWEGF URCEG UGCTEJ CNIQTKVJO =? ECP DG GORNQ[GF
7.2.5 Statistical Verification 5VCVKUVKECN XGTKſECVKQP CU CRRNKGF VQ URGCMGT XGTKſECVKQP CPF WVVGTCPEG XGTKſECVKQP ECP DG EQPUKFGTGF CU C VYQENCUU ENCUUKſECVKQP RTQDNGO YJGVJGT C URQMGP WVVGTCPEG KU HTQO VJG VTWG URGCMGT VJG VCTIGV UQWTEG QT HTQO CP KORQUVQT VJG CNVGTPCVKXG UQWTEG )KXGP CP QDUGTXCVKQP C FGEKUKQP KU VCMGP DCUGF QP VJG HQNNQYKPI EQPFKVKQPCN TKUMU FGTKXGF HTQO 'S
6JG CEVKQP EQTTGURQPFU VQ VJG FGEKUKQP QH RQUKVKXG XGTKſECVKQP KH
$TKPI CPF KPVQ CPF TGCTTCPIKPI VJG VGTOU YG VCMG CEVKQP KH
YJGTG KU C RTGUETKDGF VJTGUJQNF (WTVJGTOQTG D[ CRRN[KPI VJG $C[GU HQTOWNC YG JCXG
(QT C UGSWGPEG QH QDUGTXCVKQP YJKEJ CTG CUUWOGF VQ DG KPFGRGPFGPV
CPF KFGPVKECNN[ FKUVTKDWVGF KKF YG JCXG VJG NKMGNKJQQFTCVKQ VGUV
E\&5&3UHVV//&
6JG UCOG TGUWNV ECP CNUQ DG FGTKXGF HTQO VJG 0G[OCPP2GCTUQP FGEKUKQP HQTOWNC VKQP VJWU VJG PCOG Neymann-Pearson VGUV = ? +V ECP DG UJQYP VJCV VJG NKMGNKJQQFTCVKQ VGUV OKPKOK\GU VJG XGTKſECVKQP GTTQT HQT QPG ENCUU YJKNG OCKPVCKPKPI VJG XGTKſECVKQP GTTQT HQT VJG QVJGT ENCUU EQPUVCPV = ? +P RTCEVKEG YG EQORWVG C NQINKMGNKJQQF TCVKQ HQT XGTKſECVKQP
½ ¾
# FGEKUKQP KU OCFG CEEQTFKPI VQ VJG TWNG
#EEGRVCPEG 4GLGEVKQP
YJGTG KU C VJTGUJQNF XCNWG YJKEJ ECP DG FGVGTOKPGF VJGQTGVKECNN[ QT GZRGTKOGP VCNN[ 6JGTG CTG VYQ V[RGU QH GTTQT KP C VGUV HCNUG TGLGEVKQP KG TGLGEVKPI VJG J[RQVJGUKU YJGP KV KU CEVWCNN[ VTWG CPF HCNUG CEEGRVCPEG KG CEEGRVKPI KV YJGP KV KU CEVWCNN[ HCNUG 6JG GSWCN GTTQT TCVG ''4 KU FGſPGF CU VJG GTTQT TCVG YJGP VJG QRGTCVKPI RQKPV KU UQ EJQUGP CU VQ CEJKGXG GSWCN GTTQT RTQDCDKNKVKGU HQT VJG VYQ V[RGU QH GTTQT ''4 JCU DGGP YKFGN[ WUGF CU C XGTKſECVKQP RGTHQTOCPEG KPFKECVQT +P WVVGTCPEG XGTKſECVKQP YG CUUWOG VJCV VJG GZRGEVGF YQTF QT UWDYQTF UGSWGPEG KU MPQYP CPF VJG VCUM KU VQ XGTKH[ YJGVJGT VJG KPRWV URQMGP WVVGTCPEG OCVEJGU KV 5KOKNCTN[ KP 58 VJG VGZV QH VJG RCUURJTCUG KU MPQYP 6JG VCUM KU VQ XGTKH[ YJGVJGT VJG KPRWV URQMGP WVVGTCPEG OCVEJGU VJG IKXGP UGSWGPEG WUKPI VJG OQFGN VTCKPGF D[ VJG URGCMGTŏU XQKEG
7.3 Speaker Verification System #OQPI FKHHGTGPV URGCMGT XGTKſECVKQP U[UVGOU KPVTQFWEGF KP 5GEVKQP YG HQEWU JGTG QP VJG ſZGFRJTCUG U[UVGO = ? CPF GXCNWCVG VJG U[UVGO KP CP QRGPUGV VGUV 6JKU KU FWG VQ VJTGG TGCUQPU (KTUV C UJQTV WUGTUGNGEVGF RJTCUG KU GCU[ VQ TGOGODGT 5GEQPF C ſZGFRJTCUG U[UVGO WUWCNN[ JCU C DGVVGT RGTHQTOCPEG VJCP C VGZVRTQORVGF U[UVGO =? .CUV CP QRGPUGV GXCNWCVKQP KU OQTG CRRTQRTKCVG HQT TGCN CRRNKECVKQPU (QT GZCORNG C NCTIGUECNG VGNGDCPMKPI U[UVGO WUWCNN[ KPXQNXGU C NCTIG WUGT RQRWNCVKQP 6JG RQRWNCVKQP CNUQ EJCPIGU QP C FCKN[ DCUKU +V KU KORQUUKDNG CPF WPTGCNKUVKE VQ EQPUKFGT 58 CU C ENQUGUGV RTQDNGO #U UJQYP KP (KI C ſZGFRJTCUG U[UVGO JCU VYQ RJCUGU GPTQNNOGPV CPF VGUV (QT HGCVWTG GZVTCEVKQP VJG URGGEJ UKIPCN KU UCORNGF CV M*\ CPF RTGGORJCUK\GF WUKPI C ſTUVQTFGT ſNVGT YKVJ C EQGHſEKGPV QH 6JG UCORNGU CTG DNQEMGF KPVQ QXGTNCRRKPI HTCOGU QH OU KP FWTCVKQP CPF WRFCVGF CV OU KPVGTXCNU 'CEJ HTCOG KU YKPFQYGF YKVJ C *COOKPI YKPFQY HQNNQYGF D[ C VJ QTFGT .2% CPCN[UKU 6JG .2% EQGHſ EKGPVU CTG VJGP EQPXGTVGF VQ EGRUVTCN EQGHſEKGPVU YJGTG QPN[ VJG ſTUV EQGHſEKGPVU
E\&5&3UHVV//&
&CVCDCUG 5RGCMGTFGRGPFGPV OQFGN
+FGPVKV[ ENCKO 2JQPGOG 6TCPUETKRVKQP (GCVWTG 8GEVQTU
(QTEGF #NKIPOGPV
L(O, / t )
6CTIGV5EQTG %QORWVCVKQP
%GRUVTCN/GCP 5WDVTCEVKQP $CEMITQWPF 5EQTG %QORWVCVKQP
5RGCMGTKPFGRGPFGPV RJQPGOGOQFGNU
6JTGUJQNF
&GEKUKQP
L(O, /b)
$CEMITQWPFOQFGNU
FIGURE 7.5 A fixed-phrase speaker verification system. CTG TGVCKPGF HQT EQORWVKPI VJG HGCVWTG XGEVQT 6JG HGCVWTG XGEVQT EQPUKUVGF QH HGCVWTGU KPENWFKPI VJG EGRUVTCN EQGHſEKGPVU CPF FGNVC EGRUVTCN EQGHſEKGPVU =? &WTKPI GPTQNNOGPV .2% EGRUVTCN HGCVWTG XGEVQTU EQTTGURQPFKPI VQ VJG PQPUKNGPEG RQTVKQP QH VJG GPTQNNOGPV RCUURJTCUGU CTG WUGF VQ VTCKP C 5& NGHVVQTKIJV *// VQ TGRTGUGPV VJG XQKEG RCVVGTP +V KU ECNNGF C whole-word or whole phrase model =? +P CFFKVKQP VQ OQFGN VTCKPKPI VJG VGZV QH VJG RCUURJTCUG EQNNGEVGF HTQO VJG GPTQNNOGPV UGUUKQP KU VTCPUETKDGF KPVQ C UGSWGPEG QH RJQPGOGU YJGTG KU VJG VJ RJQPGOG CPF KU VJG VQVCN PWODGT QH RJQPGOGU KP VJG RCUURJTCUG 6JG OQFGNU CPF VJG VTCPUETKRVKQP CTG UCXGF KP VJG FCVCDCUG # FGVCKNGF DNQEM FKCITCO QH C VGUV UGUUKQP KU UJQYP KP (KI #HVGT C URGCMGT ENCKOU JKU QT JGT KFGPVKV[ VJG U[UVGO GZRGEVU VJG WUGT VQ URGCM VJG UCOG RJTCUG CU KP VJG GPTQNNOGPV UGUUKQP 6JG XQKEG YCXGHQTO KU ſTUV EQPXGTVGF VQ VJG RTGUETKDGF HGCVWTG TGRTGUGPVCVKQP +P VJG HQTEGF CNKIPOGPV DNQEM C UGSWGPEG QH URGCMGTKPFGRGPFGPV RJQPGOG OQFGNU KU EQPUVTWEVGF CEEQTFKPI VQ VJG RJQPGOKE VTCPUETKRVKQP QH VJG RCUU RJTCUG 6JG OQFGN UGSWGPEG KU VJGP WUGF VQ UGIOGPV CPF CNKIP VJG HGCVWTG XGEVQT UGSWGPEG VJTQWIJ WUG QH VJG 8KVGTDK CNIQTKVJO +P VJG EGRUVTCN OGCP UWDVTCEVKQP DNQEM UKNGPEG HTCOGU CTG TGOQXGF CPF C OGCP XGEVQT KU EQORWVGF DCUGF QP VJG TGOCKPKPI URGGEJ HTCOGU 6JG OGCP KU VJGP UWDVTCEVGF HTQO CNN URGGEJ HTCOGU =? 6JKU KU CP KORQTVCPV UVGR HQT EJCPPGN EQORGPUCVKQP +V OCMGU VJG U[UVGO OQTG TQDWUV VQ EJCPIGU KP VJG QRGTCVKPI GPXKTQPOGPV CU YGNN CU KP VJG VTCPUOKUUKQP EJCPPGN 9G PQVG VJCV VJG HQTEGF CNKIPOGPV DNQEM KU CNUQ WUGF HQT CEEWTCVG GPFRQKPV FGVGEVKQP (QT C U[UVGO YKVJ NKOKVGF EQORWVKPI RQYGT C UGRCTCVG GPFRQKPV FGVGEVKQP CNIQTKVJO ECP DG KORNGOGPVGF HQT HCUV TGURQPUG = ? +P VJG DNQEM QH VCTIGV UEQTG EQORWVCVKQP QH (KI URGGEJ HGCVWTG XGEVQTU CTG FG EQFGF KPVQ UVCVGU D[ VJG 8KVGTDK CNIQTKVJO WUKPI VJG VTCKPGF YJQNGRJTCUG OQFGN # NQINKMGNKJQQF UEQTG HQT VJG VCTIGV OQFGN KG VJG VCTIGV UEQTG KU ECNEWNCVGF CU
YJGTG
KU VJG HGCVWTG XGEVQT UGSWGPEG KU VJG VQVCN PWODGT QH XGEVQTU KP VJG
E\&5&3UHVV//&
UGSWGPEG KU VJG VCTIGV OQFGN CPF KU VJG NKMGNKJQQF UEQTG TGUWNVGF HTQO 8KVGTDK FGEQFKPI +P VJG DNQEM QH DCEMITQWPF PQPVCTIGV UEQTG EQORWVCVKQP C UGV QH URGCMGT KPFGRGP FGPV *//U KP VJG QTFGT QH VJG VTCPUETKDGF RJQPGOG UGSWGPEG KU CRRNKGF VQ CNKIP VJG KPRWV WVVGTCPEG YKVJ VJG GZRGEVGF VTCPUETKRVKQP WUKPI VJG 8KVGTDK YJGTG KU FGEQFKPI CNIQTKVJO 6JG UGIOGPVGF WVVGTCPEG KU VJG UGV QH HGCVWTG XGEVQTU EQTTGURQPFKPI VQ VJG ŏVJ RJQPGOG KP VJG RJQPGOG UGSWGPEG 6JG DCEMITQWPF QT PQPVCTIGV NKMGNKJQQF UEQTG KU VJGP EQORWVGF D[
YJGTG KU VJG UGV QH 5+ RJQPGOG OQFGNU KP VJG QTFGT QH VJG VTCPUETKDGF RJQPGOG UGSWGPEG KU VJG EQTTGURQPFKPI RJQPGOG NKMGNKJQQF UEQTG CPF KU VJG VQVCN PWODGT QH RJQPGOGU 6JG VCTIGV CPF DCEMITQWPF UEQTGU =? CTG VJGP WUGF KP VJG NKMGNKJQQFTCVKQ VGUV
YJGTG CPF CTG FGſPGF KP 'SU CPF TGURGEVKXGN[
6JG U[UVGO JCU DGGP VGUVGF QP C FCVCDCUG EQPUKUVKPI QH ſZGFRJTCUG WVVGTCPEGU 6JG FCVCDCUG YCU TGEQTFGF QXGT C NQPIFKUVCPEG VGNGRJQPG PGVYQTM +V EQPUKUVU QH URGCMGTU OCNG CPF HGOCNG 6JG ſZGF RJTCUG EQOOQP VQ CNN URGCMGTU KU ő+ RNGFIG CNNGIKCPEG VQ VJG ƀCIŒ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ſ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ſIWTCVKQP (QT OQUV ECUGU C TGCN VKOG TGURQPUG KU GZRGEVGF
E\&5&3UHVV//&
TABLE 7.1
Experimental Results in Average Equal-Error Rates (KZGF 2CUU2JTCUG 5RGCMGT 8GTKſECVKQP
9KVJQWV #FCRVCVKQP 9KVJ #FCRVCVKQP
6GUVGF QP URGCMGTU WUKPI QPG EQOOQP RCUURJTCUG
9G PQVG VJCV VJG UCOG RCUURJTCUG KU WUGF HQT CNN URGCMGTU KP QWT GXCNWCVKQP CPF VJG CDQXG TGUWNVU CTG VJG NQYGT DQWPF QH VJG RGTHQTOCPEG 6JG CEVWCN U[UVGO RGT HQTOCPEG YQWNF DG DGVVGT YJGP WUGTU EJQQUG VJGKT QYP CPF OQUV NKMGN[ FKHHGTGPV RCUURJTCUG #NUQ VQ GPUWTG VJG QRGP VGUV PCVWTG PQPG QH VJG KORQUVQTŏU FCVC YCU WUGF HQT FKUETKOKPCVKXGN[ VTCKPKPI VJG 5& VCTIGV OQFGN
7.4 Verbal Information Verification +P VJKU UGEVKQP YG KPVTQFWEG C RCVVGTP TGEQIPKVKQP VGEJPKSWG HQT XGTDCN KPHQTOCVKQP XGTKſECVKQP 8+8 CPF RTGUGPV UQOG GZRGTKOGPVCN TGUWNVU =? )GPGTCNN[ URGCMKPI VJGTG CTG VYQ YC[U VQ XGTKH[ C UKPING URQMGP WVVGTCPEG HQT 8+8 D[ CWVQOCVKE URGGEJ TGEQIPKVKQP #54 QT D[ WVVGTCPEG XGTKſECVKQP 78 9KVJ #54 VJG URQMGP KPRWV KU VTCPUETKDGF KPVQ C UGSWGPEG QH YQTFU 6JG VTCPUETKDGF YQTFU CTG VJGP EQORCTGF VQ VJG KPHQTOCVKQP RTGUVQTGF KP VJG ENCKOGF URGCMGTŏU RGTUQPCN RTQſNG 9KVJ 78 VJG URQMGP KPRWV KU XGTKſGF CICKPUV CP GZRGEVGF UGSWGPEG QH YQTFU QT UWDYQTFU YJKEJ KU VCMGP HTQO C RGTUQPCN FCVC RTQſNG QH VJG ENCKOGF KPFKXKFWCN $CUGF QP QWT GZRGTKGPEG =? CPF VJG CPCN[UKU VJG WVVGTCPEG XGTKſECVKQP CRRTQCEJ ECP IKXG WU OWEJ DGVVGT RGTHQTOCPEG VJCP VJG #54 CRRTQCEJ 6JGTGHQTG YG HQEWU QP VJG WVVGTCPEG XGTKſECVKQP CRRTQCEJ KP VJKU UVWF[ 9JGP C SWGUVKQP KU CPUYGTGF KP VJG HQTO QH C PCVWTCNN[ URQMGP WVVGTCPEG VJG MG[ KPHQTOCVKQP KP VJG RTQſNG OC[ DG GODGFFGF KP C UGPVGPEG GI ő/[ OQVJGTŏU OCKFGP PCOG KU Œ +P VJG UGPVGPEG VJG PCOG KU VJG MG[ KPHQTOCVKQP YJKEJ ECP DG GZVTCEVGF YKVJ C MG[YQTF URQVVKPI VGEJPKSWG =? *GTG YG CUUWOG VJCV VJG MG[ KPHQTOCVKQP JCU DGGP GZVTCEVGF QT VJG CPUYGTGF WVVGTCPEG EQPVCKPU QPN[ VJG MG[ KPHQTOCVKQP 6Q XGTKH[ QPG UKPING WVVGTCPEG YG GORNQ[ VJG VGEJPKSWG QH WVVGTCPEG XGTKſECVKQP YJKEJ YCU FGXGNQRGF HQT MG[YQTF URQTVKPI CPF PQPMG[YQTF TGLGEVKQP GI = ? # DNQEM FKCITCO QH VJG WVVGTCPEG XGTKſECVKQP HQT 8+8 KU UJQYP KP (KI 6JG VJTGG MG[ OQFWNGU CTG WVVGTCPEG UGIOGPVCVKQP D[ HQTEGF FGEQFKPI UWDYQTF VGUVKPI CPF WVVGTCPEG NGXGN EQPſFGPEG OGCUWTG ECNEWNCVKQP 6JG[ YKNN DG FGUETKDGF KP FGVCKN KP VJG HQNNQYKPI UWDUGEVKQPU
E\&5&3UHVV//&
+FGPVKV[ ENCKO
2JQPGUWDYQTF VTCPUETKRVKQPHQT /WTTC[*KNN O 55
2CUUWVVGTCPEG /WTTC[*KNN
(QTEGF &GEQFKPI
6CTIGVNKMGNKJQQFU 2 1^ O OO O 2 1^ %QPHKFGPEG 2JQPG DQWPFCTKGU
5+*//ŏUHQT VJGVTCPUETKRVKQP O O O
#PVK
5EQTGU
/GCUWTG
NKMGNKJQQFU
2 1^ O 2 1^ O OO %QORWVCVKQP O O O
#PVK*//ŏUHQTVJG VTCPUETKRVKQP
FIGURE 7.6 Utterance verification in verbal information verification (VIV).
7.4.1 Utterance Segmentation 9JGP C WUGT QRGPU CP CEEQWPV MG[ KPHQTOCVKQP VJCV EQPUVKVWVGU JKU QT JGT RTQſNG KU TGIKUVGTGF KP C FCVCDCUG 'CEJ RKGEG QH VJG MG[ KPHQTOCVKQP KU TGRTGUGPVGF D[ C UGSWGPEG QH YQTFU Ë YJKEJ KP VWTP KU GSWKXCNGPVN[ EJCTCEVGTK\GF D[ C EQPECVGPCVKQP QH C UGSWGPEG QH RJQPGOGU QT UWDYQTFU YJGTG KU VJG VJ UWDYQTF CPF KU VJG VQVCN PWODGT QH UWDYQTFU KP VJG MG[ YQTF UGSWGPEG 5KPEG VJG 8+8 U[UVGO QPN[ RTQORVU QPG UKPING SWGUVKQP CV C VKOG VJG U[UVGO MPQYU VJG GZRGEVGF EQTTGEV MG[ KPHQTOCVKQP VQ VJG RTQORVGF SWGUVKQP CPF VJG EQTTGURQPF KPI UWDYQTF UGSWGPEG Ë 9G VJGP CRRN[ VJG UWDYQTF OQFGNU KP VJG UCOG QTFGT QH VJG UWDYQTF UGSWGPEG Ë VQ FGEQFG VJG CPUYGT WVVGTCPEG WUKPI VJG 8KVGTDK CNIQTKVJO KPVTQFWEGF RTGXKQWUN[ 6JKU ECP DG TGRTGUGPVGF CU ÇË ½ ½¾ ½ ½ ¾ YJGTG
Ç Ç Ç Ç ½ ½¾ ½
KU C UGV QH UGIOGPVGF HGCVWTG XGEVQTU CUUQEKCVGF YKVJ UWDYQTFU CTG VJG GPFHTCOG PWODGTU QH GCEJ UWDYQTF UGIOGPVU TGURGEVKXGN[ CPF Ç ½ KU VJG UGIOGPVGF UGSWGPEG QH QDUGTXCVKQPU EQTTGURQPFKPI VQ UWDYQTF HTQO HTCOG PWODGT VQ HTCOG PWODGT YJGTG CPF 7.4.2 Subword Hypothesis Testing )KXGP C FGEQFGF UWDYQTF KP CP QDUGTXGF URGGEJ UGIOGPV Ç YG PGGF C FGEK UKQP TWNG D[ YJKEJ YG CUUKIP VJG UWDYQTF VQ GKVJGT QPG QH VJG VYQ ENCUUGU J[RQVJGUGU
E\&5&3UHVV//&
QT YJGTG OGCPU VJCV VJG QDUGTXGF URGGEJ EQPUKUVU QH VJG CEVWCN UQWPF QH UWDYQTF CPF KU VJG CNVGTPCVKXG J[RQVJGUKU 6JG OQUV RQYGTHWN VGUV KU VJG NKMGNKJQQFTCVKQ VGUV CU YG JCXG KPVTQFWEGF
YJGTG CPF CTG VJG VCTIGV *// CPF EQTTGURQPFKPI CPVK*//U HQT UWDYQTF WPKV TGURGEVKXGN[ 6JG VCTIGV OQFGN KU VTCKPGF WUKPI VJG FCVC QH UWDYQTF VJG EQTTGURQPFKPI CPVKOQFGN KU VTCKPGF WUKPI VJG FCVC QH C UGV QH UWDYQTFU Ë YJKEJ KU JKIJN[ EQPHWUCDNG YKVJ UWDYQTF =? KG Ë 6JG NQI NKMGNKJQQF TCVKQ ..4 HQT UWDYQTF KU
Ç Ç Ç
(QT PQTOCNK\CVKQP CP CXGTCIG HTCOG ..4 KU FGſPGF CU
Ç
Ç
YJGTG KU VJG NGPIVJ QH VJG URGGEJ UGIOGPV (QT GCEJ UWDYQTF C FGEKUKQP ECP DG OCFG D[ #EEGRVCPEG
4GLGEVKQP
YJGTG GKVJGT C UWDYQTFFGRGPFGPV VJTGUJQNF XCNWG QT C EQOOQP VJTGUJQNF ECP DG FGVGTOKPGF PWOGTKECNN[ QT GZRGTKOGPVCNN[
7.4.3 Confidence Measure Calculation (QT CP WVVGTCPEG NGXGN FGEKUKQP YG JCXG VQ FGſPG C HWPEVKQP VQ EQODKPG VJG TGUWNVU QH UWDYQTF VGUVU # EQPſFGPEG OGCUWTG HQT C MG[ WVVGTCPEG Ç ECP DG TGRTGUGPVGF CU Ç
YJGTG KU VJG HWPEVKQP VQ EQODKPG VJG ..4U QH CNN UWDYQTFU KP VJG MG[ WVVGTCPEG 5GXGTCN EQPſFGPEG OGCUWTGU JCXG DGGP RTQRQUGF HQT WVVGTCPEG XGTKſECVKQP = ? 9G FGPQVG VYQ QH VJGO CU CPF KP VJG HQNNQYKPI
YJGTG KU VJG VQVCN PWODGT QH PQPUKNGPEG UWDYQTFU KP VJG WVVGTCPEG CPF KU VJG VQVCN PWODGT QH HTCOGU QH VJG PQPUKNGPV RQTVKQP QH VJG WVVGTCPEG KG (WTVJGTOQTG
E\&5&3UHVV//&
*GTG KU CP CXGTCIG UEQTG QXGT CNN HTCOGU CPF CNN UWDYQTFU 'CEJ QH VJG UWD YQTF UEQTG KU YGKIJVGF D[ KVU FWTCVKQP KU CP CXGTCIG ..4 QH CNN UWDYQTFU CPF KPFGRGPFGPV QH KPFKXKFWCN FWTCVKQPU 9G PQVG VJCV UKNGPEG OQFGNU CTG WUGF FWT KPI HQTEGF CNKIPOGPV HQT WVVGTCPEG UGIOGPVCVKQP DWV QPN[ PQPUKNGPEG UWDYQTFU CTG KPXQNXGF KP EQORWVKPI VJG EQPſFGPEG OGCUWTGU (QT 8+8 YG FGſPGF C FKHHGTGPV EQPſFGPEG OGCUWTG HQT VYQ TGCUQPU (KTUV CU TG RQTVGF KP =? CPF HTQO QWT GZRGTKOGPVU VJG CDQXG EQPſFGPEG OGCUWTGU JCXG C NCTIG F[PCOKE TCPIG # RTGHGTCDNG UVCVKUVKE UJQWNF JCXG C UVCDNG NKOKVGF PWOGTKECN TCPIG UWEJ VJCV C EQOOQP VJTGUJQNF ECP DG FGVGTOKPGF HQT CNN UWDYQTFU VQ UKORNKH[ VJG QRGTCVKQP 5GEQPF FGEKUKQP VJTGUJQNFU UJQWNF DG FGVGTOKPGF VQ OGGV URGEKſECVKQPU KP FKHHGTGPV CRRNKECVKQPU +V KU FGUKTCDNG VQ DG CDNG VQ TGNCVG VJG FGUKIP URGEKſECVKQPU YKVJ VJG EQORWVGF EQPſFGPEG OGCUWTG # WUGHWN FGUKIP URGEKſECVKQP KU VJG RGTEGPVCIG QH CEEGRVCDNG UWDYQTFU KP C MG[ WVVGT CPEG 9G VJGP PGGF VQ OCMG C FGEKUKQP CV DQVJ VJG UWDYQTF CPF VJG WVVGTCPEG NGXGN #V VJG UWDYQTF NGXGN C NKMGNKJQQFTCVKQ VGUV ECP DG EQPFWEVGF VQ TGCEJ C FGEKUKQP VQ CEEGRV QT TGLGEV GCEJ UWDYQTF #V VJG WVVGTCPEG NGXGN C UKORNG WVVGTCPEG UEQTG ECP DG EQORWVGF VQ TGRTGUGPV VJG RGTEGPVCIG QH CEEGRVCDNG UWDYQTFU 6Q OCMG C FGEKUKQP CV VJG UWDYQTF NGXGN YG PGGF VQ FGVGTOKPG VJG VJTGUJQNF HQT GCEJ QH VJG UWDYQTF VGUVU +H YG JCXG VJG VTCKPKPI FCVC HQT GCEJ UWDYQTF OQFGN CPF VJG EQTTGURQPFKPI CPVKUWDYQTF OQFGN VJKU KU PQV C RTQDNGO *QYGXGT KP OCP[ ECUGU VJG FCVC OC[ PQV DG CXCKNCDNG 6JGTGHQTG YG PGGF VQ FGſPG C VGUV VJCV ECP EQPXGPKGPVN[ FGVGTOKPG VJG VJTGUJQNFU YKVJQWV WUKPI VJG VTCKPKPI FCVC (QT UWDYQTF YJKEJ KU EJCTCEVGTK\GF D[ C OQFGN YG FGſPG
YJGTG OGCPU VJG VCTIGV UEQTG KU NCTIGT VJCP VJG CPVK UEQTG CPF XKEG XGTUC (WTVJGTOQTG YG FGſPG C normalized confidence measure HQT CP WVVGTCPEG YKVJ UWDYQTFU CU
KHQVJGTYKUG
KU KP C ſZGF TCPIG QH &WG VQ VJG PQTOCNK\CVKQP KP 'S KU C UWDYQTFKPFGRGPFGPV VJTGUJQNF YJKEJ ECP DG FGVGTOKPGF UGRCTCVGN[ # UWDYQTF KU CEEGRVGF CPF EQWPVGF CU RCTV QH VJG WVVGTCPEG EQPſFGPEG OGCUWTG QPN[ KH KVU UEQTG KU ITGCVGT VJCP QT GSWCN VQ VJG VJTGUJQNF XCNWG 6JWU ECP DG KPVGTRTGVGF CU VJG RGTEGPVCIG QH CEEGRVCDNG UWDYQTFU KP CP WVVGTCPEG GI KORNKGU VJCV YJGTG
QH VJG UWDYQTFU KP VJG WVVGTCPEG CTG CEEGRVCDNG 6JGTGHQTG CP WVVGTCPEG VJTGUJQNF ECP DG FGVGTOKPGF QT CFLWUVGF DCUGF QP VJG URGEKſECVKQPU QH U[UVGO RGTHQTOCPEG CPF TQDWUVPGUU
E\&5&3UHVV//&
7.4.4 Sequential Utterance Verification (QT 8+8 VJG U[UVGO YQWNF IQ VJTQWIJ OQTG VJCP QPG SWGUVKQPCPUYGT VWTPU DGHQTG C ſPCN FGEKUKQP KU OCFG 6JWU VJG CDQXG UKPING WVVGTCPEG VGUV UVTCVGI[ PGGFU VQ DG GZVGPFGF VQ C UGSWGPEG QH UWDVGUVU UKOKNCT VQ VJG step-down procedure KP UVCVKUVKEU =? +P UWEJ C UGSWGPVKCN VGUV GCEJ QH VJG UWDVGUVU KU KPFGRGPFGPVN[ EQPUVTWEVGF CU C UKPINGWVVGTCPEG XGTKſECVKQP VGUV 9G ECP OCMG C UQHV QT FGNC[GF FGEKUKQP HQT VGUV CU HQNNQYU
#EEGRVCPEG &GNC[ IQ VQ VJG PGZV VGUV 4GLGEVKQP
YJGTG KU C EQPſFGPEG UEQTG CPF CTG VJG JKIJ CPF NQY VJTGUJQNFU HQT VGUV TGURGEVKXGN[ .GV DG VJG VCTIGV J[RQVJGUKU KP YJKEJ CNN VJG CPUYGTGF WVVGTCPEGU OCVEJ VJG MG[ KPHQTOCVKQP KP VJG RTQſNG 9G JCXG
YJGTG KU VJG VQVCN PWODGT QH UWDVGUVU CPF KU C EQORQPGPV VCTIGV J[RQVJGUKU KP VJG VJ UWDVGUV EQTTGURQPFKPI VQ VJG VJ WVVGTCPEG 6JG CNVGTPCVKXG J[RQVJGUKU KU
YJGTG KU C EQORQPGPV CNVGTPCVKXG J[RQVJGUKU EQTTGURQPFKPI VQ VJG VJ UWDVGUV 1P VJG VJ UWDVGUV C UKORNKſGF XGTUKQP QH VJG UQHV FGEKUKQP ECP DG OCFG CU
&GNC[ IQ VQ VJG PGZV VGUV 4GLGEVKQP
YJGTG KU C EQPſFGPEG UEQTG CPF KU C UKPING VJTGUJQNFU HQT VGUV #U YG JCXG KPVTQFWEGF YJGP RGTHQTOKPI C J[RQVJGUKU VGUV QPG OC[ EQOOKV QPG QH VYQ V[RGU QH GTTQTU TGLGEVKPI VJG J[RQVJGUKU YJGP KV KU VTWG false rejection (4 QT CEEGRVKPI KV YJGP KV KU HCNUG false acceptance (# 9G FGPQVG VJG (4 CPF (# GTTQT TCVGU CU CPF TGURGEVKXGN[ #P equal-error rate ''4 KU FGſPGF YJGP VJG VYQ GTTQT TCVGU CTG OCFG GSWCN D[ EJQQUKPI C RCTVKEWNCT QRGTCVKPI RQKPV HQT VJG U[UVGO KG (QT C UGSWGPVKCN VGUV YG GZVGPF VJG FGſPKVKQPU QH GTTQT TCVGU CU HQNNQYU Definition 1: False rejection error on utterances KU VJG GTTQT YJGP VJG U[UVGO TGLGEVU C EQTTGEV TGURQPUG KP CP[ QPG QH J[RQVJGUKU UWDVGUVU Definition 2: False acceptance error on utterances KU VJG GTTQT YJGP VJG U[UVGO CEEGRVU CP KPEQTTGEV UGV QH TGURQPUGU CHVGT CNN QH J[RQVJGUKU UWDVGUVU Definition 3: Equal-error rate on utterances KU VJG TCVG CV YJKEJ VJG HCNUG TGLGEVKQP GTTQT TCVG CPF VJG HCNUG CEEGRVCPEG GTTQT TCVG QP WVVGTCPEGU CTG GSWCN
E\&5&3UHVV//&
9G FGPQVG VJG CDQXG (4 CPF (# GTTQT TCVGU QP WVVGTCPEGU CU CPF TGURGEVKXGN[ .GV DG VJG TGIKQP QH EQPſFGPEG UEQTGU QH VJG VJ UWDVGUV YJGTG KU VJG TGIKQP QH EQPſFGPEG UEQTGU YJKEJ UCVKUH[ HTQO YJKEJ YG CEEGRV CPF KU VJG TGIKQP QH UEQTGU YJKEJ UCVKUH[ HTQO YJKEJ YG CEEGRV 6JG (4 CPF (# GTTQTU HQT UWDVGUV ECP DG TGRTGUGPVGF CU VJG HQNNQYKPI EQPFKVKQPCN RTQDCDKNKVKGU
CPF
TGURGEVKXGN[ (WTVJGTOQTG VJG (4 GTTQT QP WVVGTCPEGU ECP DG GXCNWCVGF CU
CPF VJG (# GTTQT QP WVVGTCPEGU KU
'SU CPF KPFKECVG CP KORQTVCPV RTQRGTV[ QH VJG UGSWGPVKCN VGUV FGſPGF CDQXG VJG OQTG VJG UWDVGUVU VJG NGUU VJG (# GTTQT CPF VJG NCTIGT VJG (4 GTTQT 6JGTG HQTG KV KU KORQTVCPV VJCV VJG VJTGUJQNF CV GXGT[ UWDVGUV KU ECTGHWNN[ EJQUGP UQ CU VQ CEJKGXG CP (4 GTTQT VJCV KU ENQUG VQ \GTQ QT C UOCNN PWODGT EQTTGURQPFKPI VQ VJG FGUKIP URGEKſECVKQP DWV CFF OQTG UWDVGUVU KP VJG UCOG YC[ CU PGGFGF WPVKN VJG TG SWKTGF U[UVGO (# GTTQT TCVG KU OGV QT VJG OCZKOWO PWODGT QH CNNQYGF UWDVGUVU KU TGCEJGF +V KU TGCUQPCDNG VQ CTTCPIG VJG UWDVGUVU KP VJG QTFGT QH FGUEGPFKPI KORQTVCPEG CPFQT FGETGCUKPI UWDVGUV GTTQT TCVGU +P QVJGT YQTFU VJG U[UVGO ſTUV RTQORVU WUGTU YKVJ VJG OQUV KORQTVCPV SWGUVKQP QT YKVJ VJG UWDVGUV VJCV YG MPQY JCU VJG NQYGUV (4 GTTQT 6JGTGHQTG KH C URGCMGT KU HCNUGN[ TGLGEVGF VJG UGUUKQP ECP DG TGUVCTVGF TKIJV CYC[ YKVJ NKVVNG KPEQPXGPKGPEG VQ VJG WUGT 'S CNUQ KPFKECVGU VJG TGCUQP VJCV VJG #54 CRRTQCEJ YQWNF PQV RGTHQTO XGT[ YGNN KP C UGSWGPVKCN VGUV #NVJQWIJ CP #54 ECP CEJKGXG NQY (4 GTTQT QP GCEJ QH VJG KPFKXKFWCN UWDVGUVU VJG QXGTCNN (4 GTTQT QP WVVGTCPEGU ECP UVKNN DG XGT[ JKIJ FWG VQ VJG HCEV VJCV VJG XGTKſECVKQP RTQEGUU KP CP #54DCUGF CRRTQCEJ WUGU YQTF EQORCTKUQP CPF FQGU PQV RGTOKV C UQHV FGEKUKQP QT FGNC[GF FGEKUKQP +P VJG RTQRQUGF WVVGTCPEG XGTKſECVKQP CRRTQCEJ VJG (4 QP GCEJ KPFKXKFWCN UWDVGUV KU
E\&5&3UHVV//&
OCFG ENQUG VQ \GTQ D[ CFLWUVKPI VJG VJTGUJQNF XCNWG YJKNG EQPVTQNNKPI VJG QXGTCNN (# GTTQT D[ CFFKPI OQTG UWDVGUVU WPVKN TGCEJKPI VJG FGUKIP URGEKſECVKQP 9G WUG VJG HQNNQYKPI GZCORNGU VQ UJQY VJG CDQXG EQPEGRV Example 1: # DCPM QRGTCVQT WUWCNN[ CUMU VYQ MKPFU QH RGTUQPCN SWGUVKQPU YJGP XGTKH[KPI C EWUVQOGT 9JGP CWVQOCVKE 8+8 KU CRRNKGF VQ VJG RTQEGFWTG VJG CXGT CIG KPFKXKFWCN GTTQT TCVGU QP VJGUG VYQ UWDVGUVU CTG CPF TGURGEVKXGN[ 6JGP HTQO 'S CPF YG MPQY VJCV VJG U[UVGO (4 CPF (# GTTQTU QP C UGSWGPVKCN VGUV CTG CPF +H VJG DCPM YCPVU VQ HWTVJGT TGFWEG VJG (# GTTQT QPG CFFK VKQPCN UWDVGUV ECP DG CFFGF VQ VJG UGSWGPVKCN VGUV 5WRRQUG VJG CFFKVKQPCN UWDVGUV JCU CPF 6JG QXGTCNN U[UVGO GTTQT TCVGU YKNN DG CPF Example 2: # UGEWTKV[ U[UVGO TGSWKTGU CPF +V KU MPQYP VJCV GCEJ UWDVGUV ECP JCXG CPF D[ CFLWUVKPI VJG VJTGUJQNFU +P VJKU ECUG YG PGGF VQ FGVGTOKPG VJG PWODGT QH UWDVGUVU VQ OGGV VJG FGUKIP URGEKſECVKQPU (TQO 'S YG JCXG
6JGP VJG CEVWCN U[UVGO (# TCVG QP VJTGG UWDVGUVU KU VJG (4 TCVG QP VJTGG VGUVU KU 6JGTGHQTG VJTGG UWDVGUVU ECP OGGV VJG TGSWKTGF RGTHQTOCPEG QP DQVJ (4 CPF (#
7.4.5 VIV Experimental Results +P VJG HQNNQYKPI GZRGTKOGPVU VJG 8+8 U[UVGO XGTKſGU URGCMGTU D[ VJTGG UGSWGPVKCN UWDVGUVU KG 6JG GZRGTKOGPVCN FCVCDCUG KPENWFGU URGCMGTU 'CEJ URGCMGT RTQXKFGF VJTGG WVVGTCPEGU CU VJG CPUYGTU VQ VJG HQNNQYKPI VJTGG SWGUVKQPU ő+P YJKEJ [GCT YGTG [QW DQTP!Œ ő+P YJKEJ EKV[ CPF UVCVG FKF [QW ITQY WR!Œ CPF ő/C[ + JCXG [QWT VGNGRJQPG PWODGT RNGCUG!Œ 6JG FCVCDCUG YG WUGF KU C DKCUGF QPG 6YGPV[ UKZ RGTEGPV QH VJG URGCMGTU JCXG DKTVJ [GCT KP VJG U CPF CTG KP VJG U 6JGTG KU QPN[ QPG FKIKV VJCV FKHHGTGPVKCVGU VJQUG DKTVJ [GCTU +P EKV[ CPF UVCVG PCOGU CTG ő0GY ,GTUG[Œ CPF QH VJG URGCMGTU WUGF GZCEVN[ VJG UCOG CPUYGT ő/WTTC[ *KNN 0GY ,GTUG[Œ 6JKTV[ GKIJV RGTEGPV QH VJG VGNGRJQPG PWODGTU UVCTV HTQO ő Œ YJKEJ OGCPU VJCV CV NGCUV QH VJG FKIKVU KP VJGKT CPUYGT HQT VJG VGNGRJQPG PWODGT CTG KFGPVKECN #NUQ UQOG QH VJG URGCMGTU JCXG HQTGKIP CEEGPV CPF UQOG EKVKGU CPF UVCVGU CTG KP HQTGKIP EQWPVTKGU 'ZKUVKPI #54 U[UVGOU ECPPQV RTQXKFG CP CEEGRVCDNG RGTHQTOCPEG +P VJKU GZRGTKOGPV C URGCMGT KU EQPUKFGTGF C VTWG URGCMGT YJGP VJG URGCMGTŏU WVVGT CPEGU CTG XGTKſGF CICKPUV JKU QT JGT FCVC RTQſNG 6JG UCOG URGCMGT KU WUGF CU CP KORQUVQT YJGP VJG WVVGTCPEGU CTG XGTKſGF CICKPUV QVJGT URGCMGTUŏ RTQſNGU 6JWU HQT GCEJ VTWG URGCMGT YG JCXG VJTGG WVVGTCPEGU HTQO VJG URGCMGT CPF WVVGTCPEGU HTQO QVJGT URGCMGTU CU KORQUVQTU
E\&5&3UHVV//&
6JG HGCVWTG XGEVQT EQPUKUVGF QH HGCVWTGU KPENWFKPI .2% EGRUVTCN EQGHſEKGPVU FGNVC EGRUVTCN EQGHſEKGPVU FGNVCFGNVC EGRUVTCN EQGHſEKGPVU GPGTI[ FGNVC GPGTI[ CPF FGNVCFGNVC GPGTI[ =? +P GXCNWCVKPI VJG UWDYQTF XGTKſECVKQP UEQTGU C UGV QH TKIJV EQPVGZVFGRGPFGPV *//U YGTG WUGF CU VJG VCTIGV RJQPG OQFGNU =? CPF C UGV QH EQPVGZVKPFGRGPFGPV CPVKRJQPG *//U CU CPVKOQFGNU =? (QT C 8+8 U[UVGO YKVJ OWNVKRNG UWDVGUVU QPG ECP WUG GKVJGT QPG UKPING VJTGUJQNF CR RNKGF INQDCNN[ VQ CNN VJG UWDVGUVU KG QT OWNVKRNG VJTGUJQNFU GCEJ CRRNKGF VQ KPFKXKFWCN SWGUVKQPU TGURGEVKXGN[ KG 6JG VJTGUJQNFU ECP DG GKVJGT EQPVGZV FGRGPFGPV QT EQPVGZV KPFGRGPFGPV 6JG[ ECP CNUQ DG GKVJGT URGCMGT FG RGPFGPV QT URGCMGT KPFGRGPFGPV # 8+8 U[UVGO ECP UVCTV HTQO C URGCMGTKPFGRGPFGPV VJTGUJQNF VJGP UYKVEJ VQ URGCMGT CPF EQPVGZVFGRGPFGPV VJTGUJQNFU CHVGT VJG U[UVGO JCU DGGP WUGF HQT UGXGTCN VKOGU D[ C WUGT 6Q GPUWTG PQ HCNUG TGLGEVKQP VJG WRRGT DQWPF QH VJG VJTGUJQNF HQT UWDVGUV QH C URGCMGT ECP DG UGNGEVGF CU
YJGTG KU VJG EQPſFGPEG UEQTG HQT WVVGTCPEG QP VJG VJ VTKCN CPF KU VJG VQVCN PWODGT QH VTKCNU VJCV VJG URGCMGT JCU RGTHQTOGF KP VJG UCOG EQPVGZV QH WVVGTCPEG 6JG VJTGUJQNFU ECP CNUQ DG WRFCVGF DCUGF QP VJG TGEGPV UEQTGU VQ CEEQOOQFCVG VJG EJCPIGU KP URGCMGTŏU XQKEG CPF GPXKTQPOGPV +P VJKU GZRGTKOGPV YG WUGF VJTGG VJTGUJQNFU CUUQEKCVGF YKVJ VJG VJTGG SWGUVKQPU HQT GCEJ URGCMGT (QNNQYKPI VJG FGUKIP UVTCVGI[ RTQRQUGF KP 5GEVKQP VJG VJTGUJQNFU YGTG FGVGTOKPGF D[ GUVKOCVKPI CU KP 'S VQ IWCTCPVGG HCNUG TGLGEVKQP TCVG TABLE 7.2
Summary of the Experimental Results on Verbal Information Verification #RRTQCEJGU 5GSWGPVKCN 7VVGTCPEG 8GTKſECVKQP
(CNUG (CNUG #EEWTCE[ 4GLGEVKQP #EEGRVCPEG
6GUVGF QP URGCMGTU YKVJ SWGUVKQPU YJKNG URGCMGTFGRGPFGPV VJTGUJQNFU YGTG CRRNKGF
# UWOOCT[ QH 8+8 HQT URGCMGT CWVJGPVKECVKQP KU UJQYP KP 6CDNG 9JGP 5& VJTGUJQNFU CTG UGV HQT GCEJ MG[ KPHQTOCVKQP ſGNF YG CEJKGXGF KPFKXKFWCN GSWCN GTTQT TCVG QP CXGTCIG 6JG TQDWUVPGUU QH VJG U[UVGO YCU CNUQ GXCNWCVGF +PVGTGUVGF TGCFGTU CTG TGHGTTGF VQ =? HQT FGVCKNU
E\&5&3UHVV//&
2CUURJTCUGUQHVJGHKTUVHGYCEEGUUGU 1RGP5GUCOG 1RGP5GUCOG 1RGP5GUCOG
5CXGHQTVTCKPKPI
8GTDCN +PHQTOCVKQP 8GTKHKECVKQP 8GTKHKGFRCUURJTCUGU HQTVTCKPKPI
#WVQOCVKE'PTQNNOGPV
*// 6TCKPKPI 5RGCMGTFGRGPFGPV *// &CVCDCUG
5RGCMGT8GTKHKECKVQP +FGPVKV[ENCKO 6GUVRCUURJTCUG 1RGP5GUCOG
5RGCMGT 8GTKHKGT
5EQTGU
FIGURE 7.7 An integrated voice authentication system combining verbal information verification and speaker verification.
7.5 Speaker Authentication by Combining SV and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ſEWNV VQ EQTTGEV QPEG C URGCMGT FGRGPFGPV OQFGN KU EQPUVTWEVGF WPNGUU OCPWCN GZCOKPCVKQP CPF XGTKſECVKQP QH FCVC VCMGU RNCEG DGHQTG OQFGN VTCKPKPI 1DXKQWUN[ 8+8 KU C PCVWTCN CPF RQYGTHWN VGEJ PKSWG HQT VJKU RWTRQUG 1PG QH VJG UQNWVKQPU KU UJQYP KP (KI &WTKPI VJG ſTUV HGY CEEGUUGU QT WUGU QH VJG U[UVGO CWVJGPVKECVKQP KU EQPFWEVGF D[ C 8+8 RTQEG FWTG 6JG WVVGTGF RCUURJTCUG OWUV RCUU 8+8 VGUVU QVJGTYKUG VJG WUGT KU RTQORVGF VQ TGRGCV 8GTKſGF WVVGTCPEGU QH VJG RCUURJTCUG CTG VJGP UCXGF CPF WUGF VQ VTCKP C
E\&5&3UHVV//&
URGCMGTFGRGPFGPV *// HQT 58 #V VJKU RQKPV VJG CWVJGPVKECVKQP U[UVGO ECP VJGP DG UYKVEJGF HTQO 8+8 VQ 58 6JGTG CTG UGXGTCN CFXCPVCIGU D[ WUKPI VJG EQODKPGF U[UVGO (KTUV VJG U[UVGO KU EQPXGPKGPV VQ WUGTU UKPEG KV FQGU PQV PGGF C HQTOCN GPTQNNOGPV UGUUKQP CPF C WUGT ECP UVCTV VQ WUG VJG U[UVGO TKIJV CHVGT JKUJGT CEEQWPV KU UGV WR 5GEQPF VJG CEQWUVKE OKUOCVEJ RTQDNGO KU VQ C EGTVCKP FGITGG OKVKICVGF UKPEG VJG VTCKPKPI FCVC OC[ EQOG HTQO FKHHGTGPV UGUUKQPU RQVGPVKCNN[ XKC FKHHGTGPV JCPFUGVU CPF EJCPPGNU 6JKTF VJG SWCNKV[ QH VJG VTCKPKPI FCVC CTG GPUWTGF UKPEG VJG VTCKPKPI RJTCUGU CTG XGTKſGF D[ 8+8 DGHQTG DGKPI WUGF VQ VTCKP VJG URGCMGTFGRGPFGPV *//U HQT VJG RCUURJTCUG (KPCNN[ QPEG VJG U[UVGO UYKVEJGU VQ 58 KV YQWNF DG FKHſEWNV HQT CP KORQUVQT VQ CEEGUU VJG CEEQWPV GXGP KH VJG KORQUVGT MPQYU VJG VTWG URGCMGTŏU RCUURJTCUG 9G EQPFWEVGF CP GZRGTKOGPV VQ XGTKH[ VJG RGTHQTOCPEG QH VJG EQODKPGF U[UVGO 6JG HGCVWTG CPF FCVCDCUG CTG VJG UCOG CU VJG URGCMGT XGTKſECVKQP U[UVGO KPVTQFWEGF KP VJG RTGXKQWU UGEVKQP 6JG GZRGTKOGPVCN FCVCDCUG EQPUKUVU QH ſZGF RJTCUG WVVGTCPEGU TGEQTFGF QXGT VJG NQPI FKUVCPEG VGNGRJQPG PGVYQTM D[ URGCMGTU OCNG CPF HGOCNG 6JG ſZGF RJTCUG EQOOQP VQ CNN URGCMGTU KU ő+ RNGFIG CNNGIKCPEG VQ VJG ƀCIŒ YKVJ CP CXGTCIG NGPIVJ QH UGEQPFU 9G CUUWOG VJG ſZGF RJTCUG KU QPG QH VJG XGTKſGF WVVGTCPEGU KP 8+8 (KXG WVVGTCPEGU QH VJG RCUURJTCUG TGEQTFGF HTQO ſXG UGRCTCVG 8+8 UGUUKQPU YGTG WUGF VQ VTCKP VJG 5& *// VJWU VJG VTCKPKPI FCVC CTG EQNNGEVGF HTQO FKHHGTGPV CEQWUVKE GPXKTQPOGPVU CPF VGNGRJQPG EJCPPGNU CV FKHHGTGPV VKOG 9G CUUWOG CNN VJG EQNNGEVGF WVVGTCPEGU JCXG DGGP XGTKſGF D[ 8+8 VQ GPUWTG VJG SWCNKV[ QH VJG VTCKPKPI FCVC (QT VGUVKPI YG WUGF WVVGTCPEGU TGEQTFGF HTQO C VTWG URGCMGT KP FKHHGTGPV UGUUKQPU CPF WVVGTCPEGU TGEQTFGF HTQO KORQUVQTU QH VJG UCOG IGPFGT KP FKHHGTGPV UGU UKQPU 6JG OQFGN UVTWEVWTG KU VJG UCOG CU VJG RTGXKQWU 58 U[UVGO (QT OQFGN CFCRVC VKQP VJG UGEQPF HQWTVJ UKZVJ CPF GKIJVJ VGUV WVVGTCPEGU HTQO VJG VGUVGF VTWG URGCMGT YGTG WUGF VQ WRFCVG VJG CUUQEKCVGF *//U HQT XGTKH[KPI UWDUGSWGPV VGUV WVVGTCPEGU KPETGOGPVCNN[ =? +P 5GEVKQP YG JCXG TGRQTVGF VJG GZRGTKOGPVCN TGUWNVU QH 8+8 KP C VGUV QH URGCMGTU 6JG U[UVGO CEJKGXGF GTTQT TCVGU YKVJ VJTGG TQWPFU QH SWGUVKQPCPUYGT VGUV KP C UGSWGPVKCN WVVGTCPEG XGTKſECVKQP RTQEGFWTG 6JGTGHQTG YG CUUWOG VJCV CNN VJG VTCKPKPI WVVGTCPEGU EQNNGEVGF D[ 8+8 CTG EQTTGEV +P QVJGT YQTFU YJKNG KORTQXGOGPV D[ TGFWEKPI CEQWUVKE OKUOCVEJ YKNN DGEQOG QDXKQWU KP VJG TGUWNV YG FKF PQV FGUKIP CP GZRGTKOGPV VQ UJQY VJG RQVGPVKCN KORTQXGOGPV KP XGTKſECVKQP RGTHQTOCPEG HTQO CP KPETGCUGF UCPKV[ EJGEM QP VJG VTCKPKPI FCVC 6JG 58 GZRGTKOGPVCN TGUWNVU YKVJQWV CPF YKVJ CFCRVCVKQP CTG NKUVGF KP 6CDNG CPF 6CDNG HQT VJG URGCMGTU TGURGEVKXGN[ 6JG PWODGTU CTG GZRTGUUGF KP VGTOU QH VJG CXGTCIG RGTEGPVCIG QH KPFKXKFWCN GSWCNGTTQT TCVG ''4 6JG ſTUV FCVC EQNWOP NKUVU VJG ''4U WUKPI KPFKXKFWCN VJTGUJQNFU CPF VJG UGEQPF FCVC EQNWOP NKUVU VJG ''4U WUKPI EQOOQP RQQNGF VJTGUJQNFU HQT CNN VGUVGF URGCMGTU 6JG DCUGNKPG U[UVGO KU VJG EQPXGPVKQPCN 58 U[UVGO KP YJKEJ C UKPING GPTQNNOGPV UGUUKQP KU WUGF +P VJG EQODKPGF U[UVGO 8+8 KU WUGF HQT VJG CWVQOCVKE GPTQNNOGPV HQT 58 #HVGT VJG 8+8 U[UVGO KU WUGF HQT ſXG VKOGU VJCV CNNQYU EQNNGEVKQP QH VTCKPKPI WVVGTCPEGU HTQO ſXG FKHHGTGPV UGUUKQPU KV UYKVEJGU VQ VJG 58 RTQEGFWTG 6JG VGUV WVVGTCPEGU HQT DQVJ VJG DCUGNKPG CPF VJG RTQRQUGF U[UVGO CTG VJG UCOG
E\&5&3UHVV//&
9KVJQWV CFCRVCVKQP VJG DCUGNKPG U[UVGO JCU CP ''4 QH CPF HQT KPFK XKFWCN CPF RQQNGF VJTGUJQNFU TGURGEVKXGN[ YJKNG VJG RTQRQUGF U[UVGO JCU CP ''4 QH CPF TGURGEVKXGN[ 9KVJ CFCRVCVKQP CU FGſPGF KP VJG NCUV UWDUGEVKQP VJG DCUGNKPG U[UVGO CEJKGXGU CP ''4 QH CPF YJKNG VJG RTQRQUGF U[UVGO CEJKGXGU CP ''4 QH CPF TGURGEVKXGN[ 6JG RTQRQUGF U[UVGO YKVJQWV CFCRVCVKQP JCU CP GXGP NQYGT ''4 VJCP VJG DCUGNKPG U[UVGO YKVJ CFCRVCVKQP 6JKU KU DGECWUG VJG 5& OQFGNU KP VJG RTQRQUGF U[UVGO YGTG VTCKPGF WUKPI VJG FCVC HTQO FKH HGTGPV UGUUKQPU YJKNG VJG DCUGNKPG U[UVGO LWUV RGTHQTOGF CP KPETGOGPVCN CFCRVCVKQP YKVJQWV TGEQPUVTWEVKPI VJG OQFGNU CHVGT EQNNGEVKPI OQTG FCVC TABLE 7.3
Experimental Results without Adaptation in Average Equal-Error Rates #NIQTKVJOU +PFKXKFWCN 6JTGUJQNFU 2QQNGF 6JTGUJQNFU 58 $CUGNKPG 8+8 58 RTQRQUGF
TABLE 7.4
Experimental Results with Adaptation in Average Equal-Error Rates #NIQTKVJOU +PFKXKFWCN 6JTGUJQNFU 2QQNGF 6JTGUJQNFU 58 $CUGNKPG 8+8 58 RTQRQUGF
6JG GZRGTKOGPVCN TGUWNVU KPFKECVG UGXGTCN CFXCPVCIGU QH VJG RTQRQUGF U[UVGO (KTUV UKPEG 8+8 ECP RTQXKFG VJG VTCKPKPI FCVC HTQO FKHHGTGPV UGUUKQPU TGRTGUGPVKPI FKHHGT GPV EJCPPGN GPXKTQPOGPVU VJG U[UVGO ECP RGTHQTO UKIPKſECPVN[ DGVVGT VJCP QPG YKVJ UKPINGUGUUKQP VTCKPKPI 5GEQPF CNVJQWIJ KV KU RQUUKDNG VQ CFCRV VJG OQFGNU QTKIKPCNN[ VTCKPGF YKVJ VJG UKPINGUGUUKQP FCVC VQ PGY VGUV GPXKTQPOGPVU VJG EQODKPGF U[UVGO CRRGCTU VQ RGTHQTO DGVVGT UVKNN 6JKU KU FWG VQ VJG HCEV VJCV C PGY OQFGN EQPUVTWEVGF YKVJ OWNVKUGUUKQP VTCKPKPI FCVC KU OQTG CEEWTCVG VJCP VJCV YKVJ KPETGOGPVCN CFCRVC VKQP WUKPI VJG OWNVKUGUUKQP FCVC .CUVN[ KP TGCNYQTNF CRRNKECVKQPU CNN VJG WVVGTCPEGU WUGF KP VTCKPKPI CPF CFCRVCVKQP ECP DG XGTKſGF D[ 8+8 DGHQTG VTCKPKPI QT CFCRVCVKQP #NVJQWIJ VJKU CFXCPVCIG ECPPQV DG QDUGTXGF KP VJKU FCVCDCUG GXCNWCVKQP KV KU ETKVK ECN KP TGCNYQTNF CRRNKECVKQPU UKPEG GXGP C VTWG URGCMGT OC[ OCMG C OKUVCMG YJKNG WVVGTKPI C RCUURJTCUG 6JG OKUVCMG YKNN PGXGT DG EQTTGEVGF QPEG KPXQNXGF KP OQFGN VTCKPKPI QT CFCRVCVKQP 8+8 ECP RTQVGEV VJG U[UVGO HTQO YTQPI VTCKPKPI FCVC +P VJKU UGEVKQP YG QPN[ RTQRQUGF QPG EQPſIWTCVKQP QH C EQODKPGF CWVJGPVKECVKQP U[UVGO (QT FKHHGTGPV CRRNKECVKQPU FKHHGTGPV EQPſIWTCVKQPU QH KPVGITCVKQP ECP DG FG UKIPGF VQ OGGV VJG URGEKſECVKQP (KPCNN[ YG PQVG VJCV KV KU VJG WUGTŏU TGURQPUKDKNKV[ VQ RTQVGEV JKU QT JGT RGTUQPCN KPHQTOCVKQP HTQO KORQUVQTU WPVKN VJG 5& OQFGN KU VTCKPGF
E\&5&3UHVV//&
CPF VJG U[UVGO KU OKITCVGF VQ CP 58 U[UVGO #HVGT OKITCVKQP CP KORQUVQT YQWNF JCXG FKHſEWNVKGU KP CEEGUUKPI VJG CEEQWPV GXGP KH VJG RCUURJTCUG KU MPQYP
7.6 Summary +P VJKU EJCRVGT YG RTGUGPVGF RCVVGTP TGEQIPKVKQP OGVJQFU KP URGCMGT CWVJGPVKECVKQP 6JG VJGQTGVKECN HQWPFCVKQP QH VJG CWVJGPVKECVKQP VGEJPKSWGU KU VJG $C[GUKCP FGEKUKQP VJGQT[ CPF J[RQVJGUKU VGUVKPI &GRGPFKPI QP CRRNKECVKQPU J[RQVJGUKU VGUVKPI ECP DG EQPFWEVGF CV RJTCUG YQTF RJQPGOG QT UWDYQTF NGXGN 1PG GZVGPUKQP VQ VJG $C[GUKCP VJGQT[ VQ CWVJGPVKECVKQP KU VJG UGSWGPVKCN XGTKſECVKQP RTQEGFWTG )KXGP C PWODGT QH VGUV WVVGTCPEGU UWDVGUVU VJG VGUV RTQEGFWTG ECP DG FGUKIPGF VQ CEJKGXG OKPKOCN QXGTCNN GTTQT TCVG 6JG UGSWGPVKCN XGTKſECVKQP RTQEGFWTG ECP CNUQ DG CRRNKGF VQ URGCMGT XGTKſECVKQP VQ TGFWEG VJG GTTQT TCVG #OQPI VJG CWVJGPVKECVKQP VGEJPKSWGU URGCMGT XGTKſECVKQP 58 KU VJG RTQEGUU QH XGT KH[KPI URGCMGTU D[ VJGKT XQKEG EJCTCEVGTKUVKEU %WTTGPVN[ VJG ſZGFRJTCUG 58 U[UVGO KU OQTG CVVTCEVKXG VQ TGCN CRRNKECVKQPU FWG VQ KVU IQQF RGTHQTOCPEG # ſZGFRJTCUG U[UVGO CNNQYU WUGTU VQ UGNGEV VJGKT RGTUQPCN RCUURJTCUG VJGTGHQTG KV KU GCU[ VQ TG OGODGT CPF EQPXGPKGPV VQ WUG 9JGP CP CEEQWPV PWODGT UWEJ CU C EQPPGEVGF FKIKV UVTKPI KU WUGF CU C RCUURJTCUG VJG WVVGTGF CEEQWPV PWODGT ECP DG TGEQIPK\GF D[ CP CWVQOCVKE URGGEJ TGEQIPKVKQP U[UVGO CPF VJGP XGTKſGF D[ C URGCMGT XGTKſECVKQP U[UVGO 6JWU QPG WVVGTCPEG ECP DG WUGF HQT DQVJ KPHQTOCVKQP TGVTKGXCN CPF CWVJGPVK ECVKQP 8GTDCN KPHQTOCVKQP XGTKſECVKQP 58 KU VQ XGTKH[ C URGCMGT D[ VJG XGTDCN EQPVGPV KP VJG WVVGTCPEG KPUVGCF QH XQKEG EJCTCEVGTKUVKEU 9G JCXG UJQYP VJCV 8+8 ECP CEJKGXG XGT[ IQQF CEEWTCE[ D[ CRRN[KPI C UGSWGPVKCN XGTKſECVKQP VGEJPKSWG *QYGXGT UKPEG 8+8 KU VQ XGTKH[ VJG XGTDCN EQPVGPV KPUVGCF QH VJG XQKEG EJCTCEVGTKUVKEU KV KU VJG WUGTUŏ TGURQPUKDKNKV[ VQ RTQVGEV VJGKT RGTUQPCN KPHQTOCVKQP HTQO KORQUVQTU 6Q KORTQXG VJG WUGT EQPXGPKGPEG CPF U[UVGO RGTHQTOCPEG YG HWTVJGT EQODKPGF XGT DCN KPHQTOCVKQP XGTKſECVKQP CPF URGCMGT XGTKſECVKQP VQ EQPUVTWEV C RTQITGUUKXG KPVG ITCVGF URGCMGT CWVJGPVKECVKQP U[UVGO +P VJG U[UVGO 8+8 KU WUGF VQ XGTKH[ C WUGT FWTKPI VJG ſTUV HGY CEEGUUGU 5KOWNVCPGQWUN[ VJG U[UVGO EQNNGEVU XGTKſGF VTCKPKPI FCVC HQT EQPUVTWEVKPI URGCMGTFGRGPFGPV OQFGNU .CVGT VJG U[UVGO OKITCVGU VQ CP 58 U[UVGO HQT CWVJGPVKECVKQP 6JG EQODKPGF U[UVGO KU EQPXGPKGPV VQ WUGTU UKPEG VJG[ ECP UVCTV VQ WUG VJG U[UVGO YKVJQWV IQKPI VJTQWIJ C HQTOCN GPTQNNOGPV UGUUKQP CPF YCKVKPI HQT OQFGN VTCKPKPI (WTVJGTOQTG UKPEG VJG VTCKPKPI FCVC OC[ DG EQNNGEVGF HTQO FKHHGTGPV EJCPPGNU KP FKHHGTGPV 8+8 UGUUKQPU VJG CEQWUVKE OKUOCVEJ RTQDNGO KU OKVKICVGF RQVGPVKCNN[ NGCFKPI VQ C DGVVGT U[UVGO RGTHQTOCPEG KP VGUV UGUUKQPU 6JG 5& *//U ECP DG WRFCVGF VQ EQXGT FKHHGTGPV CEQWUVKE GPXKTQPOGPVU YJKNG VJG U[UVGO KU KP WUG VQ HWTVJGT KORTQXG VJG U[UVGO RGTHQTOCPEG 8+8 ECP CNUQ DG WUGF VQ GPUWTG VTCKPKPI FCVC HQT 58 (QT FKHHGTGPV CRRNKECVKQPU XCTKQWU CWVJGPVKECVKQP U[UVGOU ECP DG FGUKIPGF DCUGF QP VJG VJGQT[ CPF VGEJPKSWGU RTGUGPVGF KP VJKU EJCRVGT # IQQF
E\&5&3UHVV//&
URGCMGT CWVJGPVKECVKQP U[UVGO HQT TGCN CRRNKECVKQPU EQWNF EQOG HTQO C RTQRGT KPVG ITCVKQP QH URGCMGT XGTKſECVKQP XGTDCN KPHQTOCVKQP XGTKſECVKQP URGGEJ TGEQIPKVKQP CPF VGZVVQURGGEJ U[UVGOU
References =? 6 9 #PFGTUQP An Introduction to Multivariate Statistical Analysis UGEQPF GFKVKQP ,QJP 9KNG[ 5QPU 0GY ;QTM =? $ 5 #VCN 'HHGEVKXGPGUU QH NKPGCT RTGFKEVKQP EJCTCEVGTKUVKEU QH VJG URGGEJ YCXG HQT CWVQOCVKE URGCMGT KFGPVKſECVKQP CPF XGTKſECVKQP Journal of the Acoustical Society of America Ō =? $ 5 #VCN #WVQOCVKE TGEQIPKVKQP QH URGCMGTU HTQO VJGKT XQKEGU Proceeding of the IEEE Ō =? . 4 $CJN 2 ( $TQYP 2 8 FG 5QW\C CPF 4 . /GTEGT /CZKOWO OWVWCN KPHQTOCVKQP GUVKOCVKQP QH JKFFGP /CTMQX OQFGN RCTCOGVGTU HQT URGGEJ TGEQI PKVKQP +P Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing RCIGU Ō 6QM[Q =? .' $CWO 6 2GVTKG ) 5QWNGU CPF 0 9GKUU # OCZKOK\CVKQP VGEJPKSWG QEEWTTKPI KP VJG UVCVKUVKECN CPCN[UKU QH RTQDCDKNKUVKE HWPEVKQPU QH /CTMQX EJCKPU Ann. Math. Stat. Ō =? , 2 %CORDGNN 5RGCMGT TGEQIPKVKQP # VWVQTKCN Proceedings of the IEEE Ō 5GRV =? 9 %JQW %* .GG CPF $* ,WCPI 5GIOGPVCN )2& VTCKPKPI QH *// DCUGF URGGEJ TGEQIPK\GT +P Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing RCIGU Ō 5CP (TCPEKUEQ /CTEJ =? # 2 &GORUVGT 0 / .CKTF CPF & $ 4WDKP /CZKOWO NKMGNKJQQF HTQO KPEQORNGVG FCVC XKC VJG '/ CNIQTKVJO Journal of Royal Statistical Society Ō =? 4 1 &WFC 2 ' *CTV CPF & ) 5VQTM Pattern Classification, Second Edition ,QJP 9KNG[ 5QPU 0GY ;QTM =? ) & (QTPG[ 6JG 8KVGTDK CNIQTKVJO Proceeding of IEEE Ō /CTEJ =? - (WMWPCIC Introduction to Statistical Pattern Recognition UGEQPF GFKVKQP #ECFGOKE 2TGUU +PE 0GY ;QTM =? 5 (WTWK %GRUVTCN CPCN[UKU VGEJPKSWGU HQT CWVQOCVKE URGCMGT XGTKſECVKQP IEEE Trans. Acoust., Speech, Signal Processing Ō #RTKN
E\&5&3UHVV//&
=? * )KUJ CPF / 5EJOKFV 6GZVKPFGRGPFGPV URGCMGT KFGPVKſECVKQP IEEE Signal Processing Magazine RCIGU Ō 1EV =? $* ,WCPI /CZKOWONKMGNKJQQF GUVKOCVKQP HQT OKZVWTG OWNVKXCTKCVG UVQEJCUVKE QDUGTXCVKQPU QH /CTMQX EJCKPU AT&T Technical Journal Ō ,WN[CWIWUV =? $* ,WCPI 9 %JQW CPF %* .GG /KPKOWO ENCUUKſECVKQP GTTQT TCVG OGVJQFU HQT URGGEJ TGEQIPKVKQP IEEE Trans. on Speech and Audio Process. Ō /C[ =? $* ,WCPI CPF 5 -CVCIKTK &KUETKOKPCVKXG NGCTPKPI HQT OKPKOWO GTTQT ENCU UKſECVKQP IEEE Transactions on Signal Processing Ō &G EGODGT =? 6 -CYCJCTC %* .GG CPF $* ,WCPI %QODKPKPI MG[RJTCUG FGVGEVKQP CPF UWDYQTFDCUGF XGTKſECVKQP HQT ƀGZKDNG URGGEJ WPFGTUVCPFKPI +P Proceedings of ICASSP RCIGU Ō /WPKEJ /C[ =? ( -QTMOC\UMK[ CPF $* ,WCPI &KUETKOKPCVKXG CFCRVCVKQP HQT URGCMGT XGTKſ ECVKQP +P Proceedings of Int. Conf. on Spoken Language Processing XQNWOG RCIGU Ō 2JKNCFGNRJKC =? %* .GG $* ,WCPI 9 %JQW CPF , , /QNKPC2GTG\ # UVWF[ QP VCUM KPFGRGPFGPV UWDYQTF UGNGEVKQP CPF OQFGNKPI HQT URGGEJ TGEQIPKVKQP +P Proc. of ICSLP RCIGU RR Ō 2JKNCFGNRJKC 1EV =? 3 .K # FGVGEVKQP CRRTQCEJ VQ UGCTEJURCEG TGFWEVKQP HQT *// UVCVG CNKIP OGPV KP URGCMGT XGTKſECVKQP IEEE Trans. on Speech and Audio Processing Ō ,WN[ =? 3 .K CPF $* ,WCPI 5RGCMGT XGTKſECVKQP WUKPI XGTDCN KPHQTOCVKQP XGTK ſECVKQP HQT CWVQOCVKE GPTQNNOGPV +P Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing 5GCVVNG /C[ =? 3 .K $* ,WCPI %* .GG 3 <JQW CPF ( - 5QQPI 4GEGPV CFXCPEGOGPVU KP CWVQOCVKE URGCMGT CWVJGPVKECVKQP IEEE Robotics & Automation magazine Ō /CTEJ =? 3 .K $* ,WCPI 3 <JQW CPF %* .GG 8GTDCN KPHQTOCVKQP XGTKſECVKQP +P Proceedings of EUROSPEECH RCIGU Ō 4JQFG )TGGEG 5GRV =? 3 .K $* ,WCPI 3 <JQW CPF %* .GG #WVQOCVKE XGTDCN KPHQTOCVKQP XGT KſECVKQP HQT WUGT CWVJGPVKECVKQP IEEE Trans. on Speech and Audio Processing Ō 5GRV =? 3 .K 5 2CTVJCUCTCVJ[ CPF # ' 4QUGPDGTI # HCUV CNIQTKVJO HQT UVQEJCUVKE OCVEJKPI YKVJ CRRNKECVKQP VQ TQDWUV URGCMGT XGTKſECVKQP +P Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing RCIGU Ō /WPKEJ #RTKN
E\&5&3UHVV//&
=? 3 .K 5 2CTVJCUCTCVJ[ # ' 4QUGPDGTI CPF & 9 6WHVU 0QTOCNK\GF FKU ETKOKPCPV CPCN[UKU YKVJ CRRNKECVKQP VQ C J[DTKF URGCMGTXGTKſECVKQP U[UVGO +P IEEE International Conference on Acoustics, Speech, and Signal Processing #VNCPVC /C[ =? 3 .K CPF # 6UCK # OCVEJGF ſNVGT CRRTQCEJ VQ GPFRQKPV FGVGEVKQP HQT TQDWUV URGCMGT XGTKſECVKQP +P Proceedings of IEEE Workshop on Automatic Identification 5WOOKV 0, 1EV =? 3 .K , <JGPI # 6UCK CPF 3 <JQW 4QDWUV GPFRQKPV FGVGEVKQP CPF GPGTI[ PQTOCNK\CVKQP HQT TGCNVKOG URGGEJ CPF URGCMGT TGEQIPKVKQP IEEE Trans. on Speech and Audio Processing Ō /CTEJ =? % 5 .KW %* .GG 9 %JQW $* ,WCPI CPF # ' 4QUGPDGTI # UVWF[ QP OKPKOWO GTTQT FKUETKOKPCVKXG VTCKPKPI HQT URGCMGT TGEQIPKVKQP Journal of the Acoustical Society of America Ō ,CPWCT[ =? ' .NGKFC CPF 4 % 4QUG 'HſEKGPV FGEQFKPI CPF VTCKPKPI RTQEGFWTGU HQT WV VGTCPEG XGTKſECVKQP KP EQPVKPWQWU URGGEJ TGEQIPKVKQP +P Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing RCIGU Ō #VNCPVC /C[ =? , 0G[OCP CPF ' 5 2GCTUQP 1P VJG WUG CPF KPVGTRTGVCVKQP QH EGTVCKP VGUV ETKVGTKC HQT RWTRQUG QH UVCVKUVKECN KPHGTGPEG Biometrika #2V + Ō 2V ++ =? , 0G[OCP CPF ' 5 2GCTUQP 1P VJG RTQDNGO QH VJG OQUV GHſEKGPV VGUVU QH UVCVKUVKECN J[RQVJGUGU Phil. Trans. Roy. Soc. A Ō =? ; 0QTOCPFKP 4 %CTFKP CPF 4 & /QTK *KIJRGTHQTOCPEG EQPPGEVGF FKIKV TGEQIPKVKQP WUKPI OCZKOWO OWVWCN KPHQTOCVKQP GUVKOCVKQP IEEE Trans. on Speech and Audio Processing Ō #RTKN =? 5 2CTVJCUCTCVJ[ CPF # ' 4QUGPDGTI )GPGTCN RJTCUG URGCMGT XGTKſECVKQP WU KPI UWDYQTF DCEMITQWPF OQFGNU CPF NKMGNKJQQFTCVKQ UEQTKPI +P Proceedings of ICSLP-96 2JKNCFGNRJKC 1EVQDGT =? . 4CDKPGT CPF $* ,WCPI Fundamentals of speech recognition 264 2TGP VKEG *CNN 'PINGYQQF %NKHHU 0, =? . 4 4CDKPGT , ) 9KNRQP CPF $* ,WCPI # UGIOGPVCN MOGCPU VTCKP KPI RTQEGFWTG HQT EQPPGEVGF YQTF TGEQIPKVKQP AT&T Technical Journal Ō /C[,WPG =? / ) 4CJKO %* .GG CPF $* ,WCPI 4QDWUV WVVGTCPEG XGTKſECVKQP HQT EQPPGEVGF FKIKVU TGEQIPKVKQP +P Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing RCIGU Ō &GVTQKV /C[ =? / ) 4CJKO %* .GG $* ,WCPI CPF 9 %JQW &KUETKOKPCVKXG WVVGT CPEG XGTKſECVKQP WUKPI OKPKOWO UVTKPI XGTKſECVKQP GTTQT /58' VTCKPKPI +P Proc. IEEE Int. Conf. Acoustic, Speech, Signal Processing RCIGU Ō #VNCPVC /C[
E\&5&3UHVV//&
=? & 4G[PQNFU 4QDWUV VGZVKPFGRGPFGPV URGCMGT KFGPVKſECVKQP WUKPI )CWU UKCP OKZVWTG URGCMGT OQFGNU IEEE Trans. on Speech and Audio Processing Ō =? # ' 4QUGPDGTI #WVQOCVKE URGCMGT XGTKſECVKQP C TGXKGY Proceedings of the IEEE Ō #RTKN =? # ' 4QUGPDGTI CPF , &G.QPI *//DCUGF URGCMGT XGTKſECVKQP WUKPI C VGNGRJQPG PGVYQTM FCVCDCUG QH EQPPGEVGF FKIKVCN WVVGTCPEGU 6GEJPKECN /GOQ TCPFWO $.6/ #66 $GNN .CDQTCVQTKGU &GEGODGT =? # ' 4QUGPDGTI CPF 5 2CTVJCUCTCVJ[ 5RGCMGT DCEMITQWPF OQFGNU HQT EQP PGEVGF FKIKV RCUUYQTF URGCMGT XGTKſECVKQP +P Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing RCIGU Ō #VNCPVC /C[ =? # ' 4QUGPDGTI 1 5KQJCP CPF 5 2CTVJCUCTCVJ[ 5RGCMGT XGTKſECVKQP WUKPI OKPKOWO XGTKſECVKQP GTTQT VTCKPKPI +P Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing RCIGU Ō 5GCVVNG /C[ =? # 4 5GVNWT 4 # 5WMMCT CPF , ,CEQD %QTTGEVKPI TGEQIPKVKQP GTTQTU XKC FKUETKOKPCVKXG WVVGTCPEG XGTKſECVKQP +P Proc. Int. Conf. on Spoken Language Processing RCIGU Ō 2JKNCFGNRJKC 1EV =? 1 5KQJCP # ' 4QUGPDGTI CPF 5 2CTVJCUCTCVJ[ 5RGCMGT KFGPVKſECVKQP WUKPI OKPKOWO XGTKſECVKQP GTTQT VTCKPKPI +P Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing RCIGU Ō 5GCVVNG /C[ =? ( - 5QQPI # ' 4QUGPDGTI CPF $* ,WCPI # XGEVQT SWCPVK\C VKQP CRRTQCEJ VQ URGCMGT TGEQIPKVKQP AT&T Technical Journal Ō /CTEJ#RTKN =? 4 # 5WMMCT CPF %* .GG 8QECDWNCT[ KPFGRGPFGPV FKUETKOKPCVKXG WVVGTCPEG XGTKſECVKQP HQT PQPMG[YQTF TGLGEVKQP KP UWDYQTF DCUGF URGGEJ TGEQIPKVKQP IEEE Trans. Speech and Audio Process. Ō 0QXGODGT =? 4 # 5WMMCT # 4 5GVNWT / ) 4CJKO CPF %* .GG 7VVGTCPEG XGTKſECVKQP QH MG[YQTF UVTKPI WUKPI YQTFDCUGF OKPKOWO XGTKſECVKQP GTTQT 9$/8' VTCKPKPI +P Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing RCIGU Ō #VNCPVC /C[ =? # , 8KVGTDK 'TTQT DQWPFU HQT EQPXQNWVKQPCN EQFGU CPF CP CU[ORVQVKECNN[ QRVKOCN FGEQFKPI CNIQTKVJO IEEE Transactions on Information Theory +6 Ō #RTKN =? # 9CNF Sequential analysis %JCROCP *CNN 0; =? , ) 9KNRQP .4 4CDKPGT %* .GG CPF ' )QNFOCP #WVQOCVKE TGEQIPK VKQP QH MG[YQTFU KP WPEQPUVTCKPGF URGGEJ WUKPI JKFFGP /CTMQX OQFGNU IEEE Trans. on Acoustics, Speech, and Signal Proc. Ō 0QXGODGT
E\&5&3UHVV//&
=? % ( , 9W 1P VJG EQPXGTIGPEG RTQRGTVKGU QH VJG '/ CNIQTKVJO The Annals of Statistics Ō
E\&5&3UHVV//&
8 HMMs for Language Processing Problems Richard M. Schwartz and John Makhoul BBN Technologies, Verizon
CONTENTS
+PVTQFWEVKQP 7UG QH 2TQDCDKNKVKGU 0COG 5RQVVKPI 6QRKE %NCUUKſECVKQP +PHQTOCVKQP 4GVTKGXCN 'XGPV 6TCEMKPI 7PUWRGTXKUGF 6QRKE &GVGEVKQP 5WOOCT[ 4GHGTGPEGU
6JKU EJCRVGT FGUETKDGU JKFFGP /CTMQX OQFGN *// OGVJQFU HQT XCTKQWU RTQDNGOU KP NCPIWCIG RTQEGUUKPI *//U RTQXKFG C RQYGTHWN CPF ƀGZKDNG HQTOCNKUO HQT OQF GNKPI UGSWGPEGU QH YQTFU 6JG[ CNNQY WU VQ GUVKOCVG VJG RQUVGTKQT RTQDCDKNKV[ VJCV C FQEWOGPV YQWNF DG TGNGXCPV VQ C WUGT IKXGP VJG WUGTŏU SWGT[ QT VQ EQORWVG VJG RTQDC DKNKV[ VJCV C FQEWOGPV FKUEWUUGU C RCTVKEWNCT UGV QH VQRKEU 6JG[ CNNQY WU VQ FGVGTOKPG CWVQOCVKECNN[ YJKEJ YQTFU CTG TGNCVGF VQ YJKEJ VQRKEU GXGP VJQWIJ GCEJ FQEWOGPV KU CPPQVCVGF YKVJ OWNVKRNG VQRKEU 6JG[ GXGP CNNQY WU VQ FGEQORQUG CP WPCPPQVCVGF EQTRWU QH FQEWOGPVU KPVQ KVU EQORQPGPV UGV QH VQRKE DCUKU HWPEVKQPU 9JKNG UQOG QH VJG *//U WUGF CTG GZVTGOGN[ UKORNG VJG[ CHHQTF C RCTCFKIO HQT OQFGN RCTCOGVGT GUVKOCVKQP CPF QHHGT VJG RQUUKDKNKV[ QH WUKPI OQTG RQYGTHWN OQFGNU KP VJG HWVWTG +P KPHQTOCVKQP GZVTCEVKQP YG ECP WUG *//U VQ GUVKOCVG VJG RTQDCDKNKV[ VJCV C UGSWGPEG QH YQTFU KP C RCTVKEWNCT EQPVGZV KU C PCOG QH C RCTVKEWNCT V[RG QT VJCV VYQ GPVKVKGU KP C VGZV CTG TGNCVGF KP C RCTVKEWNCT YC[ 9G FGUETKDG UKORNG *//U WUGF HQT VJGUG CPF QVJGT NCPIWCIG RTQEGUUKPI VCUMU
8.1 Introduction *//U JCXG DGGP WUGF HQT VJG NCUV FGECFG CU VJG RTGHGTTGF OGVJQF HQT URGGEJ TGEQI PKVKQP 6JKU KU DGECWUG VJG[ RTQXKFG C UKORNG CPF ƀGZKDNG OGEJCPKUO HQT OQFGN KPI UGSWGPEGU QH XCTKCDNG NGPIVJ 6JG QDUGTXCVKQP UGSWGPEG KP URGGEJ KU C JKIJ
E\&5&3UHVV//&
FKOGPUKQPCN XGEVQT VJCV KU JCTF VQ XKUWCNK\G #NVJQWIJ YG MPQY VJG EQPFKVKQPCN KP FGRGPFGPEG CUUWORVKQPU VJCV FGſPG VJG *// CTG PQV TGCNN[ VTWG HQT URGGEJ VJG QXGTYJGNOKPI CFXCPVCIGU QH WUKPI C TKIQTQWU RTQDCDKNKUVKE HQTOCNKUO UVKNN TGUWNV KP JKIJ CEEWTCE[ 1XGT VJG [GCTU TGUGCTEJGTU JCXG HQWPF XCTKQWU YC[U QH OQFGNKPI VJG FGRGPFGPEG VQ UQOG FGITGG +P EQPVTCUV VQ URGGEJ TGEQIPKVKQP RTQDNGOU KP VGZV RTQEGUUKPI CV ſTUV INCPEG UGGO FGEGRVKXGN[ UKORNG 9G ECP UGG VJG YQTFU CPF YG SWKEMN[ FGXGNQR OCP[ VJGQTKGU CU VQ JQY YQTFU EQPXG[ OGCPKPI CPF ſV VQIGVJGT VQ EQPUVTWEV OQTG EQORNGZ OGCPKPIU +V KU QHVGP RQUUKDNG VQ DWKNF RTKOKVKXG U[UVGOU YKVJ C UOCNN PWODGT QH CF JQE TWNGU 6JGTGHQTG VJGTG KU QHVGP ITGCVGT TGUKUVCPEG VQ WUKPI RTQDCDKNKUVKE OGVJQFU HQT VGZV RTQDNGOU YJGTG TWNGDCUGF OGVJQFU UGGO OQTG KPVWKVKXG +P VJKU EJCRVGT YG UJQY VJCV VJG UCOG CFXCPVCIGU VJCV JQNF HQT URGGEJ CNUQ CRRN[ VQ QVJGT NCPIWCIG RTQDNGOU 9G FKUEWUU VYQ FKHHGTGPV ENCUUGU QH NCPIWCIG RTQEGUUKPI RTQDNGOU +P VJG ſTUV ENCUU YG RGTHQTO CP QRGTCVKQP QP C YJQNG FQEWOGPV # FQEWOGPV KU CP[ UK\CDNG WPKV QH VGZV UWEJ CU C UVQT[ C OGUUCIG GVE 6JKU KPENWFGU HQT GZCORNG KPHQTOCVKQP TGVTKGXCN TGVTKGXKPI FQEWOGPVU KP TGURQPUG VQ UQOG SWGT[ CPF VQRKE ENCUUKſECVKQP CUUKIPKPI QPG QT OQTG ECVGIQTKGU HTQO C ſZGF UGV VQ C FQEWOGPV 6JG UGEQPF ENCUU QH RTQDNGOU KU EQOOQPN[ ECNNGF +PHQTOCVKQP 'ZVTCEVKQP *GTG YG CVVGORV VQ WPFGTUVCPF VJG OGCPKPI QH UQOG QH VJG VGZV KP UQOG YC[ 6JKU KPENWFGU HQT GZCORNG GZVTCEVKPI RCTVU QH URGGEJ FGVGEVKPI CPF ECVGIQTK\KPI PCOGU QT FGVGEVKPI TGNCVKQPU COQPI GPVKVKGU FGUETKDGF KP VJG VGZV 6JG HQTOGT RTQDNGO KU VTCFKVKQPCNN[ RGTHQTOGF WUKPI QPG QH UGXGTCN UKOKNCTKV[ OGCUWTGU VJCV JCXG DGGP FGXGNQRGF QXGT VJG [GCTU 6JG NCVVGT KU VTCFKVKQPCNN[ RGTHQTOGF WUKPI C UGV QH JCPF EQPUVTWEVGF TWNGU 9G EQPVTCUV VJG WUG QH *//U YKVJ VYQ QVJGT EQOOQP CRRTQCEJGU WUGF HQT NCPIWCIG RTQEGUUKPI TWNGU CPF CF JQE UKOKNCTKV[ OGCUWTGU
8.2 Use of Probabilities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ſEKGPVN[ DCF VJGP VJG TGUWNVKPI GTTQTU KP VJG RTQDCDKNKVKGU YG QDVCKP ECP OQTG VJCP QHHUGV VJG DGPGſV HQT WUKPI RTQDCDKNKVKGU KP VJG
E\&5&3UHVV//&
ſTUV RNCEG 2TQDCDKNKUVKE OGVJQFU CNOQUV CNYC[U EQOG YKVJ C OQFGN HQT VJG RTQDNGO +H YG JCXG C IQQF OQFGN HQT VJG RTQDNGO VJGP YG CNUQ OC[ IGV QVJGT CFXCPVCIGU UWEJ CU OGVJQFU VJCV NGCTP HTQO GZCORNGU TCVJGT VJCP OGVJQFU VJCV PGGF VQ DG RTQITCOOGF GZRNKEKVN[ $WV C UJQTV RTQITCO QT UGV QH TWNGU ECP UQOGVKOGU FQ DGVVGT VJCP C DCF RTQDCDKNKUVKE OQFGN 5Q VJG ſTUV CPUYGT VQ VJG SWGUVKQP QH YJGVJGT YG UJQWNF WUG RTQDCDKNKUVKE OQFGNU KU Œ+V FGRGPFU QP JQY ECTGHWN YG CTG KP WUKPI RTQDCDKNKVKGUŒ 9G JCXG HQWPF VJCV KH QPG KU ECTGHWN KV KU WUWCNN[ RQUUKDNG VQ QDVCKP JKIJ RGTHQTOCPEG
EQORCTCDNG VQ QT GZEGGFKPI VJG UVCVGQHVJGCTV WUKPI VJGUG OGVJQFU (WTVJGTOQTG YKVJ VJG UQNKF OCVJGOCVKECN DCUKU HQT YJCV KU DGKPI FQPG YG ECP QHVGP CPCN[\G YJCV KU YTQPI YKVJ QWT OQFGNU WUWCNN[ VJG CUUWORVKQPU CPF KORTQXG QP VJGO KH FGUKTGF
8.2.1 Hidden Markov Models +H YG CEEGRV VJCV YG UJQWNF WUG RTQDCDKNKUVKE OQFGNU YJ[ FQ YG WUG JKFFGP /CTMQX OQFGNU *//U HQT UQ OCP[ FKHHGTGPV RTQDNGOU! +V OC[ UGGO VQ UQOG NKMG VJG QPN[ VQQN YG JCXG KU VJG RTQXGTDKCN JCOOGT UQ GXGT[VJKPI NQQMU VQ WU NKMG C PCKN 9JKNG VJKU OKIJV DG VTWG VQ UQOG GZVGPV *//U CTG CRRTQRTKCVG HQT OQFGNKPI PQKU[ UGSWGPEGU #NN NCPIWCIG RTQDNGOU URGGEJ VGZV YTKVKPI EQPUKUV QH UGSWGPEGU 6JG UGSWGPEGU CTG QDXKQWUN[ PQKU[ QT GNUG KV YQWNF DG UKORNG VQ YTKVG RTQITCOU VQ UQNXG CNN QWT NCPIWCIG RTQDNGOU 6JG OQUV UGTKQWU RTQDNGOU VJCV YG HCEG YKVJ *//U KU VJG EQPFKVKQPCN KPFGRGPFGPEG CUUWORVKQP 6JG VTCPUKVKQP HTQO QPG UVCVG VQ VJG PGZV FGRGPFU QPN[ QP VJG UVCVG CPF PQV QP JQY YG IGV VJGTG JQY NQPI YG JCXG DGGP VJGTG QT YJCV U[ODQNU YGTG RTGXK QWUN[ GOKVVGF YJKNG YG YGTG KP VJCV UVCVG 5KOKNCTN[ VJG U[ODQN VQ DG GOKVVGF HTQO C UVCVG FGRGPFU QPN[ QP DGKPI CV VJCV UVCVG CPF PQVJKPI CDQWV VJG JKUVQT[ KPENWF KPI RTGXKQWU UVCVGU QT RTGXKQWUN[ GOKVVGF U[ODQNU 6JKU KPFGRGPFGPEG CUUWORVKQP KU RCVGPVN[ HCNUG HQT OQUV NCPIWCIG RTQDNGOU +P URGGEJ YG MPQY VJCV VJGTG KU C JKIJ EQTTGNCVKQP COQPI UWEEGUUKXG URGEVTC +P VGZV YG MPQY VJCV GCEJ YQTF FGRGPFU VQ C XGT[ NCTIG GZVGPV QP VJG RTGEGFKPI YQTFU DQVJ KOOGFKCVGN[ RTGEGFKPI CPF HWTVJGT DCEM 5Q YJ[ FQ VJGUG OQFGNU YQTM UQ YGNN! (KTUV OQUV QH VJG FGRGPFGPEG VJCV YG KIPQTG ECP DG VJQWIJV QH CU TGFWPFCPV RQUKVKXG EQTTGNCVKQP 6JCV KU KH YG VTGCV GCEJ QDUGT XCVKQP CU PGY CPF KPFGRGPFGPV VJG RTQDCDKNKV[ VJCV YG EQORWVG HQT KV OC[ DG NQYGT VJCP KV UJQWNF DG JCF YG VCMGP KPVQ CEEQWPV VJG RTGXKQWU QDUGTXCVKQPU $WV VJKU GTTQT OC[ DG VJQWIJV QH CU TGNCVKXGN[ WPKHQTO CETQUU OQUV EJQKEGU #PF VJG DGPGſVU HQT WUKPI *//U CTG NCTIG 6JG[ CHHQTF WU YKVJ C YGNNGUVCDNKUJGF UGV QH OGVJQFU CPF OCVJGOCVKEU HQT OCPKRWNCVKPI RTQDNGOU KP VJG YC[ YG PGGF VQ 2TQDCDKNKUVKE OQFGNU KP IGPGTCN CPF *//U KP RCTVKEWNCT JCXG UGXGTCN CFXCPVCIGU (KTUV CU UVCVGF CDQXG VJG[ RTQXKFG WU YKVJ C YGNNFGſPGF OCVJGOCVKECN CRRTQCEJ VQ UQNXKPI RCVVGTP TGEQIPKVKQP RTQDNGOU KP NCPIWCIG 5GEQPF VJG[ QHVGP RTQXKFG WU YKVJ UKORNG OGVJQFU HQT FGXGNQRKPI OQFGNU QP PGY FQOCKPU QT NCPIWCIGU TGSWKTKPI QPN[ C UGV QH FCVC YKVJ CPPQVCVGF CPUYGTU (KPCNN[ OCP[ NCPIWCIG RTQDNGOU VJCV FQ PQV WUG RTQDCDKNKUVKE OQFGNU WUG C UWO QH UEQTGU (QT GZCORNG VJG EQPXGPVKQPCN OGVTKE HQT EQORCTKPI FQEWOGPVU HQT +PHQTOCVKQP 4GVTKGXCN UWOU WR C UEQTG HQT GCEJ
E\&5&3UHVV//&
YQTF KP VJG FQEWOGPV VJCV OCVEJGU C SWGT[ YQTF +P QTFGT HQT VJKU UWO VQ DG CP CRRTQRTKCVG OGCUWTG QH TGNGXCPEG VJGUG UEQTGU OWUV DG NQI RTQDCDKNKVKGU +P VJG TGOCKPKPI UGEVKQPU YG FKUEWUU UGXGTCN RTQDNGOU KP VGZV RTQEGUUKPI PCOG URQV VKPI VQRKE ENCUUKſECVKQP KPHQTOCVKQP TGVTKGXCN GXGPV VTCEMKPI CPF WPUWRGTXKUGF VQRKE FKUEQXGT[ 6JG ſ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
8.3 Name Spotting 6JG QDLGEVKXG QH PCOG URQVVKPI KU VQ GZVTCEV KORQTVCPV VGTOU HTQO VJG URGGEJ CPF EQNNGEV VJGO KP C FCVCDCUG (QT GZCORNG KP PGYU KV KU WUGHWN VQ NQECVG PCOGU QH RGTUQPU RNCEGU CPF QTICPK\CVKQPU /QUV QH VJG RTGXKQWU YQTM KP VJKU CTGC JCU EQP UKFGTGF QPN[ VGZV UQWTEGU QH YTKVVGP NCPIWCIG CPF JCU EQPEGPVTCVGF QP VJG FGUKIP QH TWNGFTKXGP CNIQTKVJOU VQ NQECVG VJG PCOGU 'ZVTCEVKQP HTQO CWVQOCVKE VTCPUETKRVKQPU QH URQMGP NCPIWCIG KU OQTG FKHſEWNV VJCP YTKVVGP VGZV FWG VQ VJG CDUGPEG QH ECRKVCN K\CVKQP RWPEVWCVKQP CPF UGPVGPEG DQWPFCTKGU CU YGNN CU VJG RTGUGPEG QH TGEQIPKVKQP GTTQTU 6JGUG JCXG UKIPKſECPV FGITCFKPI GHHGEVU QP VJG RGTHQTOCPEG QH TWNGFTKXGP U[UVGOU 6Q QXGTEQOG VJGUG RTQDNGOU YG JCXG FGXGNQRGF CP *//DCUGF PCOG GZ VTCEVKQP U[UVGO ECNNGF +FGPVK(KPFGT =? 6JG VGEJPKSWG TGSWKTGU QPN[ VJCV YG RTQXKFG VTCKPKPI VGZV YKVJ VJG V[RG CPF NQECVKQP QH VJG PCOGF GPVKVKGU OCTMGF 6JG U[UVGO JCU VJG CFFKVKQPCN CFXCPVCIG VJCV KV KU GCUKN[ RQTVGF VQ QVJGT NCPIWCIGU TGSWKTKPI QPN[ C UGV QH CPPQVCVGF VTCKPKPI FCVC HTQO C PGY NCPIWCIG 6JG PCOG URQVVKPI RTQDNGO ECP DG TGFGſPGF CU JCXKPI VQ KFGPVKH[ VJG V[RG QH CNN VJG YQTFU KP C FQEWOGPV 9G OWUV ſPF CNN GZCORNGU QH PCOGU QH RGQRNG RNCEGU CPF QTICPK\CVKQPU 6JG TGOCKPKPI VGZV OWUV DG EQTTGEVN[ ENCUUKſGF CU PQV DGNQPIKPI VQ CP[ QH VJGUG V[RGU 6JG OQFGN VJCV YG WUG TGƀGEVU VJKU VCUM 6JG OQFGN EQPUKUVU QH QPG UVCVG HQT GCEJ QH VJG VJTGG PCOGF GPVKVKGU RNWU QPG UVCVG IGPGTCN NCPIWCIG HQT CNN QVJGT YQTFU KP VJG VGZV YKVJ VTCPUKVKQPU HTQO GCEJ UVCVG VQ GXGT[ QVJGT UVCVG #UUQEKCVGF YKVJ GCEJ QH VJG UVCVGU KU C DKITCO UVCVKUVKECN OQFGN QP CNN YQTFU KP VJG XQECDWNCT[ # FKHHGTGPV DKITCO OQFGN KU GUVKOCVGF HQT GCEJ QH VJG UVCVGU $[ VJKPMKPI
E\&5&3UHVV//&
QH VJKU CU C IGPGTCVKXG OQFGN VJCV IGPGTCVGU CNN VJG YQTFU KP VJG VGZV OQUV QH VJG VKOG YG CTG KP VJG ). UVCVG GOKVVKPI IGPGTCN NCPIWCIG YQTFU 9G VJGP VTCPUKVKQP VQ QPG QH VJG PCOGFGPVKV[ UVCVGU KH YG YCPV VQ IGPGTCVG C PCOG YG UVC[ KPUKFG VJG UVCVG IGPGTCVKPI VJG YQTFU HQT VJCV PCOG 6JGP YG GKVJGT VTCPUKVKQP VQ CPQVJGT PCOGF GPVKV[ UVCVG QT OQTG NKMGN[ DCEM VQ VJG ). UVCVG 6JG FGEKUKQP VQ GOKV GCEJ YQTF QT VQ VTCPUKVKQP VQ CPQVJGT UVCVG FGRGPFU QP VJG RTGXKQWU YQTF CPF VJG RTGXKQWU UVCVG +P VJKU YC[ VJG OQFGN WUGU EQPVGZV VQ JGNR FGVGEV CPF ENCUUKH[ PCOGU (QT GZCORNG VJG YQTF Œ/TŒ KP VJG ). UVCVG KU NKMGN[ VQ DG HQNNQYGF D[ C VTCPUKVKQP VQ VJG 2'4510 UVCVG #HVGT VJG RGTUQPŏU PCOG KU IGPGTCVGF C VTCPUKVKQP VQ VJG ). UVCVG KU NKMGN[ CPF IGPGTCN YQTFU NKMG ŒUCKFŒ QT ŒFGRCTVGFŒ OC[ HQNNQY 6JGUG EQPVGZVFGRGPFGPV GHHGEVU CTG KPENWFGF KP QWT OQFGN 6JG RCTCOGVGTU QH VJG OQFGN CTG GUVKOCVGF CWVQOCVKECNN[ HTQO CPPQVCVGF VTCKPKPI FCVC YJGTG VJG VJTGG UGVU QH PCOGF GPVKVKGU CTG OCTMGF KP VJG VGZV 6JGP IKXGP C VGUV UCORNG VJG OQFGN KU WUGF VQ GUVKOCVG VJG RTQDCDKNKV[ QH GCEJ YQTF DGNQPIKPI VQ QPG QH VJG VJTGG PCOGF GPVKVKGU QT VQ PQPG 9G VJGP WUG VJG 8KVGTDK CNIQTKVJO =? VQ ſPF VJG OQUV NKMGN[ UGSWGPEG QH UVCVGU VQ CEEQWPV HQT VJG VGZV 6JG TGUWNV KU VJG CPUYGT HQT VJG UGSWGPEG QH PCOGF GPVKVKGU 5KPEG QWT U[UVGO JCU DGGP VTCKPGF QP QPN[ QPG OKNNKQP YQTFU QH CPPQVCVGF FCVC HTQO DTQCFECUV PGYU OCP[ QH VJG YQTFU KP CP KPFGRGPFGPV VGUV UGV YKNN DG WPMPQYP VQ VJG PCOG URQVVKPI U[UVGO GXGP VJQWIJ VJG[ OKIJV DG MPQYP VQ VJG URGGEJ TGEQIPK\GT
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ŏ70-0190ŏ 9G VJGP CTG CDNG VQ GUVKOCVG CNN VJG RTQDCDKNKVKGU KPXQNXKPI WPMPQYP YQTFU 6JG RTQDCDKNKVKGU HQT MPQYP YQTFU CTG GUVKOCVGF HTQO CNN QH VJG FCVC &WTKPI VJG VGUVKPI RJCUG YG TGRNCEG CP[ UVTKPI VJCV KU WPMPQYP VQ VJG PCOG URQVVKPI U[UVGO D[ VJG NCDGN ŏ70 -0190ŏ CPF YG CTG VJGP CDNG VQ ſPF VJG DGUV OCVEJKPI UGSWGPEG QH UVCVGU 9G JCXG HQWPF VJCV D[ OCMKPI RTQRGT WUG QH EQPVGZV OQTG VJCP JCNH QH VJG PCOGU VJCV YGTG PQV MPQYP VQ VJG PCOG URQVVKPI U[UVGO CTG NCDGNGF EQTTGEVN[ D[ VJG U[UVGO 1PG CFXCPVCIG QH QWT CRRTQCEJ VQ KPHQTOCVKQP GZVTCEVKQP KU VJG GCUG YKVJ YJKEJ YG ECP NGCTP VJG UVCVKUVKEU HQT FKHHGTGPV UV[NGU QH VGZV (QT GZCORNG NGV WU UC[ YG YCPV VJG U[UVGO VQ YQTM QP VGZV YKVJQWV ECUG KPHQTOCVKQP KG VJG VGZV KU FKURNC[GF CU GKVJGT CNN NQYGT ECUG QT CNN WRRGT ECUG +V KU C UKORNG OCVVGT VQ TGOQXG VJG ECUG KPHQTOCVKQP HTQO QWT CPPQVCVGF VGZV CPF VJGP TGGUVKOCVG VJG OQFGNU +H YG YCPV VQ WUG +FGPVK(KPFGT QP VJG QWVRWV QH C URGGEJ TGEQIPK\GT YG GZRGEV VJCV VJG VGZV YKNN PQV QPN[ DG ECUGNGUU DWV YKNN CNUQ JCXG PQ RWPEVWCVKQP +P CFFKVKQP VJGTG YKNN DG PQ CDDTGXKCVKQPU CPF PWOGTKE XCNWGU YKNN DG URGNNGF QWV GI 69'06; (174 TCVJGT VJCP #ICKP YG ECP GCUKN[ UKOWNCVG VJKU GHHGEV QP QWT CPPQVCVGF VGZV KP QTFGT VQ NGCTP C OQFGN QH VGZV QWVRWV HTQO C URGGEJ TGEQIPK\GT 1H EQWTUG IKXGP CPPQVCVGF FCVC HTQO C PGY NCPIWCIG KV KU C UKORNG OCVVGT VQ VTCKP VJG UCOG U[UVGO VQ TGEQIPK\G PCOGF GPVKVKGU KP VJCV NCPIWCIG
E\&5&3UHVV//&
9G JCXG RGTHQTOGF UGXGTCN GZRGTKOGPVU VQ OGCUWTG VJG RGTHQTOCPEG QH +FGPVK(KPFGT KP ſPFKPI PCOGU +P CFFKVKQP YG JCXG OGCUWTGF VJG FGITCFCVKQP YJGP ECUG CPF RWPEVWCVKQP KPHQTOCVKQP KU NQUV QT YJGP HCEGF YKVJ GTTQTU HTQO CWVQOCVKE URGGEJ TGEQIPKVKQP +P OGCUWTKPI VJG CEEWTCE[ QH VJG U[UVGO DQVJ VJG V[RG QH PCOGF GPVKV[ CPF VJG URCP QH VJG EQTTGURQPFKPI YQTFU KP VJG VGZV CTG VCMGP KPVQ EQPUKFGTCVKQP 9G OGCUWTG VJG UNQV GTTQT TCVG YJGTG VJG V[RG CPF URCP QH C PCOG KU GCEJ EQWPVGF CU C UGRCTCVG UNQV D[ FKXKFKPI VJG VQVCN PWODGT QH GTTQTU KP PCOGF GPVKVKGU UWDUVKVWVKQPU FGNGVKQPU CPF KPUGTVKQPU D[ VJG VQVCN PWODGT QH VTWG PCOGF GPVKVKGU KP VJG TGHGTGPEG CPUYGTU =? +P C VGUV HTQO VJG *# $TQCFECUV 0GYU EQTRWU =? YJGTG VJG PWODGT QH V[RGU QH PCOGF GPVKVKGU YCU UGXGP TCVJGT VJCP VJG VJTGG WUGF JGTG +FGPVK(KPFGT QDVCKPGF C UNQV GTTQT TCVG QH HQT VGZV YKVJ OKZGF ECUG CPF RWPEVWCVKQP 9JGP CNN ECUG CPF RWPEVWCVKQP YGTG TGOQXGF VJG UNQV GTTQT TCVG KPETGCUGF VQ QPN[ +P TGEGPV *# GXCNWCVKQPU QP PCOG URQVVKPI YKVJ URGGEJ KPRWV CICKP YKVJ UGXGP ENCUUGU QH PCOGU VJG UNQV GTTQT TCVG HQT VJG QWVRWV QH VJG $[DNQU URGGEJ TGEQIPK\GT YCU YKVJ C URGGEJ TGEQIPKVKQP YQTF GTTQT TCVG QH =? 9JGP CNN TGEQI PKVKQP GTTQTU YGTG EQTTGEVGF YKVJQWV CFFKPI CP[ ECUG QT RWPEVWCVKQP KPHQTOCVKQP VJG UNQV GTTQT TCVG FGETGCUGF VQ +P IGPGTCN YG JCXG HQWPF VJCV VJG PCOGF GPVKV[ UNQV GTTQT TCVG KPETGCUGU NKPGCTN[ YKVJ VJG YQTF GTTQT TCVG KP CRRTQZKOCVGN[ C QPGVQQPG HCUJKQP
8.4 Topic Classification /WEJ YQTM JCU DGGP FQPG KP VQRKE ENCUUKſECVKQP YJGTG VJG OQFGNU HQT VJG FKHHGT GPV VQRKEU CTG GUVKOCVGF KPFGRGPFGPVN[ GXGP KH OWNVKRNG VQRKEU CTG CUUKIPGF VQ GCEJ FQEWOGPV 1PG PQVCDNG GZEGRVKQP KU VJG YQTM QH ;CPI CPF %JWVG =? YJQ CU RCTV QH VJGKT OQFGN VCMG KPVQ EQPUKFGTCVKQP VJG HCEV VJCV OWNVKRNG UKOWNVCPGQWU VQRKEU CTG WUWCNN[ CUUQEKCVGF YKVJ GCEJ FQEWOGPV 1WT CRRTQCEJ VQ VQRKE ENCUUKſECVKQP KU UKO KNCT KP URKTKV VQ VJCV QH ;CPI CPF %JWVG GZEGRV VJCV YG WUG C $C[GUKCP HTCOGYQTM =? KPUVGCF QH C FKUVCPEGDCUGF CRRTQCEJ 1WT VQRKE ENCUUKſECVKQP EQORQPGPV ECNNGF 1P6QRKEÌ Å KU C RTQDCDKNKUVKE *// YJQUG RCTCOGVGTU CTG GUVKOCVGF HTQO VTCKP KPI UCORNGU QH FQEWOGPVU YKVJ IKXGP VQRKE NCDGNU YJGTG VJG VQRKE NCDGNU PWODGT KP VJG VJQWUCPFU 6JG OQFGN CNNQYU GCEJ YQTF KP VJG FQEWOGPV VQ EQPVTKDWVG FKHHGTGPV COQWPVU VQ GCEJ QH VJG VQRKEU CUUKIPGF VQ VJG FQEWOGPV 6JG QWVRWV HTQO 1P6QRKE KU C TCPMQTFGTGF NKUV QH CNN RQUUKDNG VQRKEU CPF EQTTGURQPFKPI UEQTGU HQT CP[ IKXGP FQEWOGPV
E\&5&3UHVV//&
G eneralLanguage
T0 n
P(Tj|Set) story start
P (W n|Tj)
T1 story end
T2 P(Set) . .
TM
Loop
FIGURE 8.1 A hidden Markov model for topics. Each state can emit words for one topic. State T0 emits words corresponding to general language.
8.4.1 The Model 9G EJQQUG VJG UGV QH VQRKEU VJCV EQTTGURQPFU VQ C IKXGP FQEWOGPV & UWEJ VJCV VJG RQUVGTKQT RTQDCDKNKV[ KU OCZKOK\GF
(QT VJG RWTRQUG QH TCPMKPI VJG UGVU QH VQRKEU ECP DG KIPQTGF 6JG RTKQT RTQD CDKNKV[ KU TGCNN[ VJG LQKPV RTQDCDKNKV[ QH C FQEWOGPV JCXKPI CNN VJG NCDGNU KP VJG UGV YJKEJ ECP DG CRRTQZKOCVGF WUKPI VQRKE EQQEEWTTGPEG RTQDCDKNKVKGU
½ ¾
YJGTG KU VJG PWODGT QH VQRKEU KP Set CPF VJG GZRQPGPV UGTXGU VQ RNCEG QP UKOKNCT HQQVKPI VQRKE UGVU QH FKHHGTGPV UK\GU KU GUVKOCVGF D[ VCMKPI VJG RTQFWEV QH VJG OCZKOWO NKMGNKJQQF GUVKOCVGU QH CPF 6JG HQTOGT KU GUVKOCVGF CU VJG HTCEVKQP QH VJQUG FQEWOGPVU YKVJ CU C VQRKE YJKEJ CNUQ JCXG CU C VQRKE CPF VJG NCVVGT KU GUVKOCVGF CU VJG HTCEVKQP QH FQEWOGPVU YKVJ CU C VQRKE 9JCV TGOCKPU VQ DG EQORWVGF KU VJG EQPFKVKQPCN RTQDCDKNKV[ QH VJG YQTFU KP VJG FQEWOGPV IKXGP VJCV VJG FQEWOGPV KU NCDGNGF YKVJ CNN VJG VQRKEU KP 5GV
E\&5&3UHVV//&
9G OQFGN VJKU RTQDCDKNKV[ YKVJ CP *// EQPUKUVKPI QH C UVCVG HQT GCEJ QH VJG VQRKEU KP VJG UGV RNWU QPG CFFKVKQPCN VQRKE UVCVG IGPGTCN NCPIWCIG ). CU UJQYP KP (KI 6JG OQFGN ŒIGPGTCVGUŒ VJG YQTFU KP VJG FQEWOGPV QPG D[ QPG ſTUV EJQQUKPI C VQRKE FKUVTKDWVKQP HTQO YJKEJ VQ FTCY VJG PGZV YQTF CEEQTFKPI VQ VJGP EJQQUKPI C YQTF CEEQTFKPI VQ VJGP EJQQUKPI CPQVJGT VQRKE FKUVTKDWVKQP VQ FTCY HTQO GVE 6JG HQTOWNC HQT KU VJGTGHQTG
¾
YJGTG XCTKGU QXGT VJG UGV QH YQTFU KP VJG FQEWOGPV 6JG GNGOGPVU QH VJG CDQXG GSWCVKQP CTG GUVKOCVGF HTQO VTCKPKPI FCVC CU FGUETKDGF DGNQY
8.4.2 Estimating HMM Parameters 9G WUG C DKCUGF HQTO QH VJG 'ZRGEVCVKQP/CZKOK\CVKQP '/ CNIQTKVJO =? VQ ſPF IQQF GUVKOCVGU HQT VJG VTCPUKVKQP RTQDCDKNKVKGU CPF VJG GOKUUKQP RTQDCDKNKVKGU KP VJG *// KP (KI 6JG VTCPUKVKQP RTQDCDKNKVKGU CTG FGſPGF D[
·½
VKOGU CP[ YQTF KU GOKVVGF KP UVCVG OQFGN VKOGU CP[ YQTF KU GOKVVGF KP CP[ UVCVG OQFGN
YJKEJ ECP DG GUVKOCVGF CU
·½ YJGTG
¾ ¾
YKVJ
KU VJG DKCU VGTO KU VJG PWODGT QH YQTFU KP VJG FQEWOGPV & CPF
JCU
¾
KU VJG HTCEVKQP QH VJG EQWPVU HQT 9 KP & VJCV CTG CEEQWPVGF HQT D[ IKXGP VJG EWTTGPV UGV QH RCTCOGVGTU KP VJG IGPGTCVKXG OQFGN KU VJG PWODGT QH VKOGU VJCV YQTF 9 CRRGCTU KP VJG FQEWOGPV CPF KU CP KPFKECVQT HWPEVKQP TGVWTPKPI KH KVU RTGFKECVG KU VTWG CPF QVJGTYKUG 6JG DKCU VGTO KU PGGFGF VQ DKCU VJG QDUGTXCVKQPU VQYCTFU VJG ). UVCVG QVJGTYKUG VJG '/ CNIQTKVJO YQWNF TGUWNV KP C \GTQ VTCPUKVKQP RTQDCDKNKV[ VQ VJG ). UVCVG =? 6JG GHHGEV QH VJG DKCU KU VJCV VJG VTCPUKVKQP CPF GOKUUKQP RTQDCDKNKVKGU HQT VQRKE YKNN DG UGV UWEJ VJCV VJKU VQRKE CEEQWPVU HQT C HTCEVKQP QH VJG YQTFU KP VJG EQTRWU TQWIJN[ GSWCN VQ 6JG GOKUUKQP RTQDCDKNKVKGU CTG VJGP GUVKOCVGF HTQO
·½
E\&5&3UHVV//&
¾
8.4.3 Classification
6Q RGTHQTO ENCUUKſECVKQP HQT C IKXGP FQEWOGPV YG PGGF VQ ſPF VJG UGV QH VQRKEU VJCV OCZKOK\GU $WV VJG VQVCN PWODGT QH CNN RQUUKDNG UGVU KU YJKEJ KU C XGT[ NCTIG PWODGT KH VJG PWODGT QH RQUUKDNG VQRKEU / KU KP VJG VJQWUCPFU 5KPEG UEQT KPI UWEJ C NCTIG PWODGT QH RQUUKDKNKVKGU KU RTQJKDKVKXG EQORWVCVKQPCNN[ YG GORNQ[ C VYQRCUU CRRTQCEJ +P VJG ſTUV RCUU YG UGNGEV C UOCNN UGV QH VQRKEU VJCV CTG NKMGN[ VQ DG KP VJG DGUV UGV +P VJG UGEQPF RCUU YG UEQTG CNN UGVU QH VJGUG ECPFKFCVGU WUKPI 9G UGNGEV ECPFKFCVG VQRKEU KP VJG ſTUV RCUU D[ UEQTKPI GCEJ VQRKE KPFGRGPFGPVN[ CU KH KV YGTG C EQORNGVG UGV QP KVU QYP WUKPI C UNKIJV OQFKſECVKQP QH
YJGTG KU KH CPF Z QVJGTYKUG CPF UGTXGU VQ ſNVGT QWV VJG GHHGEV QH YQTFU KP FQEWOGPVU VJCV EQPUVKVWVG PGICVKXG GXKFGPEG HQT C VQRKE 6JG RCTCOGVGT JCU DGGP KPVTQFWEGF VQ DCNCPEG VJG RTKQT CICKPUV VJG IGPGTCVKXG OQFGN CPF KU QRVKOK\GF HTQO VTCKPKPI FCVC 6JG RCTCOGVGT KU VJGTG VQ ƀCVVGP KH NGUU VJCP QPG QT UJCTRGP KH ITGCVGT VJCP QPG VJG VTCPUKVKQP RTQDCDKNKV[ FKUVTKDWVKQP KP QTFGT VQ EQORGPUCVG HQT VJG KPFGRGPFGPEG CUUWORVKQP QXGT YQTFU KP VJG FQEWOGPV
8.4.4 Experiments 9G CRRNKGF VJG VYQRCUU RTQEGFWTG QH VJG 1P6QRKE ENCUUKſGT FGUETKDGF CDQXG VQ C EQTRWU QH DTQCFECUV PGYU UVQTKGU VTCPUETKDGF CPF CPPQVCVGF D[ 2TKOCT[ 5QWTEG /G FKC (QT GCEJ UVQT[ VJG CPPQVCVQTU ICXG C PWODGT QH VQRKE NCDGNU VJCV VJG[ VJQWIJV TGRTGUGPVKPI VJG VQRKEU KP VJG UVQT[ 6JG PWODGT QH VQRKEU HQT GCEJ UVQT[ YCU CP[ YJGTG DGVYGGP CPF YKVJ CP CXGTCIG QH VQRKEU RGT UVQT[ 6JG EQTRWU YCU FKXKFGF KPVQ QPG [GCT QT UVQTKGU HQT VTCKPKPI CPF QPG OQPVJ QT UVQTKGU HQT VGUV 6JG VTCKPKPI UGV EQPVCKPGF C VQVCN QH WPKSWG VQRKE NCDGNU /GCUWTKPI VJG RGTHQTOCPEG QH QWT U[UVGO CICKPUV YJCV VJG JWOCP CPPQVCVQTU YTQVG FQYP CU VJG VQRKE NCDGNU KU PQV UVTCKIJVHQTYCTF DGECWUG QWT U[UVGO IKXGU CP QTFGTGF NKUV QH CNN VQRKEU GCEJ YKVJ C UEQTG YJKNG VJG CPPQVCVQTU JCXG C UOCNN WPQTFGTGF NKUV QH VQRKEU HQT GCEJ UVQT[ 9G OGCUWTG VJG RGTHQTOCPEG CU C HWPEVKQP QH VJG PWODGT 0 QH VQRTCPMKPI VQRKEU RTQXKFGF D[ VJG U[UVGO (QT GCEJ XCNWG QH 0 YG EQORCTG VJG VQR0 VQRKEU RTQFWEGF D[ VJG U[UVGO CICKPUV VJG UGV QH VQRKEU IGPGTCVGF D[ VJG CPPQVCVQTU 6JG CEEWTCE[ YCU HQT VJG ſTUV EJQKEG CPF FGETGCUGF VQ CDQWV HQT VJG ſHVJ EJQKEG 9G JCXG KPFKECVKQPU VJCV VJG ETKVGTKC YG JCXG CFQRVGF HQT OGCUWTKPI VJG RGTHQTOCPEG QH QWT U[UVGO OC[ DG NGUU HQTIKXKPI VJCP PGEGUUCT[ 6QRKE CPPQVCVKQP KU PQV CP GCU[ VCUM HQT RGQRNG YJGP VJG PWODGT QH VQRKEU KU NCTIG RGQRNG VGPF VQ WPFGTIGPGTCVG NCDGNU HQT FQEWOGPVU DGECWUG KV KU FKHſEWNV VQ TGOGODGT UQ OCP[ VQRKEU 7RQP KP HQTOCN GZCOKPCVKQP QH UVQTKGU HQT YJKEJ VJG VQRUEQTKPI VQRKE YCU PQV KPENWFGF KP VJG NKUV IKXGP D[ VJG CPPQVCVQTU YG HQWPF VJCV YGNN QXGT QH VJG VKOG VJG VQRKE IKXGP D[ VJG EQORWVGT YCU SWKVG TGCUQPCDNG HQT VJG UVQT[ +P VJGUG ECUGU VJG JWOCP
E\&5&3UHVV//&
CPPQVCVQTU JCF UKORN[ PQV DGGP GZJCWUVKXG KP VJGKT GPWOGTCVKQP QH VJG RQUUKDNG VQRKEU HQT VJG UVQT[
8.5 Information Retrieval +PHQTOCVKQP TGVTKGXCN KU VJG VCUM QH ſPFKPI FQEWOGPVU VJCV CTG TGNGXCPV VQ C SWGT[ YJGTG VJCV SWGT[ OKIJV EQPVCKP C UOCNN PWODGT QH YQTFU V[RGF D[ C RGTUQP QT OKIJV GXGP EQPUKUV QH UGXGTCN FQEWOGPVU KPFKECVGF CU KPVGTGUVKPI 6Q RGTHQTO VJKU VCUM YG FGXGNQRGF CP KPHQTOCVKQP TGVTKGXCN +4 U[UVGO ECNNGF )QNFGP 4GVTKGXGT =? )QNFGP 4GVTKGXGT KU C PQXGN RTQDCDKNKUVKE *//DCUGF +4 U[UVGO VJCV EQORWVGU VJG RTQD CDKNKV[ VJCV C FQEWOGPV KU TGNGXCPV IKXGP C SWGT[ CPF TCPMU CNN FQEWOGPVU KP VJG EQNNGEVKQP DCUGF QP VJKU OGCUWTG 1WT CRRTQCEJ VQ +4 OKTTQTU QWT VQRKE ENCUUKſECVKQP YQTM YG CNNQY C EQTRWU QH GZCORNGU VQ FTKXG QWT UGNGEVKQP QH OQFGNU CPF QWT GUVKOC VKQP RTQEGFWTGU 6JG EQTRWU EQPUKUVU QH C UGV QH FQEWOGPVU C UGV QH PCVWTCN NCPIWCIG SWGTKGU VGPU QH YQTFU CPF C PWODGT QH TGNGXCPEG LWFIOGPVU VJCV UVCVG YJGVJGT GCEJ FQEWOGPV KU TGNGXCPV VQ VJG SWGT[ QT PQV *WOCP CPPQVCVQTU OCMG VJG TGNGXCPEG LWFI OGPVU QP UQOG UKIPKſECPV UCORNKPI QH VJG EQTRWU QH FQEWOGPVU HQT GCEJ SWGT[ 9G DWKNF C UVCVKUVKECN OQFGN ECRCDNG QH TCPMKPI VTCKPKPI FQEWOGPVU GHHGEVKXGN[ D[ VJGKT NCDGNGF TGNGXCPEG VQ IKXGP VTCKPKPI SWGTKGU
8.5.1 A Bayesian Model for IR )KXGP C SWGT[ KV UGGOU UGPUKDNG VQ TCPM VJG FQEWOGPVU KP C EQTRWU D[ VJGKT RTQDCDKN KV[ QH DGKPI TGNGXCPV =? +P QVJGT YQTFU YG YCPV VQ WUG CU QWT FQEWOGPV TCPMKPI HWPEVKQP VJG RQUVGTKQT RTQDCDKNKV[ VJG RTQDCDKNKV[ VJCV VJG FQEWOGPV & KU TGNGXCPV IKXGP SWGT[ 3 9G CICKP WUG $C[GUŏ TWNG VQ FGEQORQUG VJG RQUVGTKQT RTQDCDKNKV[
KU VJG RTKQT RTQDCDKNKV[ QH C FQEWOGPV DGKPI TGNGXCPV VQ CP[ SWGT[ 9JKNG VJKU OKIJV PQV UQWPF OGCPKPIHWN KV ECP DG C XGT[ RQYGTHWN UQWTEG QH KPHQTOCVKQP (QT GZCORNG FQEWOGPVU HTQO RCTVKEWNCT UQWTEGU CTG OQTG NKMGN[ VQ DG WUGHWN 0GYGT FQEWOGPVU OC[ DG OQTG TGNGXCPV VJCP QNF QPGU NQPI FQEWOGPVU OQTG TGNGXCPV VJCP UJQTV QPGU CPF UQ QP +P CFFKVKQP VJGTG CTG NKMGN[ VQ DG RCTVKEWNCT FQEWOGPVU VJCV GXGT[QPG YCPVU 6JGUG RTGHGTGPEGU ECP DG WPKXGTUCN QT C HWPEVKQP QH VJG ITQWR VJG WUGT KU KP QT GXGP RCTVKEWNCT VQ VJG KPFKXKFWCN WUGT KU UKORN[ VJG RTKQT RTQD CDKNKV[ QH VJG SWGT[ DGKPI RQUGF KP VJG ſTUV RNCEG #U VJKU SWCPVKV[ FQGU PQV CNVGT VJG FQEWOGPV TCPMKPI YG ECP UCHGN[ KIPQTG KV 9JCV KU NGHV KU VJG EQPFKVKQPCN RTQD CDKNKV[ QH VJG SWGT[ DGKPI RQUGF WPFGT VJG J[RQVJGUKU VJCV VJG FQEWOGPV KU TGNGXCPV 9G OQFGN VJKU TGOCKPKPI SWCPVKV[ YKVJ C FKUETGVG *// VJCV KU FGRGPFGPV QP VJG FQEWOGPV 6JKU YKNN DG C IGPGTCVKXG OQFGN YJGTG YG VJKPM QH VJG
E\&5&3UHVV//&
FQEWOGPV *// CU IGPGTCVKPI VJG SWGT[ 6JG RCTCOGVGTU QH VJG *// UJQWNF DG GUVKOCVGF KP UWEJ C YC[ CU VQ OCMG KV OQTG NKMGN[ VJCV C FQEWOGPV YKNN IGPGTCVG C SWGT[ VQ YJKEJ KV KU TGNGXCPV VJCP C SWGT[ VQ YJKEJ KV KU PQV TGNGXCPV # UKORNG HQTOWNCVKQP QH VJG TGSWKUKVG *// JCU LWUV VYQ UVCVGU NCDGNGF & CPF ). YKVJ UVCVG & TGRTGUGPVKPI VJG QRVKQP QH IGPGTCVKPI SWGT[ YQTFU D[ FTCYKPI YQTFU FKTGEVN[ HTQO VJG FQEWOGPV CPF UVCVG ). TGRTGUGPVKPI EJQQUKPI YQTFU HTQO IGPGTCN NCPIWCIG KG YKVJQWV TGICTF VQ VJG FQEWOGPV /QUV SWGTKGU EQPVCKP YQTFU VJCV CTG RTGUGPV KP TGNGXCPV FQEWOGPVU DWV CNN SWGTKGU EQPVCKP OCP[ IGPGTCN YQTFU VJCV CTG PQV TGCNN[ RCTV QH VJG URGEKſECVKQP QH TGNGXCPV FQEWOGPVU
8.5.2 Training the IR HMM 6JG RCTCOGVGTU QH VJG *// CTG VJG VTCPUKVKQP RTQDCDKNKV[ « VJG RTQDCDKNKV[ VJCV VJG SWGT[ YQTF YKNN DG EJQUGP HTQO VJG ). UVCVG CPF VJG GOKUUKQP RTQDCDKNKVKGU HQT GCEJ QH VJG YQTFU KP GCEJ UVCVG +P RTKPEKRNG YG YQWNF NKMG VQ GUVKOCVG VJGUG RCTCOGVGTU HTQO GZCORNGU WUKPI VJG '/ CNIQTKVJO +P RTCEVKEG JQYGXGT YG ſPF VJCV YG FQ PQV JCXG GPQWIJ VTCKPKPI GZCORNGU VQ ſPF IQQF GUVKOCVGU HQT VJG GOKUUKQP RTQDCDKN KVKGU 5Q YG UGV VJG GOKUUKQP RTQDCDKNKVKGU HQT VJG & CPF ). UVCVGU VQ DG VJG WPKITCO FKUVTKDWVKQPU QH VJG YQTFU KP VJG FQEWOGPV CPF VJG YJQNG EQTRWU TGURGEVKXGN[ (WT VJGT YG UGV VJG VTCPUKVKQP RTQDCDKNKVKGU VQ DG VJG UCOG « HQT CNN FQEWOGPVU CPF YG GUVKOCVG « WUKPI VJG '/ CNIQTKVJO 9G HQWPF VJCV VJG XCNWG QH « FGRGPFU QP VJG V[RG QH SWGT[ +H VJG SWGT[ KU C UJQTV RJTCUG V[RGF D[ C JWOCP VJGP C V[RKECN XCNWG HQT « KU OGCPKPI VJCV QH VJG YQTFU KP VJG SWGT[ YQWNF DG GZRGEVGF VQ DG HQWPF KP CP[ TGNGXCPV FQEWOGPV $WV KH VJG SWGT[ YGTG C NQPI FGUETKRVKQP VJGP VJG V[RKECN XCNWG QH « YCU OGCPKPI VJCV QPN[ QH VJG YQTFU YQWNF DG GZRGEVGF VQ CRRGCT KP TGNGXCPV FQEWOGPVU (KPCNN[ KH VJG SWGT[ YGTG CP GPVKTG FQEWOGPV GXGP JKIJGT XCNWGU QH « YQWNF DG CRRTQRTKCVG
8.5.3 Performance 9G JCXG VGUVGF VJKU UKORNG VYQUVCVG *// QP VJG 64'% 6GZV 4GVTKGXCN %QPHGT GPEG EQTRWU YJKEJ EQPUKUVU QH FQEWOGPVU =? 9G RTGRTQEGUU VJG EQTRWU NKIJVN[ KP QTFGT VQ URNKV FQEWOGPVU WR KPVQ YQTFU CPF VQ CNNQY OQTRJQNQIKECNN[ UKO KNCT YQTFU VQ OCVEJ 6JG UVTGCO QH EJCTCEVGTU KP GCEJ FQEWOGPV IGVU VQMGPK\GF KPVQ YQTFU 6JGP YG EQPƀCVG VGTOU D[ CRRN[KPI 2QTVGTŏU UVGOOKPI CNIQTKVJO =? 0GZV YG FKUECTF CP[VJKPI HQWPF KP C NKUV QH ŒUVQRŒ YQTFU (KPCNN[ PWOGTKE CPF PQP YQTF KVGOU CTG TGFWEGF VQ UKPING VQMGPU ŏ07/$'4ŏ ŏ&1..#4ŏ GVE 6JG VGUV EQORTKUGF SWGTKGU YKVJ CP CXGTCIG QH YQTFU RGT SWGT[ 'CEJ SWGT[ YCU RTGRTQEGUUGF KP VJG UCOG OCPPGT FGUETKDGF CDQXG HQT VJG FQEWOGPVU 6JGP HQT GCEJ SWGT[ YG EQORWVG HQT GCEJ QH VJG FQEWOGPVU 6JG TGUWNV KU VJCV VJG VQRUEQTKPI FQEWOGPV HQT GCEJ SWGT[ YCU HQWPF VQ DG TGNGXCPV QH VJG VKOG 6JG UKORNG OQFGN FGUETKDGF CDQXG JCU DGGP GZVGPFGF D[ CFFKPI OQTG UVCVGU YKVJ FKHHGTGPV SWGT[ VGTOIGPGTCVKPI OGEJCPKUOU GI U[PQP[OU DKITCOU VQRKEU WPUW RGTXKUGF TGNGXCPEG HGGFDCEM CPF D[ VJG KPENWUKQP QH FQEWOGPV RTKQTU TGUWNVKPI KP JKIJGT RGTHQTOCPEG =?
E\&5&3UHVV//&
8.6 Event Tracking #PQVJGT CRRNKECVKQP QH VQRKE ENCUUKſECVKQP KU ECNNGF GXGPV VTCEMKPI +P VJKU ECUG YG CTG IKXGP C UOCNN PWODGT QH FQEWOGPVU VJCV FGUETKDG CP GXGPV CPF YG CTG CUMGF VQ ſPF CNN QH VJG QVJGT FQEWOGPVU VJCV FGUETKDG VJG UCOG GXGPV 6JG 6QRKE &GVGEVKQP CPF 6TCEMKPI 2TQLGEV URQPUQTGF D[ VJG IQXGTPOGPV RQUGF LWUV UWEJ C RTQDNGO +P VJKU RTQDNGO YG CTG IKXGP HQWT FQEWOGPVU VJCV CTG RWTRQTVGF VQ DG CDQWV C RCTVKEWNCT GXGPV 1WT VCUM KU VQ ſPF VJG TGOCKPKPI FQEWOGPVU CDQWV VJG UCOG GXGPV $WV VJGTG CTG UQOG EQORNKECVKQPU KP VJG VCUM (KTUV YG OWUV OCMG C DKPCT[ FGEKUKQP HQT GCEJ FQEWOGPV CU UQQP CU YG NQQM CV KV 9G ECPPQV NQQM CV VJG YJQNG EQTRWU KP QTFGT VQ TCPM VJG FQEWOGPVU $WV YG CTG CNUQ IKXGP C RCTV QH VJG EQTRWU CPF VQNF VJCV OQUV QH VJGUG FQEWOGPVU FQ PQV FKUEWUU VJG UCOG GXGPV 6JWU YG OWUV FGXKUG C OGEJCPKUO HQT FGVGTOKPKPI C VJTGUJQNF QP VJG UEQTG KP QTFGT VQ MPQY YJGVJGT QT PQV VQ CEEGRV VJG FQEWOGPV # UGEQPF RCTV QH VJG RTQDNGO KU VJCV VJG PCVWTG QH VJG FQEWOGPVU KU NKMGN[ VQ EJCPIG QXGT VKOG CU VJG FKUEWUUKQP QH VJG QTKIKPCN GXGPV GXQNXGU (QT GZCORNG VJG KPKVKCN UVQTKGU OKIJV DG CDQWV VJG EQOOKUUKQP QH UQOG ETKOG 6JG NCVGT UVQTKGU OKIJV DG CDQWV VJG VTKCN QH VJG RGTUQP U CEEWUGF QH VJCV ETKOG 1PG ECP VJKPM QH VJKU RTQDNGO KP VYQ YC[U (KTUV YG ECP VJKPM QH VJG GZCORNG FQEWOGPVU CU FGſPKPI C VQRKE 9G ECP GUVKOCVG C VQRKE OQFGN KG YQTF FKUVTKDWVKQP HTQO VJGUG GZCORNG FQEWOGPVU 6JGP QWT VCUM KU VQ ſPF VJQUG FQEWOGPVU KP VJG EQTRWU VJCV JCXG C JKIJ RTQDCDKNKV[ QH FKUEWUUKPI VJCV UCOG VQRKE #NVGTPCVKXGN[ YG ECP VJKPM QH VJG GZCORNG FQEWOGPVU CU C XGT[ NQPI SWGT[ #PF YG YQWNF VJGP DG NQQMKPI HQT FQEWOGPVU VJCV CTG TGNGXCPV VQ WU IKXGP VJCV VJG SWGT[ YCU CU IKXGP 0QVG VJCV VJGUG VYQ HQTOWNCVKQPU TGUWNV KP GPVKTGN[ FKHHGTGPV OQFGNU CPF RTQEGUUGU +P VJG ſTUV ECUG YG EQORWVG C VQRKE OQFGN HTQO VJG GZCORNG FQEWOGPVU CPF VJGP WUKPI $C[GUŏ TWNG YG EQORWVG VJG RTQDCDKNKV[ QH VJG YQTFU KP GCEJ FQEWOGPV IKXGP VJCV VJG FQEWOGPV KU CDQWV VJKU VQRKE +P VJG UGEQPF ECUG YG EQORWVG C FKUVTKDWVKQP HTQO GCEJ FQEWOGPV CPF VJGP EQORWVG VJG RTQDCDKNKV[ QH VJG IKXGP ŏSWGT[ŏ IKXGP VJKU FQEWOGPV KU TGNGXCPV 6JG VQRKE ENCUUKſECVKQP OGEJCPKUO JCU UQOG CFXCPVCIGU (QT GZCORNG CU YG IQ VJTQWIJ UVQTKGU YG ſPF UGXGTCN UVQTKGU VJCV JCXG C XGT[ JKIJ NKMGNKJQQF QH DGKPI CDQWV VJG UCOG GXGPV 9G ECP KH UWHſEKGPVN[ EQPſFGPV CFF VJGUG FQEWOGPVU KPVQ VJG OQFGN CU KH VJG[ YGTG VTCKPKPI GZCORNGU HQT VJG GXGPV 6JGP YG TGVTCKP QWT OQFGN HQT VJG VQRKE CPF RTQEGGF HTQO VJGTG 6JG KPHQTOCVKQP TGVTKGXCN OQFGN ECPPQV FQ VJKU CU GCUKN[ 6JCV KU DGECWUG YG CTG KOCIKPKPI VJCV CNN QH VJG VTCKPKPI FQEWOGPVU YGTG ŏIGPGTCVGFŏ D[ GCEJ TGNGXCPV FQEWOGPV $WV CV VJG RQKPV YJGTG VJG PWODGT QH VTCKPKPI FQEWOGPVU DGEQOGU XGT[ OWEJ NCTIGT VJCP GCEJ NCVGT FQEWOGPV YG ECPPQV TGCNN[ GZRGEV VJG UKPING FQEWOGPV VQ EQPVCKP OQUV QH VJG YQTFU KP VJG VTCKPKPI FQEWOGPVU 6JG RTQDNGO QH FGEKFKPI YJGVJGT C FQEWOGPV KU TGNGXCPV VQ VJG HQWT GZCORNG FQE WOGPVU CU QRRQUGF VQ LWUV TCPMKPI VJG FQEWOGPVU KP VJG EQTRWU KU C FKHſEWNV QPG 6JKU KU DGECWUG VJG UEQTGU HQT GXGPV OQFGN CPF FQEWOGPV YKNN XCT[ CU C HWPEVKQP QH WPKORQTVCPV HCEVQTU (QT GZCORNG HQT VJG VQRKE ENCUUKſECVKQP OQFGN VJG PWODGT YQTF UEQTGU HQT C VGUV FQEWOGPV KU C HWPEVKQP QH VJG NGPIVJ QH VJG FQEWOGPV 9JKNG
E\&5&3UHVV//&
YG ECP EQORWVG VJG CXGTCIG UEQTG RGT YQTF VJG UEQTG UVKNN FGRGPFU QP VJG RCTVKEWNCT YQTFU VJCV CRRGCT KP VJG FQEWOGPV 1P VJG QVJGT JCPF VJG +PHQTOCVKQP 4GVTKGXCN UEQTG YKNN DG TGNCVKXGN[ WPDKCUGF HQT VJG FKHHGTGPV FQEWOGPVU DWV VJG UEQTG YKNN XCT[ CU C HWPEVKQP QH VJG NGPIVJ QH VJG SWGT[ VJG HQWT GZCORNG FQEWOGPVU KP VJKU ECUG CPF VJG RCTVKEWNCT YQTFU WUGF KP VJQUG FQEWOGPVU 6Q UQNXG VJKU RTQDNGO YG CFFGF CPQVJGT NGXGN QH ENCUUKſGT )KXGP C OQFGN HQT C VQRKE YG ECP EQORWVG VJG UEQTG HQT C NCTIG PWODGT QH FQEWOGPVU KP QWT EQTRWU 9G TGOQXG VJQUG FQEWOGPVU VJCV YG GKVJGT MPQY VQ DG QP VJG UCOG GXGPV QT TGEGKXG C XGT[ JKIJ UEQTG VJG[ OC[ DG CDQWV VJG UCOG GXGPV 6JGP YG OCMG C FKUVTKDWVKQP QH VJQUG UEQTGU 9G ECP VJGP UGV C VJTGUJQNF DCUGF QP VJCV FKUVTKDWVKQP (QT GZCORNG YG V[RKECNN[ UGV C VJTGUJQNF CV UKZ UVCPFCTF FGXKCVKQPU CDQXG VJG OGCP UEQTG XCNWG
6JG FKUVTKDWVKQP QH UEQTGU KU PQV TGCNN[ )CWUUKCP 9G LWUV WUG VJKU OGEJCPKUO CU C EQPXGPKGPV QPG 6JG TGUWNV KU VJCV UQOG ſZGF HTCEVKQP QH QHHVQRKE UVQTKGU YQWNF DG KPEQTTGEVN[ ENCUUKſGF CU DGKPI QP VJG GXGPV 6JCV KU YG ſZ VJG RTQDCDKNKV[ QH HCNUG CEEGRVCPEG QH C UVQT[ VJCV KU PQV CDQWV VJG FGUKTGF GXGPV 6JKU TGUWNVU KP GZVTGOGN[ EQPUKUVGPV DGJCXKQT QH VJG ENCUUKſGT CETQUU FKHHGTGPV GXGPVU CPF FKHHGTGPV FQEWOGPVU 6JG TGUWNVKPI RTQDCDKNKV[ QH OKUUKPI C FQEWOGPV VJCV KU QP VJG GXGPV KU YJCVGXGT EQOGU HTQO VJG PCVWTG QH VJG RTQDNGO $WV VJG PWODGT QH HCNUG CNCTOU KU EQORNGVGN[ RTGFKEVCDNG
8.7 Unsupervised Topic Detection +P QTFGT VQ WUG VQRKE ENCUUKſECVKQP YG RTGUWOG VJCV YG JCXG C NCTIG UGV QH FQEW OGPVU CPPQVCVGF YKVJ VQRKEU $WV HQT C NCTIG PWODGT QH VQRKEU YG PGGF CP GXGP NCTIGT EQTRWU 6JG EQUV QH CPPQVCVKPI C NCTIG EQTRWU YKVJ VQRKEU ECP DG SWKVG NCTIG (WTVJGTOQTG VJG VQRKEU EQPVKPWG VQ EJCPIG QXGT VKOG TGSWKTKPI EQPUVCPV CPPQVCVKQP (KPCNN[ YG OWUV FQ VJKU HQT GXGT[ FQOCKP CPF GXGT[ NCPIWCIG *QYGXGT KV KU RQUUKDNG VQ CPPQVCVG C EQTRWU CWVQOCVKECNN[ YKVJ VQRKEU UKOKNCT VQ VJQUG VJCV YQWNF DG ETGCVGF D[ JWOCP CPPQVCVQTU 6Q FQ UQ YG WUG VJG PCVWTG QH VQRKEU CPF YG CNUQ WUG VJG VQQNU VJCV YG ETGCVGF HQT VJG UWRGTXKUGF VQRKE ENCUUKſECVKQP RTQDNGO (KTUV YG ECP CDUVTCEV VJG FGſPKVKQP QH VQRKEU VQ DG C EQNNGEVKQP QH UGVU QH YQTFU 6JG YQTFU YKVJKP GCEJ UGV CTG UVCVKUVKECNN[ TGNCVGF 9G CUUWOG FQEWOGPVU CTG YTKVVGP D[ FTCYKPI HTQO VJGUG UGVU QH YQTFU UQOGYJCV CV TCPFQO 5GEQPF YG PGGF VQ JCXG WPFGTUVCPFCDNG PCOGU HQT GCEJ QH VJGUG VQRKEU QVJGTYKUG KV KU PQV ENGCT YJCV C RGTUQP YQWNF WUG VJGO HQT 9G JCXG FGXKUGF C OGVJQF HQT FGTKXKPI C NCTIG UGV QH PCOGF VQRKEU HTQO C EQTRWU QH FQEWOGPVU 6JG OGVJQF YQTMU KP CP[ NCPIWCIG QT FQOCKP 9G JCXG GXCNWCVGF VJG VQRKEU CPF HQWPF VJCV VJG[ CTG EQORCTCDNG KP VJGKT DGJCXKQT VQ VJG VQRKEU CUUKIPGF D[ JWOCP CPPQVCVQTU CPF HQT UQOG CRRNKECVKQPU GXGP TGUWNV KP UWRGTKQT RGTHQTOCPEG VQ JWOCPETGCVGF VQRKE UGVU $GHQTG YG UVCTV YG FGTKXG XCTKQWU RJTCUGU HTQO VJG EQTRWU UKPEG RJTCUGU QHVGP ECTT[
E\&5&3UHVV//&
OQTG QDXKQWU OGCPKPI VJCP UKPING YQTFU 9G WUG VYQ VGEJPKSWGU VQ ſPF RJTCUGU KP VJG EQTRWU 6JG ſTUV QPG KU UKORN[ VQ ſPF UVTKPIU QH YQTFU VJCV CTG ŒUVKEM[Œ 6JCV KU VJG YQTFU VGPF VQ CRRGCT VQIGVJGT OWEJ OQTG QHVGP VJCP QVJGTYKUG GZRGEVGF 6JGTG CTG UGXGTCN VGEJPKSWGU VJCV EQWNF DG WUGF HQT VJKU (QT GZCORNG QPG ECP WUG VJG OWVWCN KPHQTOCVKQP DGVYGGP YQTFU 9G JCXG WUGF C OKPKOWO FGUETKRVKQP NGPIVJ VGEJPKSWG VJCV ſPFU VJQUG UVTKPIU YJKEJ KH CFQRVGF OQUV TGFWEG VJG PWODGT QH DKVU PGGFGF VQ GPEQFG VJG GPVKTG EQTRWU 6JG UGEQPF VGEJPKSWG HQT ſPFKPI RJTCUGU WUGU VJG PCOGFGPVKV[ GZVTCEVKQP VGEJPKSWG FGUETKDGF KP VJG RTGXKQWU UGEVKQP 6JKU CUUWOGU VJCV UQOG VGZV KP VJG NCPIWCIG JCU DGGP CPPQVCVGF YKVJ GZCORNGU QH PCOGU QH FKHHGTGPV V[RGU *GTG PCOGU ECP DG IGPGTCNK\GF VQ KPENWFG CP[ V[RGU QH KPVGTGUVKPI RJTCUGU (QT GZCORNG HQT PGYU YG HQWPF VJQUG RJTCUGU VJCV YGTG PCOGU QH RGQRNG NQECVKQPU QT QTICPK\CVKQPU 6JGUG RJTCUGU CTG WUGF UKORN[ VQ FGſPG CFFKVKQPCN VGTOU 6JG QTKIKPCN YQTFU CTG MGRV DWV CFFKVKQPCN UVTKPIU CTG CFFGF HQT VJG RJTCUGU 0GZV VJG CNIQTKVJO EJQQUGU ECPFKFCVG VQRKE PCOGU 6JGUG CTG EJQUGP UKORN[ CU YQTFU QT RJTCUGU VJCV UGGO VQ JCXG JKIJ KPHQTOCVKQP EQPVGPV 5RGEKſECNN[ HQT GCEJ VGTO YQTF QT RJTCUG KP VJG FQEWOGPV YG EQORWVG KVU KPXGTUG FQEWOGPV HTGSWGPE[ CU
YJGTG 0 KU VJG VQVCN PWODGT QH FQEWOGPVU KP VJG EQTRWU CPF FH Y KU VJG PWODGT QH FQEWOGPVU KP VJG EQTRWU VJCV EQPVCKP VJG YQTF Y CV NGCUV QPEG 6JGP YKVJKP GCEJ FQEWOGPV YG EQORWVG VJG PWODGT QH VKOGU GCEJ YQTF QEEWTU OWNVKRNKGF D[ VJG +&( QH VJCV VGTO 6JG TGUWNV KU CP CF JQE OGCUWTG QH JQY NKMGN[ VJCV VGTO KU VQ DG CP KORQTVCPV EQPEGRV HQT VJG FQEWOGPV 9G EJQQUG VJG JKIJGUV ſXG UWEJ VGTOU HQT GCEJ FQEWOGPV UWDLGEV VQ VJG EQPFKVKQP VJCV GCEJ QH VJGUG VGTOU QEEWTU UQOG OKPKOWO PWODGT QH VKOGU KP VJG GPVKTG EQTRWU
9G JCXG WUGF CU C VJTGUJQNF 6JKU VJGP OGCPU VJCV VJKU VGTO KU C RNCWUKDNG PCOG QH C VQRKE CPF YG CNUQ JCXG C UGV QH FQEWOGPVU VJCV UGGO VQ EQPVCKP VJG VGTO 6JGUG CTG VCMGP VQ DG VQ DG KPKVKCN VQRKEU CPF VQRKE CUUKIPOGPVU +H YG EJCPIG VJGUG VJTGUJQNFU VJGP VJG PWODGT QH VQRKEU RTQFWEGF D[ VJG U[UVGO YKNN XCT[ $WV QWT IQCN KU VQ ſPF HWNN VQRKEU YJGTG VQRKEU CTG RTQDCDKNKV[ FKUVTKDWVKQPU QH NCTIG UGVU QH YQTFU CPF VQ CUUKIP VJGUG VQRKEU VQ CNN VJG FQEWOGPVU VJCV FKUEWUU VJGO - PQV LWUV VJQUG VJCV EQPVCKP VJG PCOG QH VJG VQRKE 5Q YG WUG VJG UCOG 1P6QRKE VTCKPKPI CNIQTKVJO FGUETKDGF KP CDQXG 6JCV KU YG VCMG VJG KPKVKCN VQRKE CUUKIPOGPVU VQ DG EQTTGEV 9G GUVKOCVG PGY OQFGNU HQT VJG VQRKEU 6JKU JCU VJG GHHGEV QH CFFKPI VJG QVJGT EQPVGPV YQTFU HTQO VJG FQEWOGPVU VQ VJG VQRKE OQFGNU FGſPGF 6JGP YG WUG VJGUG VQRKE OQFGNU VQ ENCUUKH[ CNN QH VJG FQEWOGPVU 6JG TGUWNV KU VJCV VJG VQRKEU CTG PQY CUUKIPGF VQ OCP[ OQTG FQEWOGPVU OCP[ QH YJKEJ FQ PQV EQPVCKP VJG PCOGU QH VJG VQRKEU 9G JCXG WUGF VJKU CNIQTKVJO QP VJG UCOG EQTRWU HTQO 2TKOCT[ 5QWTEG /GFKC FG UETKDGF CDQXG +P VJKU ECUG YG ETGCVGF CDQWV VQRKEU CWVQOCVKECNN[ 6JG VQRKEU ETGCVGF YGTG LWFIGF D[ RTGUGPVKPI VJG EJQUGP VQRKEU HQT UGXGTCN FQEWOGPVU VQ UWD LGEVU YJQ FGEKFGF YJGVJGT GCEJ QPG YCU CRRTQRTKCVG 9G HQWPF VJCV QP QXGT QH VJG FQEWOGPVU VJG ſTUV EJQKEG FQEWOGPV UGGOGF TGNGXCPV VQ VJG FQEWOGPV (WT VJGTOQTG QP VJG ſHVJ EJQKEG VQRKE VJG TGNGXCPEG YCU UVKNN QXGT 6JKU TGUWNV YCU
E\&5&3UHVV//&
SWKVG JKIJ KP VJCV KV YCU EQORCTCDNG VQ QT GXGP JKIJGT VJCP VJCV QH VJG VQRKE OQFGNU VTCKPGF HTQO JWOCP CPPQVCVKQP 6JG ITGCV CFXCPVCIG QH WPUWRGTXKUGF VQRKE FKUEQXGT[ KU VJCV KV ECP GCUKN[ DG CRRNKGF VQ CP[ PGY FQOCKP QT NCPIWCIG +V ECP CNUQ ſPF XGT[ NCTIG PWODGTU QH VQRKEU +P HCEV YG JCXG CRRNKGF KV VQ #TCDKE PGYU UVQTKGU CPF HQWPF VJCV VJG RGTHQTOCPEG YCU EQORCTCDNG VQ VJCV QP 'PINKUJ UVQTKGU 6JKU OCFG KV RQUUKDNG VQ RGTHQTO VQRKE ENCUUK ſECVKQP QP VJG #TCDKE PGYU UVQTKGU UKPEG KV YQWNF JCXG DGGP HCT VQQ GZRGPUKXG HQT WU VQ CPPQVCVG VGPU QH VJQWUCPFU QH PGYU UVQTKGU YKVJ VQRKEU OCPWCNN[
8.8 Summary 9G JCXG FGUETKDGF UGXGTCN FKHHGTGPV VGZV RTQEGUUKPI VGEJPKSWGU VJCV ECP DG RGTHQTOGF WUKPI *//U +P OQUV ECUGU VJG TGUWNVKPI CNIQTKVJOU YGTG CU IQQF CU QT DGVVGT VJCP VJG GZKUVKPI VGEJPKSWGU HQT RGTHQTOKPI VJCV HWPEVKQP (WTVJGTOQTG VJG VGEJPKSWGU EQWNF DG WUGF VTKXKCNN[ KP CP[ NCPIWCIG QT FQOCKP (KPCNN[ YG JCXG UGGP VJCV VJG VGEJPKSWGU ECP QHVGP DG GZVGPFGF VQ TGUWNV KP JKIJGT CEEWTCE[ QT VQ RGTHQTO TGNCVGF HWPEVKQPU TGNCVKXGN[ GCUKN[ DGECWUG QH VJGKT WPFGTN[KPI UKORNKEKV[
References =? &/ $KMGN 4 5EJYCTV\ CPF 4/ 9GKUEJGFGN Œ#P #NIQTKVJO VJCV .GCTPU 9JCVŏU KP C 0COGŒ /CEJKPG .GCTPKPI -NWYGT #ECFGOKE 2WDNKUJGTU 8QN RR =? )& (QTPG[ Œ6JG 8KVGTDK #NIQTKVJOŒ 2TQE +''' 8QN RR =? , /CMJQWN ( -WDCNC 4 5EJYCTV\ CPF 4 9GKUEJGFGN Œ2GTHQTOCPEG /GC UWTGU HQT +PHQTOCVKQP 'ZVTCEVKQPŒ *# 9QTMUJQR QP $TQCFECUV 0GYU 7P FGTUVCPFKPI *GTPFQP 8# /QTICP -CWHOCPP 2WDNKUJGTU RR /CTEJ =? 6JG FCVC WUGF KP VJG XCTKQWU GZRGTKOGPVU TGRQTVGF KP VJKU RCRGT KU CXCKN CDNG HTQO VJG .KPIWKUVKE &CVC %QPUQTVKWO .&% 7PKXGTUKV[ QH 2GPPU[NXCPKC JVVRYYYNFEWRGPPGFW =? & /KNNGT 4 5EJYCTV\ 4 9GKUEJGFGN CPF 4 5VQPG Œ0COGF 'PVKV[ 'ZVTCE VKQP HTQO $TQCFECUV 0GYUŒ *# 9QTMUJQR QP $TQCFECUV 0GYU 7PFGT UVCPFKPI *GTPFQP 8# /QTICP -CWHOCPP 2WDNKUJGTU RR /CTEJ =? ; ;CPI CPF %) %JWVG Œ#P 'ZCORNG$CUGF /CRRKPI /GVJQF HQT 6GZV %CV
E\&5&3UHVV//&
GIQTK\CVKQP CPF 4GVTKGXCNŒ #%/ 6TCPU +PHQTOCVKQP 5[UVGOU 8QN 0Q RR ,WN[ =? 4 5EJYCTV\ 6 +OCK ( -WDCNC . 0IW[GP CPF , /CMJQWN Œ# /CZKOWO .KMGNKJQQF /QFGN HQT 6QRKE %NCUUKſECVKQP QH $TQCFECUV 0GYUŒ 'WTQURGGEJ ŏ 4JQFGU )TGGEG RR 5GRV =? #2 &GORUVGT 0/ .CKTF CPF &$ 4WDKP Œ/CZKOWO .KMGNKJQQF HTQO +P EQORNGVG &CVC XKC VJG '/ #NIQTKVJOŒ , 4Q[CN 5VCVKUVKECN 5QEKGV[ $ 8QN 0Q RR =? & /KNNGT 6 .GGM CPF 4 5EJYCTV\ Œ$$0 CV 64'% 7UKPI *KFFGP /CTMQX /QFGNU HQT +PHQTOCVKQP 4GVTKGXCNŒ 6JG 5GXGPVJ 6GZV 4GVTKGXCN %QPHGTGPEG
64'% 0+56 5RGEKCN 2WDNKECVKQP RR ,WN[ =? /' /CTQP CPF -. -WJPU Œ1P TGNGXCPEG RTQDCDKNKUVKE KPFGZKPI CPF KPHQT OCVKQP TGVTKGXCNŒ , #UUQE %QORWVKPI /CEJKPGT[ 8QN RR =? /( 2QTVGT Œ#P #NIQTKVJO HQT 5WHſZ 5VTKRRKPIŒ 2TQITCO 8QN 0Q RR =? &4* /KNNGT 6 .GGM CPF 4/ 5EJYCTV\ Œ# *KFFGP /CTMQX /QFGN +PHQT OCVKQP 4GVTKGXCN 5[UVGOŒ PF +PV #%/ 5+)+4 %QPH 4GUGCTEJ CPF &GXGN QROGPV KP +PHQTOCVKQP 4GVTKGXCN $GTMGNG[ %# RR #WI
E\&5&3UHVV//&
9 Statistical Language Models With Embedded Latent Semantic Knowledge Jerome R. Bellegarda Apple Computer, Inc.
CONTENTS
+PVTQFWEVKQP .CVGPV 5GOCPVKE #PCN[UKU .5# (GCVWTG 5RCEG 5GOCPVKE %NCUUKſECVKQP 0ITCO .5# .CPIWCIG /QFGNKPI 5OQQVJKPI 'ZRGTKOGPVU +PJGTGPV 6TCFG1HHU %QPENWUKQP 4GHGTGPEGU
9.1 Introduction 6JG $C[GUKCP CRRTQCEJ RGTXCUKXG KP VQFC[ŏU URGGEJ TGEQIPKVKQP U[UVGOU GPVCKNU VJG EQPUVTWEVKQP QH C RTKQT OQFGN QH VJG NCPIWCIG CU RGTVCKPU VQ VJG FQOCKP QH KPVGTGUV 6JG TQNG QH VJKU RTKQT KP GUUGPEG KU VQ SWCPVKH[ YJKEJ YQTF UGSWGPEGU CTG CEEGRVCDNG KP C IKXGP NCPIWCIG HQT C IKXGP VCUM CPF YJKEJ CTG PQV +V OWUV VJGTGHQTG GPECRUWNCVG CU OWEJ CU RQUUKDNG QH VJG U[PVCEVKE UGOCPVKE CPF RTCIOCVKE EJCTCEVGTKUVKEU QH VJG FQOCKP = ? +P VJG RCUV VYQ FGECFGU KV JCU DGEQOG KPETGCUKPIN[ EQOOQP VQ FQ UQ VJTQWIJ UVCVKUVKECN ITCO NCPIWCIG OQFGNKPI ./ YJGTG GCEJ YQTF KU RTG FKEVGF EQPFKVKQPGF QP VJG EWTTGPV EQPVGZV QP C NGHV VQ TKIJV DCUKU = ? #NVJQWIJ YKFGURTGCF VJKU UQNWVKQP KU PQV YKVJQWV FTCYDCEMU RTQOKPGPV COQPI VJG EJCNNGPIGU HCEGF D[ ITCO OQFGNKPI KU VJG KPJGTGPV NQECNKV[ QH KVU UEQRG FWG VQ VJG NKOKVGF COQWPV QH EQPVGZV CXCKNCDNG HQT RTGFKEVKPI GCEJ YQTF
9.1.1 Scope Locality %GPVTCN VQ VJKU KUUWG KU VJG EJQKEG QH YJKEJ JCU KORNKECVKQPU KP VGTOU QH RTGFKEVKXG RQYGT CPF RCTCOGVGT TGNKCDKNKV[ .CTIG XCNWGU QH CTG FGUKTCDNG HQT VJG HQTOGT DWV NQY XCNWGU QH CTG PGEGUUCT[ HQT VJG NCVVGT UGG HQT GZCORNG = ? 6JKU KP
E\&5&3UHVV//&
VWTP KORQUGU CP CTVKſEKCNN[ NQECN JQTK\QP VQ VJG OQFGN KORGFKPI KVU CDKNKV[ VQ ECRVWTG NCTIGURCP TGNCVKQPUJKRU KP VJG NCPIWCIG 6Q KNNWUVTCVG EQPUKFGT KP GCEJ QH VJG VYQ GSWKXCNGPV RJTCUGU stocks fell sharply as a result of the announcement stocks, as a result of the announcement, sharply fell
VJG RTQDNGO QH RTGFKEVKPI VJG YQTF őfellŒ HTQO VJG YQTF őstocksŒ +P VJKU ECP DG FQPG YKVJ VJG JGNR QH C DKITCO ./ YJKEJ KU UVTCKIJVHQTYCTF YKVJ VJG MKPF QH TGUQWTEGU EWTTGPVN[ CXCKNCDNG =? +P JQYGXGT VJG XCNWG YQWNF DG PGEGUUCT[ C TCVJGT WPTGCNKUVKE RTQRQUKVKQP CV VJG RTGUGPV VKOG .CTIGN[ DGECWUG QH VJKU KPCDKNKV[ VQ TGNKCDN[ ECRVWTG NCTIGURCP DGJCXKQT ITCO RGTHQTOCPEG JCU GUUGPVKCNN[ TGCEJGF C RNCVGCW =? 6JKU QDUGTXCVKQP JCU URCTMGF KPVGTGUV KP C XCTKGV[ QH EQWPVGTOGCUWTGU KPXQNXKPI HQT KPUVCPEG information aggregation QT span extension =? +PHQTOCVKQP CIITGICVKQP KP ETGCUGU VJG TGNKCDKNKV[ QH C YQTF RTGFKEVKQP D[ VCMKPI CFXCPVCIG QH GZGORNCTU QH QVJGT YQTFU VJCV DGJCXG őNKMGŒ VJKU YQTF KP VJG RCTVKEWNCT EQPVGZV EQPUKFGTGF 6JG VTCFG QHH V[RKECNN[ KU JKIJGT TQDWUVPGUU CV VJG GZRGPUG QH C NQUU KP TGUQNWVKQP 6JKU EJCRVGT KU OQTG ENQUGN[ CNKIPGF YKVJ URCP GZVGPUKQP YJKEJ GZVGPFU CPFQT EQORNGOGPVU VJG ITCO RCTCFKIO YKVJ KPHQTOCVKQP GZVTCEVGF HTQO NCTIGURCP WPKVU KG EQORTKUKPI C NCTIG PWODGT QH YQTFU 6JG VTCFGQHH JGTG KU KP VJG EJQKEG QH WPKVU EQPUKFGTGF HQT VJG CPCN[UKU QH NQPI FKUVCPEG FGRGPFGPEKGU 6JGUG WPKVU VGPF VQ DG GKVJGT U[PVCEVKE QT UGOCPVKE KP PCVWTG
9.1.2 Syntactically–Driven Span Extension #UUWOKPI C UWKVCDNG RCTUGT KU CXCKNCDNG HQT VJG FQOCKP EQPUKFGTGF U[PVCEVKE KPHQT OCVKQP ECP DG WUGF VQ KPEQTRQTCVG NCTIGURCP EQPUVTCKPVU KPVQ VJG TGEQIPKVKQP *QY VJGUG EQPUVTCKPVU CTG KPEQTRQTCVGF XCTKGU HTQO GUVKOCVKPI ITCO RTQDCDKNKVKGU HTQO ITCOOCTIGPGTCVGF FCVC =? VQ EQORWVKPI C NKPGCT KPVGTRQNCVKQP QH VJG VYQ OQFGNU =? /QUV TGEGPVN[ U[PVCEVKE KPHQTOCVKQP JCU DGGP WUGF URGEKſECNN[ VQ FGVGTOKPG GSWKXCNGPEG ENCUUGU QP VJG ITCO JKUVQT[ TGUWNVKPI KP UQECNNGF FGRGPFGPE[ = ? QT UVTWEVWTGF = ? ./U +P VJCV HTCOGYQTM GCEJ WPKV KU VJG JGCFYQTF QH VJG RJTCUG URCPPGF D[ VJG CUUQEKCVGF RCTUG UWDVTGG 6JG UVCPFCTF ITCO ./ KU VJGP OQFKſGF VQ QRGTCVG IKXGP VJG NCUV headwords CU QRRQUGF VQ VJG NCUV words #U C TGUWNV VJG UVTWEVWTG QH VJG OQFGN KU PQ NQPIGT RTGFGVGTOKPGF YJKEJ YQTFU UGTXG CU RTGFKEVQTU FGRGPFU QP VJG FGRGPFGPE[ ITCRJ YJKEJ KU C JKFFGP XCTK CDNG =? +P VJG GZCORNG CDQXG VJG VQR VYQ JGCFYQTFU KP VJG FGRGPFGPE[ ITCRJ YQWNF DG őstocksŒ CPF őfellŒ KP DQVJ ECUGU VJGTGD[ UQNXKPI VJG RTQDNGO 6JG OCKP ECXGCV KP UWEJ OQFGNKPI KU VJG TGNKCPEG QP VJG RCTUGT CPF RCTVKEWNCTN[ VJG KORNKEKV CUUWORVKQP VJCV VJG EQTTGEV RCTUG YKNN KP HCEV DG CUUKIPGF C JKIJ RTQDCDKNKV[ =? 6JG DCUKE HTCOGYQTM YCU TGEGPVN[ GZVGPFGF VQ QRGTCVG GHſEKGPVN[ KP C NGHVVQ TKIJV OCPPGT = ? VJTQWIJ ECTGHWN QRVKOK\CVKQP QH DQVJ EJCTV RCTUKPI =? CPF UGCTEJ OQFWNGU #NUQ PQVGYQTVJ[ KU C UQOGYJCV EQORNGOGPVCT[ NKPG QH TGUGCTEJ =? YJKEJ GZRNQKVU VJG U[PVCEVKE UVTWEVWTG EQPVCKPGF KP VJG UGPVGPEGU RTKQT VQ VJG QPG HGCVWTKPI VJG YQTF DGKPI RTGFKEVGF
E\&5&3UHVV//&
9.1.3 Semantically–Driven Span Extension *KIJ NGXGN UGOCPVKE KPHQTOCVKQP ECP CNUQ DG WUGF VQ KPEQTRQTCVG NCTIGURCP EQP UVTCKPVU KPVQ VJG TGEQIPKVKQP 5KPEG D[ PCVWTG UWEJ KPHQTOCVKQP KU FKHHWUGF CETQUU VJG GPVKTG VGZV DGKPI ETGCVGF VJKU TGSWKTGU VJG FGſPKVKQP QH C document CU C UGOCPVKECNN[ JQOQIGPGQWU UGV QH UGPVGPEGU 6JGP GCEJ FQEWOGPV ECP DG EJCTCEVGTK\GF D[ FTCY KPI HTQO C RQUUKDN[ NCTIG UGV QH VQRKEU WUWCNN[ RTGFGſPGF HTQO C JCPFNCDGNNGF JKGTCTEJ[ YJKEJ EQXGTU VJG TGNGXCPV UGOCPVKE FQOCKP = ? 6JG OCKP WPEGT VCKPV[ KP VJKU CRRTQCEJ KU VJG ITCPWNCTKV[ TGSWKTGF KP VJG VQRKE ENWUVGTKPI RTQEGFWTG =? 6Q KNNWUVTCVG KP CPF GXGP RGTHGEV MPQYNGFIG QH VJG IGPGTCN VQRKE
OQUV NKMGN[ őstock market trendsŒ FQGU PQV JGNR OWEJ #P CNVGTPCVKXG UQNWVKQP KU VQ WUG NQPI FKUVCPEG FGRGPFGPEKGU DGVYGGP YQTF RCKTU YJKEJ UJQY UKIPKſECPV EQTTGNCVKQP KP VJG VTCKPKPI EQTRWU +P VJG CDQXG GZCORNG UWRRQUG VJCV VJG VTCKPKPI FCVC TGXGCNU C UKIPKſECPV EQTTGNCVKQP DGVYGGP őstocksŒ CPF őfellŒ 6JGP VJG RTGUGPEG QH őstocksŒ KP VJG FQEWOGPV EQWNF CWVQOCVKECNN[ VTKIIGT őfellŒ ECWUKPI KVU RTQDCDKNKV[ GUVKOCVG VQ EJCPIG $GECWUG YQTF RTQZKOKV[ KU PQY KTTGNGXCPV VJG VYQ RJTCUGU YQWNF NGCF VQ VJG UCOG TGUWNV +P VJKU CRRTQCEJ VJG RCKT
stocks fell KU UCKF VQ HQTO C YQTF VTKIIGT RCKT =? +P RTCEVKEG YQTF RCKTU YKVJ JKIJ OWVWCN KPHQTOCVKQP CTG UGCTEJGF HQT KPUKFG C YKPFQY QH ſZGF FWTCVKQP 7PHQT VWPCVGN[ VTKIIGT RCKT UGNGEVKQP KU C EQORNGZ KUUWG FKHHGTGPV RCKTU FKURNC[ OCTMGFN[ FKHHGTGPV DGJCXKQT YJKEJ NKOKVU VJG RQVGPVKCN QH NQY HTGSWGPE[ YQTF VTKIIGTU =? 5VKNN UGNHVTKIIGTU JCXG DGGP UJQYP VQ DG RCTVKEWNCTN[ RQYGTHWN CPF TQDWUV =? YJKEJ WPFGTUEQTGU VJG FGUKTCDKNKV[ QH GZRNQKVKPI EQTTGNCVKQPU DGVYGGP VJG EWTTGPV YQTF CPF HGCVWTGU QH VJG FQEWOGPV JKUVQT[ 4GEGPV YQTM JCU UQWIJV VQ GZVGPF VJG YQTF VTKIIGT EQPEGRV D[ WUKPI C OQTG EQO RTGJGPUKXG HTCOGYQTM VQ JCPFNG VJG VTKIIGT RCKT UGNGEVKQP = ? 6JKU KU DCUGF QP C RCTCFKIO QTKIKPCNN[ HQTOWNCVGF KP VJG EQPVGZV QH KPHQTOCVKQP TG VTKGXCN ECNNGF latent semantic analysis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
9.1.4 Organization 6JG HQEWU QH VJKU EJCRVGT KU QP UGOCPVKECNN[FTKXGP URCP GZVGPUKQP QPN[ CPF OQTG URGEKſECNN[ QP JQY VQ GZRNQKV VJG .5# RCTCFKIO VQ KORTQXG UVCVKUVKECN ./ 6JG OCKP QDLGEVKXGU CTG K VQ TGXKGY VJG FCVCFTKXGP GZVTCEVKQP QH NCVGPV UGOCPVKE KPHQT
E\&5&3UHVV//&
OCVKQP KK VQ CUUGUU RQVGPVKCN WUCIG KP URQMGP NCPIWCIG RTQEGUUKPI KKK VQ FGUETKDG KPVGITCVKQP YKVJ EQPXGPVKQPCN ITCO ./ KX VQ GZCOKPG VJG DGJCXKQT QH VJG TG UWNVKPI J[DTKF OQFGNU KP URGGEJ TGEQIPKVKQP GZRGTKOGPVU CPF X VQ FKUEWUU HCEVQTU YJKEJ KPƀWGPEG RGTHQTOCPEG 6JG EJCRVGT KU QTICPK\GF CU HQNNQYU +P VJG PGZV VYQ UGEVKQPU YG IKXG CP QXGTXKGY QH .5# HGCVWTG GZVTCEVKQP CPF VJG TGUWNVKPI .5# HGCVWTG URCEG 5GEVKQP GZRNQTGU VJG CRRNKECDKNKV[ QH VJKU HTCOGYQTM HQT IGPGTCN UGOCPVKE ENCUUKſECVKQP +P 5GEVKQP YG UJKHV VJG HQEWU VQ .5#DCUGF UVCVKUVKECN ./ HQT NCTIG XQECDWNCT[ TGEQIPKVKQP 5GE VKQP FGUETKDGU VJG XCTKQWU UOQQVJKPI RQUUKDKNKVKGU CXCKNCDNG VQ OCMG .5#DCUGF ./U OQTG TQDWUV +P 5GEVKQP YG KNNWUVTCVG UQOG QH VJG DGPGſVU CUUQEKCVGF YKVJ J[DTKF ITCO .5# OQFGNKPI QP C UWDUGV QH VJG 9CNN 5VTGGV ,QWTPCN 95, VCUM (KPCNN[ 5GEVKQP FKUEWUUGU VJG KPJGTGPV VTCFGQHHU CUUQEKCVGF YKVJ VJG CRRTQCEJ CU GXKFGPEGF D[ VJG KPƀWGPEG QH VJG FCVC UGNGEVGF VQ VTCKP VJG .5# EQORQPGPV QH VJG OQFGN
9.2 Latent Semantic Analysis
.GV DG UQOG WPFGTN[KPI XQECDWNCT[ CPF C VTCKPKPI VGZV EQTRWU EQORTKUKPI CTVKENGU FQEWOGPVU TGNGXCPV VQ UQOG FQOCKP QH KPVGTGUV NKMG DWUKPGUU PGYU HQT GZCORNG KP VJG ECUG QH VJG 0#$ EQTRWU =? 6JG .5# RCTCFKIO FGſPGU C OCRRKPI DGVYGGP VJG FKUETGVG UGVU CPF C EQPVKPWQWU XGEVQT URCEG YJGTGD[ GCEJ YQTF KP KU TGRTGUGPVGF D[ C XGEVQT KP CPF GCEJ FQEWOGPV KP KU TGRTGUGPVGF D[ C XGEVQT KP
9.2.1 Feature Extraction
6JG UVCTVKPI RQKPV KU VJG EQPUVTWEVKQP QH C OCVTKZ QH EQQEEWTTGPEGU DGVYGGP YQTFU CPF FQEWOGPVU +P OCTMGF EQPVTCUV YKVJ ITCO OQFGNKPI YQTF QTFGT KU KI PQTGF YJKEJ KU QH EQWTUG KP NKPG YKVJ VJG UGOCPVKE PCVWTG QH VJG CRRTQCEJ =? 6JKU OCMGU KV CP KPUVCPEG QH VJG UQECNNGF őDCIQHYQTFUŒ RCTCFKIO YJKEJ FKUTGICTFU EQNNQECVKQPCN KPHQTOCVKQP KP YQTF UVTKPIU VJG EQPVGZV HQT GCEJ YQTF GUUGPVKCNN[ DG KU CEEWOWNCVGF EQOGU VJG GPVKTG FQEWOGPV KP YJKEJ KV CRRGCTU 6JWU VJG OCVTKZ HTQO VJG CXCKNCDNG VTCKPKPI FCVC D[ UKORN[ MGGRKPI VTCEM QH YJKEJ YQTF KU HQWPF KP YJCV FQEWOGPV 6JKU VGPFU VQ KPXQNXG UQOG CRRTQRTKCVG HWPEVKQP QH VJG YQTF EQWPV KP GCEJ FQEWOGPV =? 8CTKQWU KORNGOGPVCVKQPU JCXG DGGP KPXGUVKICVGF D[ VJG KPHQTOCVKQP TGVTKGXCN EQOOWPKV[ UGG HQT GZCORNG =? 'XKFGPEG RQKPVU VQ VJG FGUKTCDKNKV[ QH PQTOCNK\ KPI HQT FQEWOGPV NGPIVJ CPF YQTF GPVTQR[ 6JWU C UWKVCDNG GZRTGUUKQP HQT VJG EGNN QH KU
E\&5&3UHVV//&
YJGTG KU VJG PWODGT QH VKOGU QEEWTU KP KU VJG VQVCN PWODGT QH YQTFU RTGUGPV KP CPF KU VJG PQTOCNK\GF GPVTQR[ QH KP VJG EQTRWU 6JG INQDCN YGKIJVKPI KORNKGF D[ TGƀGEVU VJG HCEV VJCV VYQ YQTFU CRRGCTKPI YKVJ VJG UCOG EQWPV KP FQ PQV PGEGUUCTKN[ EQPXG[ VJG UCOG COQWPV QH KPHQTOCVKQP CDQWV VJG FQEWOGPV VJKU KU UWDQTFKPCVGF VQ VJG FKUVTKDWVKQP QH VJG YQTFU KP VJG EQNNGEVKQP +H YG FGPQVG D[ VJG VQVCN PWODGT QH VKOGU QEEWTU KP VJG GZRTGUUKQP HQT KU GCUKN[ UGGP VQ DG
$[ FGſPKVKQP YKVJ GSWCNKV[ KH CPF QPN[ KH CPF TGURGEVKXGN[ # XCNWG QH ENQUG VQ KPFKECVGU C YQTF FKUVTKDWVGF CETQUU OCP[ FQE WOGPVU VJTQWIJQWV VJG EQTRWU YJKNG C XCNWG QH ENQUG VQ OGCPU VJCV VJG YQTF KU RTGUGPV QPN[ KP C HGY URGEKſE FQEWOGPVU 6JG INQDCN YGKIJV KU VJGTGHQTG C OGCUWTG QH VJG KPFGZKPI RQYGT QH VJG YQTF
9.2.2 Singular Value Decomposition 6JG
YQTFFQEWOGPV OCVTKZ FGſPGU VYQ XGEVQT TGRTGUGPVCVKQPU HQT VJG YQTFU CPF VJG FQEWOGPVU 'CEJ YQTF ECP DG WPKSWGN[ CUUQEKCVGF YKVJ C TQY XGEVQT QH FKOGPUKQP CPF GCEJ FQEWOGPV ECP DG WPKSWGN[ CUUQEKCVGF YKVJ C EQNWOP XGEVQT QH FKOGPUKQP 7PHQTVWPCVGN[ VJKU KU WPRTCEVKECN HQT VJTGG TGNCVGF TGCUQPU (KTUV VJG FKOGPUKQPU CPF ECP DG GZVTGOGN[ NCTIG UGEQPF VJG XGEVQTU CPF CTG V[RKECNN[ XGT[ URCTUG CPF VJKTF VJG VYQ URCEGU CTG FKUVKPEV HTQO QPG CPQVJGT 6Q CFFTGUU VJGUG KUUWGU QPG UQNWVKQP KU VQ RGTHQTO VJG QTFGT UKPIWNCT XCNWG FG EQORQUKVKQP 58& QH CU =?
YJGTG KU VJG
NGHV UKPIWNCT OCVTKZ YKVJ TQY XGEVQTU KU VJG FKCIQPCN OCVTKZ QH UKPIWNCT XCNWGU KU VJG
TKIJV UKPIWNCT OCVTKZ YKVJ TQY XGEVQTU KU VJG QTFGT QH VJG FGEQORQUKVKQP CPF FGPQVGU OCVTKZ VTCPURQUKVKQP #U KU YGNN KU VJG DGUV TCPM CRRTQZKOCVKQP VQ VJG YQTFFQEWOGPV OCVTKZ HQT MPQYP CP[ WPKVCTKN[ KPXCTKCPV PQTO EH GI =? (WTVJGTOQTG DQVJ NGHV CPF TKIJV UKPIWNCT OCVTKEGU CPF CTG EQNWOPQTVJQPQTOCN KG VJG KFGPVKV[ OCVTKZ QH QTFGT 6JWU VJG EQNWOP XGEVQTU QH CPF GCEJ FGſPG CP QTVJQTPQTOCN DCUKU HQT VJG URCEG QH FKOGPUKQP URCPPGF D[ VJG ŏU CPF ŏU 7RQP RTQLGEVKPI VJG TQY XGEVQTU QH KG YQTFU QPVQ VJG QTVJQTPQTOCN DCUKU HQTOGF D[ VJG EQNWOP XGEVQTU QH VJG TQY XGEVQT EJCTCEVGTK\GU VJG RQUKVKQP QH YQTF KP VJG WPFGTN[KPI FKOGPUKQPCN URCEG HQT 5KOKNCTN[ WRQP RTQLGEVKPI VJG EQNWOP XGEVQTU QH KG FQEWOGPVU QPVQ VJG QTVJQPQTOCN DCUKU HQTOGF D[ VJG EQNWOP XGEVQTU QH VJG TQY XGEVQT EJCTCEVGTK\GU VJG RQ UKVKQP QH FQEWOGPV KP VJG UCOG URCEG HQT 9G TGHGT VQ GCEJ QH VJG
E\&5&3UHVV//&
UECNGF XGEVQTU CU C word vector WPKSWGN[ CUUQEKCVGF YKVJ YQTF KP VJG XQECDWNCT[ CPF GCEJ QH VJG UECNGF XGEVQTU CU C document vector WPKSWGN[ CUUQEKCVGF YKVJ FQEWOGPV KP VJG EQTRWU 6JWU FGſPGU C VTCPUHQT OCVKQP DGVYGGP JKIJFKOGPUKQPCN FKUETGVG GPVKVKGU Î CPF Ì CPF C NQYFKOGPUKQPCN EQPVKPWQWU XGEVQT URCEG Ë VJG FKOGPUKQPCN .5# URCEG URCPPGF D[ VJG ŏU CPF ŏU 6JG FKOGPUKQP KU DQWPFGF HTQO CDQXG D[ VJG WPMPQYP TCPM QH VJG OCVTKZ CPF HTQO DGNQY D[ VJG COQWPV QH FKUVQTVKQP VQNGTCDNG KP VJG FGEQORQUKVKQP +V KU ECRVWTGU VJG OCLQT UVTWEVWTCN CUUQEKCVKQPU KP CPF FGUKTCDNG VQ UGNGEV UQ VJCV
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
9.2.3 General Behavior $[ EQPUVTWEVKQP VJG őENQUGPGUUŒ QH XGEVQTU KP VJG .5# URCEG Ë KU FGVGTOKPGF D[ VJG QXGTCNN RCVVGTP QH VJG NCPIWCIG WUGF KP Ì CU QRRQUGF VQ URGEKſE EQPUVTWEVU *GPEG VYQ YQTFU YJQUG TGRTGUGPVCVKQPU CTG őENQUGŒ KP UQOG UWKVCDNG OGVTKE VGPF VQ CRRGCT KP VJG UCOG MKPF QH FQEWOGPVU YJGVJGT QT PQV VJG[ CEVWCNN[ QEEWT YKVJKP KFGPVKECN YQTF EQPVGZVU KP VJQUG FQEWOGPVU %QPXGTUGN[ VYQ FQEWOGPVU YJQUG TGRTGUGPVC VKQPU CTG őENQUGŒ VGPF VQ EQPXG[ VJG UCOG UGOCPVKE OGCPKPI YJGVJGT QT PQV VJG[ EQPVCKP VJG UCOG YQTF EQPUVTWEVU 9G ECP VJGTGHQTG GZRGEV YQTFU CPF FQEWOGPVU VJCV CTG UGOCPVKECNN[ NKPMGF VQ CNUQ DG őENQUGŒ KP VJG .5# URCEG Ë 1H EQWTUG VJG QRVKOCNKV[ QH VJKU HTCOGYQTM ECP DG FGDCVGF UKPEG VJG WPFGTN[KPI ¾ PQTO OC[ PQV DG VJG DGUV EJQKEG YJGP KV EQOGU VQ NKPIWKUVKE RJGPQOGPC (QT GZCORNG VJG -WNNDCEM.GKDNGT FKXGTIGPEG RTQXKFGU C OQTG GNGICPV RTQDCDKNKUVKE KPVGTRTGVCVKQP QH =? CNDGKV CV VJG GZRGPUG QH TGSWKTKPI C EQPFKVKQPCN KPFGRGP FGPEG CUUWORVKQP QP VJG YQTFU CPF VJG FQEWOGPVU =? 6JKU ECXGCV PQVYKVJUVCPF KPI VJG EQTTGURQPFGPEG DGVYGGP ENQUGPGUU KP .5# URCEG CPF UGOCPVKE TGNCVGFPGUU KU YGNN FQEWOGPVGF +P CRRNKECVKQPU UWEJ CU KPHQTOCVKQP TGVTKGXCN ſNVGTKPI KPFWEVKQP CPF XKUWCNK\CVKQP VJG .5# HTCOGYQTM JCU TGRGCVGFN[ RTQXGP TGOCTMCDN[ GHHGEVKXG KP ECRVWTKPI UGOCPVKE KPHQTOCVKQP = ? 5WEJ DGJCXKQT YCU TGEGPVN[ KNNWUVTCVGF KP =? KP VJG EQPVGZV QH CP CTVKſ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őKPVTCVQRKEŒ KH
E\&5&3UHVV//&
+PVTC6QRKE1TKIKPCN5RCEG
.QI 2TQDCDKNKV[
+PVTC6QRKE .5#5RCEG +PVGT6QRKE 1TKIKPCN5RCEG
+PVGT6QRKE .5#5RCEG
&KUVCPEG 'ZRGEVGF&KUVTKDWVKQPU+P1TKIKPCN5RCEGCPF.5#5RCEG
FIGURE 9.1 Improved Topic Separability in LSA Space.
VJG VYQ FQEWOGPVU YGTG IGPGTCVGF HTQO VJG UCOG VQRKE CPF őKPVGTVQRKEŒ QVJGTYKUG +V ECP DG UGGP VJCV KP VJG .5# URCEG VJG CXGTCIG FKUVCPEG DGVYGGP KPVGTVQRKE RCKTU UVC[U CDQWV VJG UCOG YJKNG VJG CXGTCIG FKUVCPEG DGVYGGP KPVTCVQRKE RCKTU KU FTCOCV KECNN[ TGFWEGF +P CFFKVKQP VJG KPVTCVQRKE UVCPFCTF FGXKCVKQP CNUQ DGEQOGU UWDUVCP VKCNN[ UOCNNGT #U C TGUWNV UGRCTCDKNKV[ DGVYGGP KPVTC CPF KPVGTVQRKE RCKTU KU OWEJ DGVVGT KP VJG .5# URCEG VJCP KP VJG QTKIKPCN URCEG +PVGTGUVKPIN[ VJKU JQNFU FGURKVG C UJCTR KPETGCUG KP VJG KPVGTVQRKE UVCPFCTF FGXKCVKQP YJKEJ DQFGU YGNN HQT VJG IGP GTCN CRRNKECDKNKV[ QH VJG OGVJQF #PCNQIQWU QDUGTXCVKQPU ECP DG OCFG TGICTFKPI VJG FKUVCPEG DGVYGGP YQTFU CPFQT DGVYGGP YQTFU CPF FQEWOGPVU
9.3 LSA Feature Space +P VJG EQPVKPWQWU XGEVQT URCEG Ë QDVCKPGF CDQXG GCEJ YQTF ¾ Î KU TGRTGUGPVGF CPF GCEJ FQEWOGPV ¾ D[ VJG CUUQEKCVGF YQTF XGEVQT QH FKOGPUKQP Ì KU TGRTGUGPVGF D[ VJG CUUQEKCVGF FQEWOGPV XGEVQT QH FKOGPUKQP 6JKU QRGPU WR VJG QRRQTVWPKV[ VQ CRRN[ HCOKNKCT ENWUVGTKPI VGEJPKSWGU KP Ë CU NQPI CU C FKUVCPEG OGCUWTG EQPUKUVGPV YKVJ VJG 58& HQTOCNKUO KU FGſPGF QP VJG XGEVQT URCEG 5KPEG VJG OCVTKZ GODQFKGU D[ EQPUVTWEVKQP CNN UVTWEVWTCN CUUQEKCVKQPU
E\&5&3UHVV//&
DGVYGGP YQTFU CPF FQEWOGPVU KV HQNNQYU VJCV HQT C IKXGP VTCKPKPI EQTRWU EJCTCEVGTK\GU CNN EQQEEWTTGPEGU DGVYGGP YQTFU CPF EJCTCEVGTK\GU CNN EQ QEEWTTGPEGU DGVYGGP FQEWOGPVU
9.3.1 Word Clustering
WUKPI VJG 58& GZRTGUUKQP YG QDVCKP JGPEGHQTVJ KIPQTKPI 'ZRCPFKPI VJG FKUVKPEVKQP DGVYGGP CPF
¾
5KPEG KU FKCIQPCN C PCVWTCN OGVTKE VQ EQPUKFGT HQT VJG őENQUGPGUUŒ DGVYGGP YQTFU KU VJGTGHQTG VJG EQUKPG QH VJG CPING DGVYGGP CPF
¾
HQT CP[ # XCNWG QH OGCPU VJG VYQ YQTFU CNYC[U QEEWT KP VJG UCOG UGOCPVKE EQPVGZV YJKNG C XCNWG QH OGCPU VJG VYQ YQTFU
CTG WUGF KP KPETGCUKPIN[ FKHHGTGPV UGOCPVKE EQPVGZVU 9JKNG FQGU PQV FGſPG C DQPC ſFG FKUVCPEG OGCUWTG KP VJG URCEG KV GCU[ NGCFU VQ QPG (QT GZCORNG QXGT VJG KPVGTXCN VJG OGCUWTG
EQU ½
TGCFKN[ UCVKUſGU VJG RTQRGTVKGU QH C FKUVCPEG QP #V VJKU RQKPV KV KU UVTCKIJVHQTYCTF VQ RTQEGGF YKVJ VJG ENWUVGTKPI QH VJG YQTF XGEVQTU WUKPI CP[ QH C XCTKGV[ QH CN IQTKVJOU UGG HQT KPUVCPEG =? 6JG QWVEQOG KU C UGV QH ENWUVGTU YJKEJ ECP DG VJQWIJV QH CU TGXGCNKPI C RCTVKEWNCT NC[GT QH UGOCPVKE MPQYNGFIG KP VJG URCEG
9.3.2 Word Cluster Example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ődrawingŒ HTQO ENWUVGT CPF VJG RTGUGPV VGPUG XGTD őruleŒ HTQO ENWUVGT 6JKU KU CP KPUVCPEG QH polysemy ődrawingŒ CPF őruleŒ CTG OQTG NKMGN[ VQ CRRGCT KP VJG VTCKPKPI VGZV YKVJ VJGKT CNVGTPCVKXG
E\&5&3UHVV//&
Cluster 1
#PF[ CPVKSWG CPVKSWGU CTV CTVKUV CTVKUVŏU CTVKUVU CTVYQTMU CWEVKQPGGTU %JTKUVKGŏU EQNNGEVQT FTCYKPIU ICNNGT[ )QIJ HGVEJGF J[UVGTKC OCUVGTRKGEG OWUGWOU RCKPVGT RCKPVKPI RCKPVKPIU 2KECUUQ 2QNNQEM TGRTQFWEVKQP 5QVJGD[ŏU XCP 8KPEGPV 9CTJQN Cluster 2
CRRGCN CRRGCNU CVVQTPG[ CVVQTPG[ŏU EQWPVU EQWTV EQWTVŏU EQWTVU EQPFGOPGF EQPXKEVKQPU ETKOKPCN FGEKUKQP FGHGPF FGHGPFCPV FKUOKUUGU FKUOKUUGF JGCTKPI JGTG KPFKEVGF KPFKEVOGPV KPFKEVOGPVU LWFIG LWFKEKCN LWFKEKCT[ LWT[ LWTKGU NCYUWKV NGPKGPE[ QXGTVWTPGF RNCKPVKHHU RTQUGEWVG RTQUGEWVKQP RTQUGEWVKQPU RTQUGEWVQTU TWNGF TWNKPI UGPVGPEGF UGPVGPEKPI UWKPI UWKV UWKVU YKVPGUU
FIGURE 9.2 Word Cluster Example (After [2]). OGCPKPIU CU KP ődrawing a conclusionŒ CPF őbreaking a ruleŒ TGURGEVKXGN[ VJWU TGUWNVKPI KP FKHHGTGPV ENWUVGT CUUKIPOGPVU (KPCNN[ UQOG YQTFU UGGO VQ EQPVTKDWVG QPN[ OCTIKPCNN[ VQ VJG ENWUVGTU HQT GZCORNG őhysteriaŒ HTQO ENWUVGT CPF őhereŒ HTQO ENWUVGT 6JGUG CTG VJG WPCXQKFCDNG QWVNKGTU CV VJG RGTKRJGT[ QH VJG ENWUVGTU
9.3.3 Document Clustering 2TQEGGFKPI KP C UKOKNCT HCUJKQP CV VJG FQEWOGPV NGXGN YG QDVCKP
¾
YJKEJ HQT
NGCFU VQ VJG UCOG HWPEVKQPCN HQTO CU ¾
9G EQPENWFG VJCV VJG FKUVCPEG KU GSWCNN[ XCNKF HQT DQVJ YQTF CPF FQEWOGPV ENWUVGTKPI 6JG TGUWNVKPI UGV QH ENWUVGTU ECP DG XKGYGF CU TGXGCNKPI CPQVJGT NC[GT QH UGOCPVKE MPQYNGFIG KP VJG URCEG
+P HCEV VJG OGCUWTG KU RTGEKUGN[ VJG QPG WUGF KP VJG UVWF[ TGRQTVGF KP (KIWTG 6JWU VJG FKUVCPEGU
QP VJG ZCZKU QH (KIWTG CTG
E\&5&3UHVV//&
GZRTGUUGF KP TCFKCPU
2TQDCDKNKV[
0CVWTCN5EKGPEG #RRNKGF5EKGPEG 5QEKCN5EKGPEG +OCIKPCVKXG
5WDFQOCKP %NWUVGT+PFGZ 2TQDCDKNKV[&KUVTKDWVKQPUQH(QWT$0%6QRKEU#ICKPUV.5#&QEWOGPV%NWUVGTU
FIGURE 9.3 Document Cluster Example.
9.3.4 Document Cluster Example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
E\&5&3UHVV//&
9.4 Semantic Classification 6Q UWOOCTK\G VJG NCVGPV UGOCPVKE HTCOGYQTM JCU C PWODGT QH KPVGTGUVKPI RTQRGT VKGU KPENWFKPI K C UKPING XGEVQT TGRTGUGPVCVKQP HQT DQVJ YQTFU CPF FQEWOGPVU KP VJG UCOG EQPVKPWQWU XGEVQT URCEG KK CP WPFGTN[KPI VQRQNQIKECN UVTWEVWTG TGƀGEV KPI UGOCPVKE UKOKNCTKV[ KKK C YGNNOQVKXCVGF PCVWTCN OGVTKE VQ OGCUWTG VJG FKUVCPEG DGVYGGP YQTFU CPF DGVYGGP FQEWOGPVU KP VJCV URCEG CPF KX C TGNCVKXGN[ NQY FKOGP UKQPCNKV[ YJKEJ OCMGU ENWUVGTKPI OGCPKPIHWN CPF RTCEVKECN 6JGUG RTQRGTVKGU ECP DG GZRNQKVGF KP UGXGTCN CTGCU QH URQMGP NCPIWCIG RTQEGUUKPI +P VJKU UGEVKQP YG CFFTGUU VJG OQUV KOOGFKCVG FQOCKP QH CRRNKECVKQP YJKEJ HQNNQYU FKTGEVN[ HTQO VJG RTGXKQWU ENWUVGTKPI FKUEWUUKQP FCVCFTKXGP UGOCPVKE ENCUUKſECVKQP = ?
9.4.1 Framework Extension 5GOCPVKE ENCUUKſECVKQP FGVGTOKPGU HQT C IKXGP FQEWOGPV YJKEJ QPG QH UGXGTCN RTG FGſPGF VQRKEU VJG FQEWOGPV KU OQUV ENQUGN[ CNKIPGF YKVJ +P EQPVTCUV YKVJ VJG ENWU VGTKPI UGVWR FKUEWUUGF CDQXG UWEJ FQEWOGPV YKNN PQV PQTOCNN[ JCXG DGGP UGGP KP VJG VTCKPKPI EQTRWU *GPEG YG ſTUV PGGF VQ GZVGPF VJG .5# HTCOGYQTM CEEQTFKPIN[ #U KV VWTPU QWV WPFGT TGNCVKXGN[ OKNF CUUWORVKQPU ſPFKPI C TGRTGUGPVCVKQP HQT C PGY FQEWOGPV KP VJG URCEG Ë KU UVTCKIJVHQTYCTF .GV WU FGPQVG VJG PGY FQEWOGPV D[ Ô YJGTG VJG VKNFG U[ODQN TGƀGEVU VJG HCEV VJCV +V KU QDVCKPGF D[ EQPUVTWEVKPI C HGCVWTG XGEVQT EQPVCKPKPI HQT GCEJ YQTF KP VJG WPFGTN[KPI XQECDWNCT[ VJG YGKIJVGF EQWPVU YKVJ 6JKU XGEVQT Ô CU C EQNWOP XGEVQT QH FKOGPUKQP ECP DG VJQWIJV QH CU CP CFFKVKQPCN EQNWOP QH VJG OCVTKZ 6JWU RTQXKFGF VJG OCVTKEGU CPF FQ PQV EJCPIG VJG 58& GZRCPUKQP
KORNKGU
Ô ÔÌ YJGTG VJG FKOGPUKQPCN XGEVQT ÔÌ CEVU CU CP CFFKVKQPCN EQNWOP QH VJG OCVTKZ Ì 6JKU KP VWTP NGCFU VQ VJG FGſPKVKQP
Ô Ô ÔÌ
Ô KPFGGF UGGP VQ DG HWPEVKQPCNN[ UKOKNCT VQ C FQEWOGPV XGEVQT EQTTG 6JG XGEVQT URQPFU VQ VJG TGRTGUGPVCVKQP QH VJG PGY FQEWOGPV KP VJG URCEG Ë Ô KU TGHGTTGF VQ CU C 6Q EQPXG[ VJG HCEV VJCV KV YCU PQV RCTV QH VJG 58& GZVTCEVKQP pseudo document vector 4GECNN VJCV VJG VTWPECVGF 58& RTQXKFGU D[ FGſPKVKQP C RCTUKOQPKQWU FGUETKRVKQP QH VJG NKPGCT URCEG URCPPGF D[ +H VJG PGY FQEWOGPV EQPVCKPU NCPIWCIG RCVVGTPU YJKEJ CTG KPEQPUKUVGPV YKVJ VJQUG GZVTCEVGF HTQO VJG 58& GZRCPUKQP YKNN PQ NQPIGT CRRN[ 5KOKNCTN[ KH VJG CFFKVKQP QH Ô ECWUGU VJG OCLQT UVTWEVWTCN CUUQEKCVKQPU KP VQ UJKHV KP UQOG UWDUVCPVKCN OCPPGT VJG RCTUKOQ PKQWU FGUETKRVKQP YKNN DGEQOG KPCFGSWCVG 6JGP CPF YKNN PQ NQPIGT DG XCNKF KP YJKEJ ECUG KV YQWNF DG PGEGUUCT[ VQ TGEQORWVG VQ ſPF C RTQRGT TGRTGUGPVCVKQP
E\&5&3UHVV//&
HQT Ý +H QP VJG QVJGT JCPF VJG PGY FQEWOGPV IGPGTCNN[ EQPHQTOU VQ VJG TGUV QH VJG EQTRWU VJGP KP YKNN DG C TGCUQPCDNG TGRTGUGPVCVKQP HQT KV 1PEG VJG TGRTGUGPVCVKQP KU QDVCKPGF VJG őENQUGPGUUŒ DGVYGGP VJG PGY FQEW OGPV CPF CP[ FQEWOGPV ENWUVGT ECP VJGP DG GZRTGUUGF CU ECNEWNCVGF HTQO KP VJG RTGXKQWU UGEVKQP
9.4.2 Semantic Inference 6JKU ECP DG TGCFKN[ GZRNQKVGF KP UWEJ EQOOCPFCPFEQPVTQN VCUMU CU FGUMVQR WUGT KPVGTHCEG EQPVTQN =? QT CWVQOCVGF ECNN TQWVKPI =? 5WRRQUG VJCV GCEJ FQEWOGPV ENWUVGT ECP DG WPKSWGN[ CUUQEKCVGF YKVJ C RCTVKEWNCT CEVKQP KP VJG VCUM 6JGP VJG EGPVTQKF QH GCEJ ENWUVGT ECP DG XKGYGF CU VJG semantic anchor QH VJKU CEVKQP KP VJG .5# URCEG #P WPMPQYP YQTF UGSWGPEG VTGCVGF CU C PGY őFQEWOGPVŒ ECP VJWU DG OCRRGF QPVQ CP CEVKQP D[ GXCNWCVKPI VJG FKUVCPEG DGVYGGP VJCV őFQEWOGPVŒ CPF GCEJ UGOCPVKE CPEJQT 9G TGHGT VQ VJKU CRRTQCEJ CU semantic inference = ? +P EQPVTCUV YKVJ WUWCN KPHGTGPEG GPIKPGU EH =? UGOCPVKE KPHGTGPEG VJWU FGſPGF FQGU PQV TGN[ QP HQTOCN DGJCXKQTCN RTKPEKRNGU GZVTCEVGF HTQO C MPQYNGFIG DCUG +PUVGCF VJG FQOCKP MPQYNGFIG KU CWVQOCVKECNN[ GPECRUWNCVGF KP VJG .5# URCEG KP C FCVC FTKXGP HCUJKQP 6Q KNNWUVTCVG EQPUKFGT CP CRRNKECVKQP YKVJ CEVKQPU FQEWOGPVU GCEJ CUUQ EKCVGF YKVJ C WPKSWG EQOOCPF K őwhat is the timeŒ KK őwhat is the dayŒ KKK őwhat time is the meetingŒ CPF KX őcancel the meetingŒ 6JKU UKORNG GZCORNG YKVJ C XQECDWNCT[ QH QPN[ YQTFU KU FGUKIPGF UWEJ VJCV őwhatŒ CPF őisŒ CNYC[U EQ QEEWT őtheŒ CRRGCTU KP CNN HQWT EQOOCPFU QPN[ KK CPF KX EQPVCKP C WPKSWG YQTF CPF K KU C RTQRGT UWDUGV QH KKK %QPUVTWEVKPI VJG YQTFFQEWOGPV OCVTKZ CU CDQXG CPF RGTHQTOKPI VJG 58& YG QDVCKP VJG FKOGPUKQPCN URCEG UJQYP KP (KIWTG YJGTG KU FGRKEVGF VJG TGRTGUGPVCVKQP QH GCEJ YQTF CPF GCEJ EQOOCPF KP VJG CRRNKECVKQP 6JG VYQ YQTFU YJKEJ GCEJ WPKSWGN[ KFGPVKH[ C EQOOCPFōődayŒ HQT KK CPF őcancelŒ HQT KXōGCEJ JCXG C JKIJ EQQTFKPCVG QP C FKHHGTGPV CZKU %QPXGTUGN[ VJG YQTF őtheŒ YJKEJ EQPXG[U PQ KPHQTOCVKQP CDQWV VJG KFGPVKV[ QH C EQOOCPF KU NQECVGF CV VJG QTKIKP 1P VJG QVJGT JCPF VJG UGOCPVKE CPEJQTU HQT KK CPF KX HCNN őENQUGŒ VQ VJG YQTFU YJKEJ RTGFKEV VJGO DGUVōődayŒ CPF őcancelŒ TGURGEVKXGN[ 5KOK NCTN[ VJG UGOCPVKE CPEJQTU HQT K CPF KKK HCNN KP VJG XKEKPKV[ QH VJGKT OGCPKPIHWN EQORQPGPVUōőwhat–isŒ CPF őtimeŒ HQT K CPF őtimeŒ CPF őmeetingŒ HQT KKKōYKVJ VJG YQTF őtimeŒ YJKEJ QEEWTU KP DQVJ KPFGGF CRRGCTKPI őENQUGŒ VQ DQVJ 0QY UWR RQUG VJCV C WUGT UC[U UQOGVJKPI QWVUKFG QH VJG VTCKPKPI UGVWR UWEJ CU őwhen is the meetingŒ TCVJGT VJCP őwhat time is the meetingŒ 6JKU PGY YQTF UVTKPI QT XCTKCPV KU TGRTGUGPVGF KP VJG URCEG D[ VJG JQNNQY VTKCPING KP (KIWTG YJKEJ KU ENQUGUV VQ Ý (QT GZCORNG UWRRQUG VTCKPKPI YCU ECTTKGF QWV HQT C DCPMKPI CRRNKECVKQP KPXQNXKPI VJG YQTF őDCPMŒ VCMGP KP C ſPCPEKCN EQPVGZV 0QY UWRRQUG KU IGTOCPG VQ C ſUJKPI CRRNKECVKQP YJGTG őDCPMŒ KU TGHGTTGF VQ KP VJG EQPVGZV QH C TKXGT QT C NCMG %NGCTN[ VJG ENQUGPGUU QH őDCPMŒ VQ GI őOQPG[Œ CPF őCEEQWPVŒ YQWNF DG KTTGNGXCPV VQ %QPXGTUGN[ CFFKPI VQ YQWNF NKMGN[ ECWUG UWEJ UVTWEVWTCN CUUQEKCVKQPU VQ UJKHV UWDUVCPVKCNN[ CPF RGTJCRU GXGP FKUCRRGCT CNVQIGVJGT
E\&5&3UHVV//&
5GEQPF58&&KOGPUKQP
YJCVKUVJGFC[ FC[
YQTF EQOOCPF PGYXCTKCPV
YJCV KU YJCVVKOGKUVJGOGGVKPI
YJCVKU VJGVKOG
VKOG YJGPKUVJGOGGVKPI
OGGVKPI
ECPEGN
VJG
ECPEGN VJG OGGVKPI
(KTUV58&&KOGPUKQP 6YQ&KOGPUKQPCN+NNWUVTCVKQPQH.5#5RCEG
FIGURE 9.4 An Example of Semantic Inference for Command and Control (Ê ). VJG TGRTGUGPVCVKQP QH EQOOCPF KKK 6JWU VJG PGY XCTKCPV CRRGCTU OQUV UGOCPVKECNN[ TGNCVGF VQ KKK CPF VJG EQTTGEV CEVKQP ECP DG CWVQOCVKECNN[ KPHGTTGF 6JKU ECP DG VJQWIJV QH CU C YC[ VQ RGTHQTO őDQVVQOWRŒ PCVWTCN NCPIWCIG WPFGT UVCPFKPI $[ TGRNCEKPI VJG VTCFKVKQPCN TWNGDCUGF OCRRKPI DGVYGGP WVVGTCPEG CPF CEVKQP D[ UWEJ FCVCFTKXGP ENCUUKſECVKQP UGOCPVKE KPHGTGPEG OCMGU KV RQUUKDNG VQ TG NCZ UQOG QH VJG V[RKECN EQOOCPFCPFEQPVTQN KPVGTCEVKQP EQPUVTCKPVU (QT GZCORNG KV QDXKCVGU VJG PGGF VQ URGEKH[ TKIKF NCPIWCIG EQPUVTWEVU VJTQWIJ C FQOCKPURGEKſE
CPF VJWU V[RKECNN[ JCPFETCHVGF ſPKVG UVCVG ITCOOCT 6JKU KU VWTP CNNQYU VJG GPF WUGT OQTG ƀGZKDKNKV[ KP GZRTGUUKPI VJG FGUKTGF EQOOCPFSWGT[ YJKEJ VGPFU VQ TG FWEG VJG CUUQEKCVGF EQIPKVKXG NQCF CPF VJGTGD[ GPJCPEG WUGT UCVKUHCEVKQP =?
9.4.3 Caveats #U CP KPUVCPEG QH VJG őDCIQHYQTFUŒ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őnotŒ KP C ETWEKCN RNCEG 6JG TGURGEVKXG XGEVQT TGRTGUGPVCVKQPU EQWNF EQPEGKXCDN[ DG TGNCVKXGN[ ENQUG
E\&5&3UHVV//&
KP VJG .5# URCEG CPF [GV JCXG XCUVN[ FKHHGTGPV KPVGPFGF EQPUGSWGPEGU 9QTUG [GV UQOG EQOOCPFU OC[ FKHHGT QPN[ VJTQWIJ YQTF QTFGT %QPUKFGT HQT KPUVCPEG VJG VYQ /CE15 EQOOCPFU change popup to window change window to popup
YJKEJ CTG QDXKQWUN[ KORQUUKDNG VQ FKUCODKIWCVG UKPEG VJG[ CTG OCRRGF QPVQ VJG exact same point KP .5# URCEG #U KV VWTPU QWV KV KU RQUUKDNG VQ JCPFNG UWEJ ECUGU VJTQWIJ CP GZVGPUKQP QH VJG DC UKE .5# HTCOGYQTM WUKPI YQTF CIINQOGTCVKQP 6JG KFGC KU VQ OQXG HTQO YQTFU CPF FQEWOGPVU VQ YQTF VWRNGU CPF VWRNG FQEWOGPVU YJGTG GCEJ YQTF VWRNG KU VJG CIINQOGTCVKQP QH UWEEGUUKXG YQTFU CPF GCEJ VWRNG FQEWOGPV KU PQY GZ RTGUUGF KP VGTOU QH CNN VJG YQTF VWRNGU KV EQPVCKPU &GURKVG VJG TGUWNVKPI KPETGCUG KP EQORWVCVKQPCN EQORNGZKV[ VJKU GZVGPUKQP KU RTCEVKECN KP VJG EQPVGZV QH UGOCPVKE ENCUUKſECVKQP DGECWUG QH VJG TGNCVKXGN[ OQFGUV FKOGPUKQPU KPXQNXGF CU EQORCTGF VQ NCTIG XQECDWNCT[ TGEQIPKVKQP (WTVJGT FGVCKNU YQWNF DG DG[QPF VJG UEQRG QH VJKU OCPWUETKRV DWV VJG TGCFGT KU TGHGTTGF VQ =? HQT C EQORNGVG FGUETKRVKQP
9.5 N-gram+LSA Language Modeling #PQVJGT OCLQT CTGC QH CRRNKECVKQP QH VJG .5# HTCOGYQTM KU KP UVCVKUVKECN ./ YJGTG KV ECP TGCFKN[ UGTXG CU C RCTCFKIO HQT UGOCPVKECNN[FTKXGP URCP GZVGPUKQP $GECWUG QH VJG NKOKVCVKQP LWUV FKUEWUUGF JQYGXGT KV KU DGUV CRRNKGF KP EQPLWPEVKQP YKVJ VJG UVCPFCTF ITCO CRRTQCEJ 6JKU UGEVKQP FGUETKDGU JQY VJKU ECP DG FQPG
9.5.1 LSA Component .GV FGPQVG VJG YQTF CDQWV VQ DG RTGFKEVGF CPF ½ VJG CFOKUUKDNG .5# JKUVQT[
EQPVGZV HQT VJKU RCTVKEWNCT YQTF 6JKU PQVCVKQP VTCPUNCVGU C ECWUCNKV[ TGUVTKEVKQP QH VJG EQPVGZV VQ ½ VJG EWTTGPV FQEWOGPV UQ HCT KG WR VQ YQTF ½ 6JWU KP IGPGTCN VGTOU VJG .5# ./ RTQDCDKNKV[ KU IKXGP D[ 2T
½
2T
½
YJGTG VJG EQPFKVKQPKPI QP TGƀGEVU VJG HCEV VJCV VJG RTQDCDKNKV[ FGRGPFU QP VJG RCTVKEWNCT XGEVQT URCEG CTKUKPI HTQO VJG 58& TGRTGUGPVCVKQP +P VJKU GZRTGUUKQP 2T ½ KU EQORWVGF FKTGEVN[ HTQO VJG TGRTGUGPVCVKQPU QH CPF ½ KP VJG URCEG KG KV KU KPHGTTGF HTQO VJG őENQUGPGUUŒ DGVYGGP VJG CUUQEKCVGF YQTF XGEVQT CPF RUGWFQ FQEWOGPV XGEVQT KP 9G VJGTGHQTG JCXG VQ URGEKH[ DQVJ VJG CRRTQRTK CVG RUGWFQ FQEWOGPV TGRTGUGPVCVKQP CPF VJG TGNGXCPV RTQDCDKNKV[ OGCUWTG
E\&5&3UHVV//&
9.5.1.1 2UGWFQ &QEWOGPV 4GRTGUGPVCVKQP 6Q EQOG WR YKVJ C RUGWFQ FQEWOGPV TGRTGUGPVCVKQP YG NGXGTCIG VJG TGUWNVU QH 5GE VKQP YKVJ UQOG UNKIJV OQFKſECVKQPU FWG VQ VJG VKOGXCT[KPI PCVWTG QH VJG URCP EQPUKFGTGF (TQO ½ NGCFU VQ VJG TGRTGUGPVCVKQP
½
½
½
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½ CPF FKHHGT QPN[ KP QPG EQQTFKPCVG EQTTGURQPFKPI VQ VJG KPFGZ #UUWOG HWTVJGT VJCV VJG VTCKPKPI EQTRWU KU NCTIG GPQWIJ UQ VJCV VJG PQTOCNK\GF GPVTQR[ FQGU PQV EJCPIG CRRTGEKCDN[ YKVJ VJG CFFKVKQP QH GCEJ RUGWFQ FQEWOGPV 6JKU OCMGU KV RQUUKDNG HTQO VQ GZRTGUU CU
½
YJGTG VJG őŒ CRRGCTU CV EQQTFKPCVG 6JKU KU VWTP KORNKGU HTQO
½
#U C TGUWNV VJG RUGWFQ FQEWOGPV XGEVQT CUUQEKCVGF YKVJ VJG NCTIGURCP EQPVGZV ECP DG GHſEKGPVN[ WRFCVGF FKTGEVN[ KP VJG .5# URCEG 9.5.1.2 .5# 2TQDCDKNKV[ 6Q URGEKH[ C UWKVCDNG őENQUGPGUUŒ OGCUWTG YG PQY HQNNQY C TGCUQPKPI UKOKNCT VQ VJCV QH 5GEVKQP 5KPEG D[ EQPUVTWEVKQP VJG OCVTKZ GODQFKGU UVTWEVWTCN CUUQEKCVKQPU DGVYGGP YQTFU CPF FQEWOGPVU CPF D[ FGſPKVKQP C PCVWTCN OGVTKE VQ EQPUKFGT HQT VJG őENQUGPGUUŒ DGVYGGP YQTF CPF FQEWOGPV KU VJG EQUKPG QH VJG CPING DGVYGGP ½¾ CPF ½¾ #RRN[KPI VJG UCOG TGCUQPKPI VQ RUGWFQ FQEWOGPVU YG CTTKXG CV
½ ¾ ½
½¾
½
½ ½¾ ½ ½¾
HQT CP[ KPFGZKPI C YQTF KP VJG VGZV FCVC # XCNWG QH ½ OGCPU VJCV ½ KU C UVTQPI UGOCPVKE RTGFKEVQT QH YJKNG C XCNWG QH ½
E\&5&3UHVV//&
OGCPU VJCV VJG JKUVQT[ ECTTKGU KPETGCUKPIN[ NGUU KPHQTOCVKQP CDQWV VJG EWTTGPV YQTF +PVGTGUVKPIN[ KU HWPEVKQPCNN[ GSWKXCNGPV VQ CPF DWV KPXQNXGU UECNKPI D[ ½¾ KPUVGCF QH #U DGHQTG VJG OCRRKPI ECP DG WUGF VQ VTCPUHQTO KPVQ C TGCN FKUVCPEG OGCUWTG 6Q GPCDNG VJG EQORWVCVKQP QH 2T ½ KV TGOCKPU VQ IQ HTQO VJCV FKUVCPEG OGC UWTG VQ CP CEVWCN RTQDCDKNKV[ OGCUWTG 1PG RQUUKDNG UQNWVKQP KU HQT VJG FKUVCPEG OGCUWTG VQ KPFWEG C HCOKN[ QH GZRQPGPVKCN FKUVTKDWVKQPU YKVJ RGTVKPGPV OCTIKPCNKV[ EQPUVTCKPVU +P RTCEVKEG KV OC[ PQV DG PGEGUUCT[ VQ KPEWT VJKU FGITGG QH EQORNGZ KV[ %QPUKFGTKPI VJCV ½ KU QPN[ C RCTVKCN FQEWOGPV CP[YC[ GZCEVN[ YJCV MKPF QH FKUVTKDWVKQP KU KPFWEGF KU RTQDCDN[ NGUU EQPUGSWGPVKCN VJCP GPUWTKPI VJCV VJG RUGWFQ FQEWOGPV KU RTQRGTN[ UEQRGF EH 5GEVKQP DGNQY $CUKECNN[ CNN VJCV KU PGGFGF KU C őTGCUQPCDNGŒ RTQDCDKNKV[ FKUVTKDWVKQP VQ CEV CU C RTQZ[ HQT VJG VTWG WPMPQYP OGCUWTG 9G VJGTGHQTG QRV VQ WUG VJG GORKTKECN OWNVKXCTKCVG FKUVTKDWVKQP EQPUVTWEVGF D[ CNNQECV KPI VJG VQVCN RTQDCDKNKV[ OCUU KP RTQRQTVKQP VQ VJG FKUVCPEGU QDUGTXGF FWTKPI VTCKPKPI +P GUUGPEG VJKU TGFWEGU VJG EQORNGZKV[ VQ C UKORNG JKUVQITCO PQTOCNK\CVKQP CV VJG GZRGPUG QH KPVTQFWEKPI C RQVGPVKCN őSWCPVK\CVKQPNKMGŒ GTTQT 1H EQWTUG UWEJ GTTQT ECP DG OKPKOK\GF VJTQWIJ C XCTKGV[ QH JKUVQITCO UOQQVJKPI VGEJPKSWGU #NUQ PQVG VJCV VJG F[PCOKE TCPIG QH VJG FKUVTKDWVKQP V[RKECNN[ PGGFU VQ DG EQPVTQNNGF D[ C RC TCOGVGT VJCV KU QRVKOK\GF GORKTKECNN[ GI D[ CP GZRQPGPV QP VJG FKUVCPEG VGTO CU FKUEWUUGF KP =? +PVWKVKXGN[ 2T ½ TGƀGEVU VJG őTGNGXCPEGŒ QH YQTF VQ VJG CFOKUUKDNG JKU VQT[ CU QDUGTXGF VJTQWIJ ½ #U UWEJ KV YKNN DG JKIJGUV HQT YQTFU YJQUG OGCPKPI CNKIPU OQUV ENQUGN[ YKVJ VJG UGOCPVKE HCDTKE QH ½ KG TGNGXCPV őEQPVGPVŒ YQTFU CPF NQYGUV HQT YQTFU YJKEJ FQ PQV EQPXG[ CP[ RCTVKEWNCT KPHQTOCVKQP CDQWV VJKU HCDTKE GI őHWPEVKQPŒ YQTFU NKMG őtheŒ 6JKU DGJCXKQT KU GZCEVN[ VJG QRRQUKVG QH VJCV QDUGTXGF YKVJ VJG EQPXGPVKQPCN ITCO HQTOCNKUO YJKEJ VGPFU VQ CUUKIP JKIJGT RTQDCDKNKVKGU VQ HTGSWGPV HWPEVKQP YQTFU VJCP VQ TCTGT EQPVGPV YQTFU *GPEG VJG CVVTCEVKXG U[PGTI[ RQVGPVKCN DGVYGGP VJG VYQ RCTCFKIOU
9.5.2 Integration with N-grams 'ZRNQKVKPI VJKU RQVGPVKCN TGSWKTGU NGXGTCIKPI VJG DGPGſVU QH DQVJ KP C EQPUVTWEVKXG OCPPGT 6JKU MKPF QH KPVGITCVKQP ECP QEEWT KP C PWODGT QH YC[U UWEJ CU UKORNG KP VGTRQNCVKQP = ? QT YKVJKP VJG OCZKOWO GPVTQR[ HTCOGYQTM = ? #NVGT PCVKXGN[ WPFGT TGNCVKXGN[ OKNF CUUWORVKQPU KV KU CNUQ RQUUKDNG VQ FGTKXG CP KPVGITCVGF HQTOWNCVKQP FKTGEVN[ HTQO VJG GZRTGUUKQP HQT VJG QXGTCNN ./ RTQDCDKNKV[ 9G UVCTV YKVJ VJG FGſPKVKQP ´·µ ½
2T
´ µ ´ µ ½ ½
2T
YJGTG ½ FGPQVGU CU DGHQTG UQOG UWKVCDNG CFOKUUKDNG JKUVQT[ HQT YQTF CPF VJG UWRGTUETKRVU ´µ ´µ CPF ´·µ TGHGT VQ VJG ITCO EQORQPGPV ½ ¾ ·½ YKVJ VJG .5# EQORQPGPV ½ CPF VJG KPVGITCVKQP VJGTGQH TG
E\&5&3UHVV//&
URGEVKXGN[ 6JKU GZRTGUUKQP ECP DG TGYTKVVGP CU 2T
´·µ ½
´µ½ ´µ 2T ½ ½ ´ µ
2T
¾Î
½ ´ µ
YJGTG VJG UWOOCVKQP KP VJG FGPQOKPCVQT GZVGPFU QXGT CNN YQTFU KP CPF TGCTTCPIKPI VJG PWOGTCVQT QH KU UGGP VQ DG
'ZRCPFKPI
´µ ´µ 2T ½ ½ ´µ ´µ ´µ 2T ½ 2T ½ ½ 2T ½ ¾ ·½ 2T ½ ½ ¾ ·½
0QY YG OCMG VJG CUUWORVKQP VJCV VJG RTQDCDKNKV[ QH VJG FQEWOGPV JKUVQT[ IKXGP VJG EWTTGPV YQTF KU PQV CHHGEVGF D[ VJG KOOGFKCVG EQPVGZV RTGEGFKPI KV 6JKU TGƀGEVU VJG HCEV VJCV HQT C IKXGP YQTF FKHHGTGPV U[PVCEVKE EQPUVTWEVU KOOGFKCVG EQPVGZV ECP DG WUGF VQ ECTT[ VJG UCOG OGCPKPI FQEWOGPV JKUVQT[ 6JKU KU QDXKQWUN[ TGCUQPCDNG HQT EQPVGPV YQTFU *QY OWEJ KV OCVVGTU HQT HWPEVKQP YQTFU KU NGUU ENGCT =? DWV YG EQPLGEVWTG VJCV KH VJG FQEWOGPV JKUVQT[ KU NQPI GPQWIJ VJG UGOCPVKE CPEJQTKPI KU UWHſEKGPVN[ UVTQPI HQT VJG CUUWORVKQP VQ JQNF #U C TGUWNV VJG KPVGITCVGF RTQDCDKNKV[ DGEQOGU ´·µ 2T ½
2T ½ ¾ 2T ½ ¾
¾Î
·½ 2T ½ ·½ 2T ½
+H 2T ½ KU XKGYGF CU C RTKQT RTQDCDKNKV[ QP VJG EWTTGPV FQEWOGPV JKUVQT[ VJGP UKORN[ VTCPUNCVGU VJG ENCUUKECN $C[GUKCP GUVKOCVKQP QH VJG ITCO NQECN RTQDCDKNKV[ WUKPI C RTKQT FKUVTKDWVKQP QDVCKPGF HTQO INQDCN .5# 6JG GPF TGUWNV KP GHHGEV KU C OQFKſGF ITCO ./ KPEQTRQTCVKPI NCTIGURCP UGOCPVKE KPHQTOCVKQP 6JG FGRGPFGPEG QH QP VJG .5# RTQDCDKNKV[ ECNEWNCVGF GCTNKGT ECP DG GZRTGUUGF GZRNKEKVN[ D[ WUKPI $C[GUŏ TWNG VQ IGV 2T ½ KP VGTOU QH 2T ½ 5KPEG VJG SWCPVKV[ 2T ½ XCPKUJGU HTQO DQVJ PWOGTCVQT CPF FGPQOKPCVQT YG CTG NGHV YKVJ ´·µ 2T ½
2T ½ 2T 2T ½ 2T ½ ¾ ·½ ¾Î 2T ½ ¾ ·½
E\&5&3UHVV//&
YJGTG 2T KU UKORN[ VJG UVCPFCTF WPKITCO RTQDCDKNKV[ 0QVG VJCV VJKU GZRTGUUKQP KU OGCPKPIHWN HQT CP[ Þ
9.5.3 Context Scope Selection +P RTCEVKEG GZRTGUUKQPU NKMG Ō CTG QHVGP UNKIJVN[ OQFKſGF UQ VJCV C TGN CVKXG YGKIJV ECP DG RNCEGF QP GCEJ EQPVTKDWVKQP JGTG VJG ITCO CPF .5# RTQD CDKNKVKGU 7UWCNN[ VJKU KU FQPG XKC GORKTKECNN[ FGVGTOKPGF YGKIJVKPI EQGHſEKGPVU +P VJG RTGUGPV ECUG UWEJ YGKIJVKPI KU OQVKXCVGF D[ VJG HCEV VJCV KP VJG őRTKQTŒ RTQDCDKNKV[ 2T ½ EQWNF EJCPIG UWDUVCPVKCNN[ CU VJG EWTTGPV FQEWOGPV WPHQNFU 6JWU TCVJGT VJCP WUKPI CTDKVTCT[ YGKIJVU CP CNVGTPCVKXG CRRTQCEJ KU VQ F[PCOKECNN[ VCKNQT VJG FQEWOGPV JKUVQT[ ½ UQ VJCV VJG ITCO CPF .5# EQPVTKDWVKQPU TGOCKP GORKTKECNN[ DCNCPEGF 6JKU CRRTQCEJ TGHGTTGF VQ CU EQPVGZV UEQRG UGNGEVKQP KU OQTG ENQUGN[ CNKIPGF YKVJ VJG .5# HTCOGYQTM DGECWUG QH VJG WPFGTN[KPI EJCPIG KP DGJCXKQT DGVYGGP VTCKPKPI CPF TGEQIPKVKQP &WTKPI VTCKPKPI VJG UEQRG KU ſZGF VQ DG VJG EWTTGPV FQEWOGPV &WTKPI TGEQIPKVKQP JQYGXGT VJG EQPEGRV QH őEWTTGPV FQEWOGPVŒ KU KNNFGſPGF DGECWUG K KVU NGPIVJ ITQYU YKVJ GCEJ PGY YQTF CPF KK KV KU PQV PGEGUUCTKN[ ENGCT CV YJKEJ RQKPV EQORNGVKQP QEEWTU #U C TGUWNV C FGEKUKQP JCU VQ DG OCFG TGICTFKPI YJCV VQ EQPUKFGT őEWTTGPVŒ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
½
YJGTG VJG RCTCOGVGT KU EJQUGP CEEQTFKPI VQ VJG GZRGEVGF JGVGTQIGPGKV[ QH VJG UGU UKQP +P VGTOU QH EQORWVCVKQPCN GHHQTV VJG QPNKPG EQUV KPEWTTGF FWTKPI TGEQIPKVKQP EQO RTKUGU K VJG EQPUVTWEVKQP QH VJG RUGWFQ FQEWOGPV TGRTGUGPVCVKQP KP CU IGPGTCNN[ FQPG XKC KK VJG EQORWVCVKQP QH VJG .5# RTQDCDKNKV[ 2T ½ KP CPF KKK VJG KPVGITCVKQP RTQRGT KP +V ECP DG UJQYP EH = ? VJCV VJG VQVCN EQUV QH VJGUG QRGTCVKQPU RGT YQTF CPF RUGWFQ FQEWOGPV KU ¾ 6JKU KU QDXKQWUN[ OQTG GZRGPUKXG VJCP VJG WUWCN VCDNG NQQMWR TGSWKTGF KP EQPXGPVKQPCN ITCO ./ ;GV HQT V[RKECN XCNWGU QH VJG TGUWNVKPI QXGTJGCF KU CTIWCDN[ SWKVG OQFGUV 6JKU CNNQYU J[DTKF ITCO .5# ./ VQ DG DTQWIJV VQ DGCT KP GCTN[ UVCIGU QH VJG UGCTEJ =?
KU YKVJQWV NQUU QH IGPGTCNKV[ 9JGP VJG TKIJV JCPF UKFG QH
FGIGPGTCVGU VQ VJG .5# RTQDCDKNKV[ CNQPG UKPEG VJG ITCO JKUVQT[ DGEQOGU PWNN $WV VJG KPVGITCVGF JKUVQT[ CNUQ FGIGPGTCVGU VQ VJG .5# JKUVQT[ CNQPG GHHGEVKXGN[ TGFWEKPI VQ Þ /QTGQXGT VJG CUUWORVKQP VJCV
E\&5&3UHVV//&
9.6 Smoothing 5KPEG VJG FGTKXCVKQP QH FQGU PQV FGRGPF QP C RCTVKEWNCT HQTO QH VJG .5# RTQD CDKNKV[ KV KU RQUUKDNG VQ VCMG CFXCPVCIG QH VJG CFFKVKQPCN NC[GT U QH MPQYNGFIG WPEQX GTGF GCTNKGT VJTQWIJ YQTF KP 5GEVKQP CPF FQEWOGPV KP 5GEVKQP ENWUVGT KPI $CUKECNN[ YG ECP GZRGEV YQTFU CPFQT FQEWOGPVU TGNCVGF VQ VJG EWTTGPV FQEW OGPV VQ EQPVTKDWVG YKVJ OQTG U[PGTI[ CPF WPTGNCVGF YQTFU CPFQT FQEWOGPVU VQ DG DGVVGT FKUEQWPVGF %NWUVGTKPI VJGTGHQTG RTQXKFGU C EQPXGPKGPV UOQQVJKPI OGEJCPKUO KP VJG .5# URCEG = ?
9.6.1 Word Smoothing 7UKPI VJG UGV QH YQTF ENWUVGTU RTQFWEGF KP 5GEVKQP NGCFU VQ YQTFDCUGF UOQQVJKPI +P VJKU ECUG YG GZRCPF CU HQNNQYU 2T
½
2T 2T
YJKEJ ECTTKGU QXGT VQ KP C UVTCKIJVHQTYCTF OCPPGT +P VJG RTQDCDKNKV[ 2T KU SWCNKVCVKXGN[ UKOKNCT VQ CPF ECP VJGTGHQTG DG QDVCKPGF YKVJ VJG JGNR QH D[ UKORN[ TGRNCEKPI VJG TGRTGUGPVCVKQP QH VJG YQTF D[ VJCV QH VJG EGPVTQKF QH YQTF ENWUVGT +P EQPVTCUV VJG RTQDCDKNKV[ 2T FGRGPFU QP VJG őENQUGPGUUŒ QH TGNCVKXG VQ VJKU YQTF EGPVTQKF 6Q FGTKXG KV YG VJGTGHQTG JCXG VQ TGN[ QP VJG GORKTKECN OWNVKXCTKCVG FKUVTKDWVKQP KPFWEGF PQV D[ VJG FKUVCPEG QDVCKPGF HTQO DWV D[ VJCV QDVCKPGF HTQO VJG OGCUWTG OGPVKQPGF KP 5GEVKQP 0QVG VJCV C FKUVKPEV FKUVTKDWVKQP ECP DG KPHGTTGF QP GCEJ QH VJG ENWUVGTU VJWU CNNQYKPI WU VQ EQORWVG CNN SWCPVKVKGU 2T HQT CPF 6JG DGJCXKQT QH VJG OQFGN FGRGPFU QP VJG PWODGT QH YQTF ENWUVGTU FGſPGF KP VJG URCEG 6YQ URGEKCN ECUGU CTKUG CV VJG GZVTGOGU QH VJG ENWUVGT TCPIG +H VJGTG CTG CU OCP[ ENCUUGU CU YQTFU KP VJG XQECDWNCT[ VJGP YKVJ VJG EQPXGPVKQP VJCV Æ UKORN[ TGFWEGU VQ 0Q UOQQVJKPI KU KPVTQFWEGF UQ VJG RTGFKEVKXG RQYGT QH VJG OQFGN UVC[U VJG UCOG CU DGHQTG %QPXGTUGN[ KH CNN VJG YQTFU CTG KP C UKPING ENCUU VJG OQFGN DGEQOGU OCZKOCNN[ UOQQVJ VJG KPƀWGPEG QH URGEKſ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ſE CPF HCKN VQ TGƀGEV VJG QXGTCNN UG OCPVKE HCDTKE QH 6JWU VJGTG OWUV GZKUV C ENWUVGT UGV UK\G YJGTG VJG FGITGG
E\&5&3UHVV//&
QH UOQQVJKPI CPF VJGTGHQTG VJG CUUQEKCVGF RTGFKEVKXG RQYGT KU QRVKOCN HQT VJG VCUM EQPUKFGTGF 6JKU JCU KPFGGF DGGP XGTKſGF GZRGTKOGPVCNN[ EH =?
9.6.2 Document Smoothing
'ZRNQKVKPI KPUVGCF VJG UGV QH FQEWOGPV ENWUVGTU RTQFWEGF KP 5GE VKQP NGCFU VQ FQEWOGPVDCUGF UOQQVJKPI 6JG GZRCPUKQP KU UKOKNCT 2T
½
2T
2T
YKVJ FQEWOGPV ENWUVGTU PQY TGRNCEKPI VJG YQTF ENWUVGTU 6JKU VKOG KV KU VJG RTQDCDKNKV[ 2T YJKEJ KU SWCNKVCVKXGN[ UKOKNCT VQ CPF ECP VJGTGHQTG DG QDVCKPGF YKVJ VJG JGNR QH #U HQT VJG RTQDCDKNKV[ 2T KV FGRGPFU QP VJG őENQUGPGUUŒ QH TGNCVKXG VQ VJG EGPVTQKF QH FQEWOGPV ENWUVGT 6JWU KV ECP DG QDVCKPGF VJTQWIJ VJG GORKTKECN OWNVKXCTKCVG FKUVTKDWVKQP KPFWEGF D[ VJG FKUVCPEG FGTKXGF HTQO KP 5GEVKQP #ICKP VJG DGJCXKQT QH VJG OQFGN FGRGPFU QP VJG PWODGT QH FQEWOGPV ENWUVGTU FGſPGF KP VJG URCEG %QORCTGF VQ JQYGXGT KU OQTG FKHſEWNV VQ KPVGTRTGV CV VJG GZVTGOGU QH VJG ENWUVGT TCPIG KG CPF +H HQT GZCORNG FQGU PQV TGFWEG VQ DGECWUG JCU PQV DGGP UGGP KP VJG VTCKPKPI FCVC CPF VJGTGHQTG ECPPQV DG KFGPVKſGF YKVJ CP[ QH VJG GZKUVKPI ENWUVGTU 5KOKNCTN[ VJG HCEV VJCV CNN VJG FQEWOGPVU CTG KP C UKPING ENWUVGT FQGU PQV KORN[ VJG FGITGG QH FGIGPGTCVGPGUU QDUGTXGF RTGXKQWUN[ DGECWUG VJG ENWUVGT KVUGNH KU UVTQPIN[ KPFKECVKXG QH VJG IGPGTCN FKUEQWTUG FQOCKP YJKEJ YCU PQV IGPGTCNN[ VTWG QH VJG őXQECDWNCT[ ENWUVGTŒ CDQXG *GPEG FGRGPFKPI QP VJG UK\G CPF UVTWEVWTG QH VJG EQTRWU VJG OQFGN OC[ YGNN DG KORQTVCPV VQ ECRVWTG IGPGTCN FKUEQWTUG GHHGEVU 6Q UGG VJCV YG CRRN[ KP VQ QDVCKP KP
2T 2T
¾Î
2T
2T 2T
2T
UKPEG VJG SWCPVKV[ 2T XCPKUJGU HTQO DQVJ PWOGTCVQT CPF FGPQOKPCVQT +P VJKU GZRTGUUKQP TGHGTU VQ VJG UKPING FQEWOGPV ENWUVGT GPEQORCUUKPI CNN FQEWOGPVU KP VJG .5# URCEG +P ECUG VJG EQTRWU KU HCKTN[ JQOQIGPGQWU YKNN DG C OQTG TGNK CDNG TGRTGUGPVCVKQP QH VJG WPFGTN[KPI HCDTKE QH VJG FQOCKP VJCP CPF VJGTGHQTG CEV CU C TQDWUV RTQZ[ HQT VJG EQPVGZV QDUGTXGF +PVGTGUVKPIN[ COQWPVU VQ GUVK OCVKPI C őEQTTGEVKQPŒ HCEVQT HQT GCEJ YQTF YJKEJ FGRGPFU QPN[ QP VJG QXGTCNN VQRKE QH VJG EQNNGEVKQP 6JKU KU UKOKNCT VQ YJCV KU FQPG KP VJG ECEJG CRRTQCEJ VQ ./ CFCR VCVKQP UGG HQT GZCORNG = ? GZEGRV VJCV JGTG CNN YQTFU CTG VTGCVGF CU VJQWIJ VJG[ YGTG CNTGCF[ KP VJG ECEJG
E\&5&3UHVV//&
/QTG IGPGTCNN[ CU VJG PWODGT QH FQEWOGPV ENCUUGU KPETGCUGU VJG EQPVTKDWVKQP QH 2T VGPFU VQ KPETGCUG VQ VJG GZVGPV VJCV C OQTG JQOQIGPGQWU VQRKE DQQUVU VJG GHHGEVU QH CP[ TGNCVGF EQPVGPV YQTFU 1P VJG QVJGT JCPF VJG EQPVTKDWVKQP QH 2T ½ VGPFU VQ FGETGCUG DGECWUG VJG ENWUVGTU TGRTGUGPV OQTG CPF OQTG URG EKſE VQRKEU YJKEJ KPETGCUGU VJG EJCPEG VJCV VJG RUGWFQ FQEWOGPV ½ DGEQOGU CP QWVNKGT 6JWU CICKP VJGTG GZKUVU C ENWUVGT UGV UK\G YJGTG VJG FGITGG QH UOQQVJKPI KU QRVKOCN HQT VJG VCUM EQPUKFGTGF EH =?
9.6.3 Joint Smoothing (KPCNN[ CP GZRTGUUKQP CPCNQIQWU VQ CPF ECP CNUQ DG FGTKXGF VQ VCMG CFXCPVCIG QH DQVJ YQTF CPF FQEWOGPV ENWUVGTU 6JKU NGCFU VQ C OKZVWTG RTQDCDKNKV[ URGEKſGF D[ 2T
½
2T
2T
YJKEJ HQT VTCEVCDKNKV[ ECP DG CRRTQZKOCVGF CU 2T
2T
2T 2T
+P VJKU GZRTGUUKQP VJG ENWUVGTU CPF CTG CU RTGXKQWUN[ CU CTG VJG SWCPVKVKGU 2T CPF 2T #U HQT VJG RTQDCDKNKV[ 2T KV KU SWCNKVCVKXGN[ UKOKNCT VQ CPF ECP VJGTGHQTG DG QDVCKPGF CEEQTFKPIN[ 6Q UWOOCTK\G CP[ QH VJG GZRTGUUKQPU QT ECP DG WUGF VQ EQORWVG 6JKU TGUWNVU KP HQWT HCOKNKGU QH J[DTKF ITCO .5# ./U #UUQEK CVGF YKVJ VJGUG FKHHGTGPV HCOKNKGU CTG XCTKQWU VTCFGQHHU VQ DGEQOG CRRCTGPV DGNQY
9.7 Experiments
6JG RWTRQUG QH VJKU UGEVKQP KU VQ KNNWUVTCVG VJG DGJCXKQT QH J[DTKF ITCO .5# OQF GNKPI QP C NCTIG XQECDWNCT[ TGEQIPKVKQP VCUM Ü 6JG IGPGTCN FQOCKP EQPUKFGTGF YCU DWUKPGUU PGYU CU TGƀGEVGF KP VJG 95, RQTVKQP QH VJG 0#$ EQTRWU 6JKU YCU EQPXG PKGPV HQT EQORCTKUQP RWTRQUGU UKPEG EQPXGPVKQPCN ITCO ./U CTG TGCFKN[ CXCKNCDNG VTCKPGF QP GZCEVN[ VJG UCOG FCVC =?
Ü 6JG TGCFGT KU TGHGTTGF VQ =? HQT CFFKVKQPCN TGUWNVU KP VJKU CRRNKECVKQP CPF VQ =? HQT GZRGTKOGPVU KPXQNXKPI
UGOCPVKE KPHGTGPEG
E\&5&3UHVV//&
9.7.1 Experimental Conditions
6JG VGZV EQTRWU Ì WUGF KP VJKU VTCKPKPI YCU EQORQUGF QH CDQWV FQEW OGPVU URCPPKPI VJG [GCTU VQ EQORTKUKPI CRRTQZKOCVGN[ OKNNKQP YQTFU 6JG XQECDWNCT[ Î YCU EQPUVTWEVGF D[ VCMKPI VJG OQUV HTGSWGPV YQTFU QH VJG 0#$ EQTRWU CWIOGPVGF D[ UQOG YQTFU HTQO CP GCTNKGT TGNGCUG QH VJG 95, EQT RWU HQT C VQVCN QH YQTFU 6JG VGUV UGV EQPUKUVGF QH C VGUV EQT RWU QH UGPVGPEGU WVVGTGF D[ PCVKXG URGCMGTU QH 'PINKUJ +P CNN GZRGTKOGPVU CEQWUVKE VTCKPKPI YCU RGTHQTOGF WUKPI UGPVGPEGU QH FCVC WVVGTGF D[ URGCMGTU
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ōOKPKOK\KPI ·½ VJG NCTIGUV UKPIWNCT XCNWG PQV TGVCKPGFōCPF PQKUG UWRRTGUUKQPōOKPKOK\KPI VJG TCVKQ DGVYGGP QTFGT CPF QTFGT VTCEGU 6JKU NGF VQ C XGEVQT URCEG Ë QH FKOGPUKQP (QNNQYKPI 5GEVKQP YG VJGP WUGF VJKU .5# URCEG VQ EQPUVTWEV VJG WPUOQQVJGF .5# OQFGN 9G CNUQ EQPUVTWEVGF VJG XCTKQWU ENWUVGTGF .5# OQFGNU RTGUGPVGF KP 5GEVKQP VQ KORNGOGPV YQTF UOQQVJKPI DCUGF QP FQEWOGPV UOQQVJKPI DCUGF QP CPF LQKPV UOQQVJKPI DCUGF QP 9G GZRGTKOGPVGF YKVJ FKHHGT GPV XCNWGU HQT VJG PWODGT QH YQTF CPFQT FQEWOGPV ENWUVGTU EH =? CPF GPFGF WR WUKPI YQTF ENWUVGTU CPF FQEWOGPV ENWUVGT (KPCNN[ WUKPI YG EQODKPGF GCEJ QH VJGUG OQFGNU YKVJ GKVJGT VJG UVCPFCTF 95, DKITCO QT VJG UVCP FCTF 95, VTKITCO 6JG TGUWNVKPI J[DTKF ITCO .5# ./U FWDDGF DK.5# CPF VTK.5# OQFGNU TGURGEVKXGN[ YGTG VJGP WUGF KP NKGW QH VJG UVCPFCTF 95, DKITCO CPF VTKITCO OQFGNU
9.7.2 Experimental Results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
E\&5&3UHVV//&
TABLE 9.1
9QTF 'TTQT 4CVG 9'4 4GUWNVU 7UKPI *[DTKF $K.5# CPF 6TK.5# /QFGNU 9QTF 'TTQT 4CVG 9'4 4GFWEVKQP %QPXGPVKQPCN )TCO *[DTKF 0Q 5OQQVJKPI *[DTKF &QEWOGPV 5OQQVJKPI *[DTKF 9QTF 5OQQVJKPI *[DTKF ,QKPV 5OQQVJKPI
$KITCO
6TKITCO
CEJKGXGF D[ DK.5# 6JKU KU OQUV NKMGN[ TGNCVGF VQ VJG ITGCVGT RTGFKEVKXG RQYGT QH VJG VTKITCO EQORCTGF VQ VJG DKITCO YJKEJ OCMGU VJG .5# EQPVTKDWVKQP QH VJG J[DTKF ./ EQORCTCVKXGN[ UOCNNGT 6JKU KU EQPUKUVGPV YKVJ VJG HCEV VJCV VJG NCVGPV UGOCPVKE KPHQTOCVKQP FGNKXGTGF D[ VJG .5# EQORQPGPV YQWNF GXGPVWCNN[ DG UWDUWOGF D[ CP ITCO YKVJ C NCTIG GPQWIJ +PVGTGUVKPIN[ KP DQVJ ECUGU VJG CXGTCIG 9'4 TGFWEVKQP KU HCT HTQO EQPUVCPV CETQUU KPFKXKFWCN UGUUKQPU TGƀGEVKPI VJG XCT[KPI TQNG RNC[GF D[ INQDCN UGOCPVKE EQPUVTCKPVU HTQO QPG UGV QH URQMGP WVVGTCPEGU VQ CPQVJGT
9.7.3 Context Scope Selection +V KU KORQTVCPV VQ GORJCUK\G VJCV VJG TGEQIPKVKQP VCUM EJQUGP CDQXG TGRTGUGPVU C UG XGTG VGUV QH VJG .5# EQORQPGPV QH VJG J[DTKF ./ $[ FGUKIP VJG VGUV EQTRWU KU EQPUVTWEVGF YKVJ PQ OQTG VJCP VJTGG QT HQWT EQPUGEWVKXG UGPVGPEGU GZVTCEVGF HTQO C UKPING CTVKENG 1XGTCNN KV EQORTKUGU FKUVKPEV FQEWOGPV HTCIOGPVU YJKEJ OGCPU VJCV GCEJ URGCMGT URGCMU QP VJG CXGTCIG CDQWV FKHHGTGPV őOKPKFQEWOGPVUŒ #U C TGUWNV VJG EQPVGZV GHHGEVKXGN[ EJCPIGU GXGT[ YQTFU QT UQ YJKEJ OCMGU KV UQOG YJCV EJCNNGPIKPI VQ DWKNF C XGT[ CEEWTCVG RUGWFQ FQEWOGPV TGRTGUGPVCVKQP 6JKU KU C UKVWCVKQP YJGTG KV KU ETKVKECN HQT VJG .5# EQORQPGPV VQ CRRTQRTKCVGN[ HQTIGV VJG EQPVGZV CU KV WPHQNFU VQ CXQKF TGN[KPI QP CP QDUQNGVG TGRTGUGPVCVKQP 6Q QDVCKP VJG TGUWNVU QH 6CDNG YG WUGF VJG GZRQPGPVKCN HQTIGVVKPI UGVWR QH YKVJ C XCNWG ß +P QTFGT VQ CUUGUU VJG KPƀ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ß 6Q ſZ KFGCU VJKU OGCPU VJCV VJG YQTF YJKEJ QEEWTTGF YQTFU CIQ KU FKUEQWPVGF VJTQWIJ C YGKIJV QH CDQWV
E\&5&3UHVV//&
TABLE 9.2
+PƀWGPEG QH %QPVGZV 5EQRG 5GNGEVKQP QP 9QTF 'TTQT 4CVG 9QTF 'TTQT 4CVG 9'4 4GFWEVKQP
$K.5# YKVJ 9QTF 5OQQVJKPI
VJG QXGTCNN RGTHQTOCPEG KU UWDUVCPVKCNN[ NGUU VJCP VJG EQORCTCDNG QPG QDUGTXGF KP 6CDNG EQORCTGF VQ 9'4 TGFWEVKQP 6JKU KU EQPUKUVGPV YKVJ VJG EJCT CEVGTKUVKEU QH VJG VCUM CPF WPFGTUEQTGU VJG TQNG QH FKUEQWPVKPI CU C UWKVCDNG EQWPVGT DCNCPEG VQ HTGSWGPV EQPVGZV EJCPIGU 2GTHQTOCPEG TCRKFN[ KORTQXGU CU FGETGCUGU HTQO VQ RTGUWOCDN[ DGECWUG VJG RUGWFQ FQEWOGPV TGRTGUGPVCVKQP IGVU NGUU CPF NGUU EQPVCOKPCVGF YKVJ QDUQNGVG FCVC +H HQTIGVVKPI DGEQOGU VQQ CIITGU UKXG JQYGXGT VJG RGTHQTOCPEG UVCTVU FGITCFKPI CU VJG GHHGEVKXG EQPVGZV PQ NQPIGT JCU CP GSWKXCNGPV NGPIVJ YJKEJ KU UWHſEKGPV HQT VJG VCUM CV JCPF *GTG VJKU JCRRGPU HQT
9.8 Inherent Trade-Offs +P VJG RTGXKQWU UGEVKQP DQVJ .5# CPF ITCO EQORQPGPVU QH VJG J[DTKF ./ YGTG VTCKPGF QP GZCEVN[ VJG UCOG FCVC 6JKU KU PQV C TGSWKTGOGPV JQYGXGT YJKEJ TCKUGU VJG SWGUVKQP QH JQY ETKVKECN VJG UGNGEVKQP QH VJG .5# VTCKPKPI FCVC KU VQ VJG RGTHQTOCPEG QH VJG TGEQIPK\GT 6JKU KU RCTVKEWNCTN[ KPVGTGUVKPI UKPEG .5# KU MPQYP VQ DG YGCMGT QP JGVGTQIGPGQWU EQTRQTC UGG HQT GZCORNG =?
9.8.1 Cross-Domain Training 6Q CUEGTVCKP VJG OCVVGT YG YGPV DCEM VQ CP .5# EQORQPGPV KPXQNXKPI VJG QTKIKPCN WPUOQQVJGF OQFGN 9G MGRV VJG UCOG WPFGTN[KPI XQECDWNCT[ Î NGHV VJG DK ITCO EQORQPGPV WPEJCPIGF CPF TGRGCVGF VJG .5# VTCKPKPI QP PQP95, FCVC HTQO VJG UCOG IGPGTCN RGTKQF 6JTGG EQTRQTC QH KPETGCUKPI UK\G YGTG EQPUKFGTGF CNN EQT TGURQPFKPI VQ #UUQEKCVGF 2TGUU #2 FCVC K Ì ½ EQORQUGF QH ½ FQEW OGPVU HTQO EQORTKUKPI CRRTQZKOCVGN[ OKNNKQP YQTFU KK Ì ¾ EQORQUGF QH ¾ FQEWOGPVU HTQO CPF EQORTKUKPI CRRTQZKOCVGN[ OKN NKQP YQTFU CPF KKK Ì ¿ EQORQUGF QH ¿
FQEWOGPVU HTQO
E\&5&3UHVV//&
TABLE 9.3
/QFGN 5GPUKVKXKV[ VQ .5# 6TCKPKPI &CVC 9QTF 'TTQT 4CVG 9'4 4GFWEVKQP ̽ ½ ̾ ¾ Ì¿ ¿
$K.5# YKVJ 0Q 5OQQVJKPI
EQORTKUKPI CRRTQZKOCVGN[ OKNNKQP YQTFU +P GCEJ ECUG .5# VTCKPKPI RTQEGGFGF CU FGUETKDGF KP 5GEVKQP 6JG TGUWNVU CTG TGRQTVGF KP 6CDNG 6YQ VJKPIU CTG KOOGFKCVGN[ CRRCTGPV (KTUV VJG RGTHQTOCPEG KORTQXGOGPV KP CNN ECUGU KU OWEJ UOCNNGT VJCP VJG TGFWEVKQP QDUGTXGF KP 6CDNG QP VJG CXGT CIG VJG J[DTKF OQFGN VTCKPGF QP #2 FCVC KU CDQWV HQWT VKOGU NGUU GHHGEVKXG VJCP VJCV VTCKPGF QP 95, FCVC 6JKU UWIIGUVU C TGNCVKXGN[ JKIJ .5# UGPUKVKXKV[ VQ VJG FQOCKP EQPUKFGTGF 6Q RWV VJKU QDUGTXCVKQP KP RGTURGEVKXG TGECNN VJCV K D[ FGſ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ƀGEVGF HQT GZCORNG KP VJG EJQKEG QH EQPVGPV YQTFU CPFQT YQTF EQ QEEWTTGPEGU 1P VJG RQUKVKXG UKFG VJKU DQFGU YGNN HQT TCRKF CFCRVCVKQP VQ ETQUU FQOCKP FCVC RTQXKFGF C UWKVCDNG CFCRVCVKQP HTCOGYQTM ECP DG FGTKXGF
9.8.2 Discussion 6JG HCEV VJCV VJG J[DTKF ITCO .5# CRRTQCEJ KU UGPUKVKXG VQ EQORQUKVKQP UV[NG WPFGTUEQTGU VJG TGNCVKXGN[ PCTTQY UGOCPVKE URGEKſEKV[ QH VJG .5# RCTCFKIO 9JKNG ITCOU CNUQ UWHHGT HTQO CP[ OKUOCVEJ DGVYGGP VTCKPKPI CPF TGEQIPKVKQP .5# NGCFU VQ C RQVGPVKCNN[ OQTG UGXGTG GZRQUWTG DGECWUG VJG URCEG Ë TGƀGEVU GXGP NGUU QH VJG RTCIOCVKE EJCTCEVGTKUVKEU HQT VJG VCUM EQPUKFGTGF 2GTJCRU YJCV KU TGSWKTGF KU VQ GZRNKEKVN[ KPENWFG CP őCWVJQTUJKR UV[NGŒ EQORQPGPV KPVQ VJG .5# HTCOGYQTM +P CP[ GXGPV QPG JCU VQ DG EQIPK\CPV QH VJKU KPVTKPUKE NKOKVCVKQP CPF OKVKICVG KV VJTQWIJ ECTGHWN CVVGPVKQP VQ VJG GZRGEVGF FQOCKP QH WUG
+P =? HQT GZCORNG KV JCU DGGP UWIIGUVGF VQ FGſPG CP UVQEJCUVKE OCVTKZ C OCVTKZ YKVJ PQP PGICVKXG GPVTKGU CPF TQY UWOU GSWCN VQ VQ CEEQWPV HQT VJG YC[ UV[NG OQFKſGU VJG HTGSWGPE[ QH YQTFU 6JKU UQNWVKQP JQYGXGT OCMGU VJG CUUWORVKQPōPQV CNYC[U XCNKFōVJCV VJKU KPƀWGPEG KU KPFGRGPFGPV QH VJG WPFGTN[KPI UWDLGEV OCVVGT
E\&5&3UHVV//&
#PQVJGT ECXGCV KU VJG HCEV VJCV .5# KU KPJGTGPVN[ OQTG CFGRV CV JCPFNKPI EQPVGPV YQTFU VJCP HWPEVKQP YQTFU #U KU YGNNMPQYP C UWDUVCPVKCN RTQRQTVKQP QH URGGEJ TGEQIPKVKQP GTTQTU EQOG HTQO HWPEVKQP YQTFU DGECWUG QH VJGKT VGPFGPE[ VQ DG UJQTVGT PQV YGNN CTVKEWNCVGF CPF CEQWUVKECNN[ EQPHWUCDNG +P IGPGTCN .5#ŏU EQPVTKDWVKQP VQ ſZKPI UWEJ RTQDNGOU YKNN DG NKOKVGF 6JKU UWIIGUVU VJCV GXGP YKVJKP C YGNN URGEKſGF FQOCKP U[PVCEVKECNN[FTKXGP URCP GZVGPUKQP VGEJPKSWGU OC[ DG C PGEGUUCT[ EQORNGOGPV VQ VJG J[DTKF CRRTQCEJ 1P VJCV UWDLGEV PQVG HTQO 5GEVKQP VJCV VJG KPVGITCVGF JKUVQT[ EQWNF GCUKN[ DG OQFKſGF VQ TGƀGEV C JGCFYQTFDCUGF ITCO CU QRRQUGF VQ C EQPXGPVKQPCN ITCO JKUVQT[ YKVJQWV KPXCNKFCVKPI VJG FGTKXCVKQP QH 6JWU VJGTG KU PQ VJGQTGVKECN DCTTKGT VQ VJG KPVGITCVKQP QH NCVGPV UGOCPVKE KPHQTOCVKQP YKVJ UVTWEVWTGF ./U UWEJ CU FGUETKDGF KP = ? 5KOKNCTN[ VJGTG KU PQ TGCUQP YJ[ VJG .5# RCTCFKIO EQWNF PQV DG WUGF KP EQPLWPEVKQP YKVJ VJG KPVGITCVKXG CRRTQCEJGU QH VJG MKPF RTQRQUGF KP = ? QT GXGP YKVJKP VJG ECEJG CFCRVKXG HTCOGYQTM = ?
9.9 Conclusion 5VCVKUVKECN ITCOU CTG D[ PCVWTG NKOKVGF VQ VJG ECRVWTG QH NKPIWKUVKE RJGPQOGPC URCP PKPI CV OQUV YQTFU 6JKU EJCRVGT JCU HQEWUGF QP C UGOCPVKECNN[FTKXGP URCP GZVGP UKQP HTCOGYQTM DCUGF QP VJG .5# RCTCFKIO KP YJKEJ JKFFGP UGOCPVKE TGFWPFCPEKGU CTG VTCEMGF CETQUU UGOCPVKECNN[ JQOQIGPGQWU FQEWOGPVU 6JKU CRRTQCEJ NGCFU VQ C
EQPVKPWQWU XGEVQT TGRTGUGPVCVKQP QH GCEJ FKUETGVG YQTF CPF FQEWOGPV KP C URCEG QH TGNCVKXGN[ OQFGUV FKOGPUKQP KP YJKEJ UWKVCDNG OGVTKEU ECP DG FGſPGF HQT YQTF FQEWOGPV YQTFYQTF CPF FQEWOGPVFQEWOGPV EQORCTKUQPU #U YGNNMPQYP ENWU VGTKPI CNIQTKVJOU ECP VJGP DG CRRNKGF GHſEKGPVN[ VJKU OCMGU KV RQUUKDNG VQ WPEQXGT KP C FCVCFTKXGP HCUJKQP OWNVKRNG RCTCNNGN NC[GTU QH UGOCPVKE MPQYNGFIG YKVJ XCTKCDNG ITCPWNCTKV[ #P KORQTVCPV RTQRGTV[ QH VJKU XGEVQT TGRTGUGPVCVKQP KU VJCV KV TGƀGEVU VJG OCLQT UG OCPVKE CUUQEKCVKQPU KP VJG VTCKPKPI EQTRWU CU FGVGTOKPGF D[ VJG QXGTCNN RCVVGTP QH VJG NCPIWCIG CU QRRQUGF VQ URGEKſ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
E\&5&3UHVV//&
GTCNN[ UGOCPVKECNN[FTKXGP URCP GZVGPUKQPU NKMG VJG QPG RTQRQUGF JGTG TWP VJG TKUM QH NCEMNWUVGT KORTQXGOGPV YJGP KV EQOGU VQ HWPEVKQP YQTF TGEQIPKVKQP 6JKU WPFGT UEQTGU VJG PGGF HQT CP CNNGPEQORCUUKPI UVTCVGI[ KPXQNXKPI U[PVCEVKECNN[ OQVKXCVGF CRRTQCEJGU CU YGNN
References =? ,4 $GNNGICTFC Context-Dependent Vector Clustering for Speech Recognition %JCRVGT KP #WVQOCVKE 5RGGEJ CPF 5RGCMGT 4GEQIPKVKQP #FXCPEGF 6QRKEU %* .GG (- 5QQPI CPF -- 2CNKYCN 'FU -NWYGT #ECFGOKE 2WDNKUJGTU 0; RR Ō /CTEJ =? ,4 $GNNGICTFC A Multi-Span Language Modeling Framework for Large Vocabulary Speech Recognition +''' 6TCPU 5RGGEJ #WFKQ 2TQE 8QN 0Q RR Ō 5GRVGODGT =? ,4 $GNNGICTFC Large Vocabulary Speech Recognition With Multi-Span Statistical Language Models +''' 6TCPU 5RGGEJ #WFKQ 2TQE 8QN 0Q RR Ō ,CPWCT[ =? ,4 $GNNGICTFC Exploiting Latent Semantic Information in Statistical Language Modeling 2TQE +''' 5RGE +UUWG 5RGGEJ 4GEQI 7PFGTUVCPFKPI $* ,WCPI CPF 5 (WTWK 'FU 8QN 0Q RR Ō #WIWUV =? ,4 $GNNGICTFC Robustness in Statistical Language Modeling: Review and Perspectives %JCRVGT KP 4QDWUVPGUU KP .CPIWCIG CPF 5RGGEJ 6GEJPQN QI[ ,% ,WPSWC CPF ),/ XCP 0QQTF 'FU -NWYGT #ECFGOKE 2WDNKUJGTU &QTVTGEJV 6JG 0GVJGTNCPFU RR Ō (GDTWCT[ =? ,4 $GNNGICTFC Fast Update of Latent Semantic Spaces Using a Linear Transform Framework KP 2TQE +PV %QPH #EQWUV 5RGGEJ 5KI 2TQE 1T NCPFQ (. /C[ =? ,4 $GNNGICTFC ,9 $WV\DGTIGT ;. %JQY 0$ %QEECTQ CPF & 0CKM A Novel Word Clustering Algorithm Based on Latent Semantic Analysis KP 2TQE +PV %QPH #EQWUV 5RGGEJ 5KI 2TQE #VNCPVC )# RR +Ō+ /C[ =? ,4 $GNNGICTFC CPF -'# 5KNXGTOCP Toward Unconstrained Command and Control: Data-Driven Semantic Inference KP 2TQE +PV %QPH 5RQMGP .CP IWCIG 2TQE $GKLKPI %JKPC RR +Ō+ 1EVQDGT =? ,4 $GNNGICTFC CPF -'# 5KNXGTOCP Natural Language Spoken Interface Control Using Data-Driven Semantic Inference +''' 6TCPU 5RGGEJ #WFKQ 2TQE KP RTGUU
E\&5&3UHVV//&
=? /9 $GTT[ Large–Scale Sparse Singular Value Computations +PV , 5WRGT EQOR #RRN 8QN 0Q RR Ō =? /9 $GTT[ 56 &WOCKU CPF )9 1ŏ$TKGP Using Linear Algebra for Intelligent Information retrieval 5+#/ 4GXKGY 8QN 0Q RR Ō =? / $GTT[ CPF # 5COGJ An Overview of Parallel Algorithms for the Singular Value and Dense Symmetric Eigenvalue Problems , %QORWVCVKQPCN #RRNKGF /CVJ 8QN RR Ō =? $ %CTRGPVGT CPF , %JWŌ%CTTQNN Natural Language Call Routing: A Robust, Self–Organized Approach KP 2TQE +PV %QPH 5RQMGP .CPIWCIG 2TQE 5[FPG[ #WUVTCNKC RR Ō &GEGODGT =? % %JGNDC & 'PING ( ,GNKPGM 8 ,KOGPG\ 5 -JWFCPRWT . /CPIW * 2TKPV\ '5 4KUVCF 4 4QUGPHGNF # 5VQNEMG CPF & 9W Structure and Performance of a Dependency Language Model KP 2TQE (KHVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN 4JQFGU )TGGEG 8QN RR Ō 5GRVGODGT =? % %JGNDC CPF ( ,GNKPGM Recognition Performance of a Structured Language Model KP 2TQE 5KZVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN $WFCRGUV *WP ICT[ 8QN RR Ō 5GRVGODGT =? 5 %JGP Building Probabilistic Models for Natural Language 2J& 6JGUKU *CTXCTF 7PKXGTUKV[ %CODTKFIG /# =? , %JWŌ%CTTQNN CPF $ %CTRGPVGT Dialog Management in Vector–Based Call Routing KP 2TQE %QPH #UUQE %QORWV .KPIWKUVKEU #%.%1.+0) /QPVTGCN %CPCFC RR Ō =? 24 %NCTMUQP CPF #, 4QDKPUQP Language Model Adaptation Using Mixtures and an Exponentially Decaying Cache KP 2TQE +PV %QPH #EQWUV 5RGGEJ 5KIPCN 2TQE /WPKEJ )GTOCP[ 8QN RR Ō /C[ =? 0 %QEECTQ CPF & ,WTCHUM[ Towards Better Integration of Semantic Predictors in Statistical Language Modeling KP 2TQE +PV %QPH 5RQMGP .CPIWCIG 2TQE 5[FPG[ #WUVTCNKC RR Ō &GEGODGT =? ,- %WNNWO CPF 4# 9KNNQWIJD[ Lanczos Algorithms for Large Symmetric Eigenvalue Computations – Vol. 1 Theory %JCRVGT 4GCN 4GEVCPIWNCT /CVTK EGU $TKEMJCWUGT $QUVQP /# =? 4 &G /QTK Recognizing and Using Knowledge Structures in Dialog Systems KP 2TQE #WV 5RGGEJ 4GEQI 7PFGTUVCPFKPI 9QTMUJQR -G[UVQPG %1 RR Ō &GEGODGT =? 5 &GGTYGUVGT 56 &WOCKU )9 (WTPCU 6- .CPFCWGT CPF 4 *CTUJOCP Indexing by Latent Semantic Analysis , #O 5QE +PHQTO 5EKGPEG 8QN RR Ō
E\&5&3UHVV//&
=? 5 &GNNC 2KGVTC 8 &GNNC 2KGVTC 4 /GTEGT CPF 5 4QWMQU Adaptive Language Model Estimation Using Minimum Discrimination Estimation KP 2TQE +PV %QPH #EQWUV 5RGGEJ 5KIPCN 2TQEGUUKPI 5CP (TCPEKUEQ %# 8QN + RR Ō #RTKN =? 56 &WOCKU Improving the Retrieval of Information from External Sources $GJCXKQT 4GU /GVJQFU +PUVTWO %QORWVGTU 8QN 0Q RR Ō =? 56 &WOCKU Latent Semantic Indexing (LSI) and TREC–2 KP 2TQE 5GEQPF 6GZV 4GVTKGXCN %QPHGTGPEG 64'%Ō & *CTOCP 'F 0+56 2WD Ō RR Ō =? / (GFGTKEQ CPF 4 &G /QTK Language Modeling %JCRVGT KP 5RQMGP &K CNQIWGU YKVJ %QORWVGTU 4 &G /QTK 'F #ECFGOKE 2TGUU .QPFQP 7- RR Ō =? 29 (QNV\ CPF 56 &WOCKU Personalized Information Delivery: An Analysis of Information Filtering Methods %QOOWP #%/ 8QN 0Q RR Ō =? 20 )CTPGT On Topic Identification and Dialogue Move Recognition %QO RWVGT 5RGGEJ CPF .CPIWCIG 8QN 0Q RR Ō =? & )KNFGC CPF 6 *QHOCPP Topic–Based Language Modeling Using EM KP 2TQE 5KZVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN $WFCRGUV *WPICT[ 8QN RR Ō 5GRVGODGT =? ) )QNWD CPF % 8CP .QCP Matrix Computations ,QJPU *QRMKPU $CNVKOQTG /& 5GEQPF 'F =? ; )QVQJ CPF 5 4GPCNU Document Space Models Using Latent Semantic Analysis KP 2TQE (KHVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN 4JQFGU )TGGEG 8QN RR Ō 5GRVGODGT =? 6 *QHOCPP Probabilistic Latent Semantic Analysis KP 2TQE (KHVGGPVJ %QPH 7PEGTVCKPV[ KP #+ 5VQEMJQNO 5YGFGP ,WN[ =? 6 *QHOCPP Probabilistic Topic Maps: Navigating Through Large Text Collections KP .GEVWTG 0QVGU %QOR 5EKGPEG 0Q RR Ō 5RTKPIGTŌ 8GTNCI *GKFGNDGTI )GTOCP[ ,WN[ =? 4 +[GT CPF / 1UVGPFQTH Modeling Long Distance Dependencies in Language: Topic Mixtures Versus Dynamic Cache Models +''' 6TCPU 5RGGEJ #WFKQ 2TQE 8QN 0Q ,CPWCT[ =? ( ,GNKPGM Self–Organized Language Modeling for Speech Recognition KP 4GCFKPIU KP 5RGGEJ 4GEQIPKVKQP # 9CKDGN CPF -( .GG 'FU /QTICP -CWHOCPP 2WDNKUJGTU RR Ō
E\&5&3UHVV//&
=? ( ,GNKPGM CPF % %JGNDC Putting Language into Language Modeling KP 2TQE 5KZVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN $WFCRGUV *WPICT[ 8QN RR -0Ō-0 5GRVGODGT =? & ,WTCHUM[ % 9QQVGTU , 5GICN # 5VQNEMG ' (QUNGT ) 6CLEJOCP CPF 0 /QTICP Using a Stochastic Context–Free Grammar as a Language Model for Speech Recognition KP 2TQE +PV %QPH #EQWUV 5RGGEJ 5KIPCN 2TQE &GVTQKV /+ 8QN + RR Ō /C[ =? 5 -JWFCPRWT Putting Language Back into Language Modeling RTGUGPVGF CV 9QTMUJQRŌ 5RQMGP .CPI 4GEQ 7PFGTUVCPFKPI 5WOOKV 0, (GDTWCT[ =? 4 -PGUGT Statistical Language Modeling Using a Variable Context KP 2TQE +PV %QPH 5RQMGP .CPIWCIG 2TQE RR Ō 2JKNCFGNRJKC 2# 1EVQDGT =? ( -WDCNC ,4 $GNNGICTFC ,4 %QJGP & 2CNNGVV &$ 2CWN / 2JKNNKRU 4 4CLCUGMCTCP ( 4KEJCTFUQP / 4KNG[ 4 4QUGPHGNF 4 4QVJ CPF / 9GKPVTCWD The Hub and Spoke Paradigm for CSR Evaluation KP 2TQE #42# 5RGGEJ CPF 0CVWTCN .CPIWCIG 9QTMUJQR /QTICP -CWHOCPP 2WDNKUJ GTU RR Ō /CTEJ =? 4 -WJP CPF 4 &G /QTK A Cache-based Natural Language Method for Speech Recognition +''' 6TCPU 2CVVGTP #PCN /CEJ +PVGN 8QN 2#/+Ō 0Q RR Ō ,WPG =? ,& .CHHGTV[ CPF $ 5WJO Cluster Expansion and Iterative Scaling for Maximum Entropy Language Models KP /CZKOWO 'PVTQR[ CPF $C[GUKCP /GVJQFU - *CPUQP CPF 4 5KNXGT 'FU -NWYGT #ECFGOKE 2WDNKUJGTU 0QTYGNN /# =? 6- .CPFCWGT CPF 56 &WOCKU Solution to Plato’s Problem: The Latent Semantic Analysis Theory of Acquisition, Induction, and Representation of Knowledge KP 2U[EJQNQIKECN 4GXKGY 8QN 0Q RR Ō =? 6- .CPFCWGT & .CJCO $ 4GJFGT CPF /' 5EJTGKPGT How Well Can Passage Meaning Be Derived Without Using Word Order: A Comparison of Latent Semantic Analysis and Humans KP 2TQE %QPH %QIPKV 5EKGPEG 5QE /CJYCJ 0, RR Ō =? 4 .CW 4 4QUGPHGNF CPF 5 4QWMQU Trigger–Based Language Models: A Maximum Entropy Approach KP 2TQE +PV %QPH #EQWUV 5RGGEJ 5KIPCN 2TQE /KPPGCRQNKU /0 RR ++Ō /C[ =? * 0G[ 7 'UUGP CPF 4 -PGUGT On Structuring Probabilistic Dependences in Stochastic Language Modeling %QORWVGT 5RGGEJ CPF .CPIWCIG 8QN RR Ō
E\&5&3UHVV//&
=? 6 0KGUNGT CPF 2 9QQFNCPF A Variable–Length Category–Based N–Gram Language Model KP 2TQE +PV %QPH #EQWUV 5RGGEJ 5KI 2TQE #VNCPVC )# RR +Ō+ /C[ =? %* 2CRCFKOKVTKQW 2 4CIJCXCP * 6COCMK CPF 5 8GORCNC Latent Semantic Indexing: A Probabilistic Analysis KP 2TQE VJ #%/ 5[OR 2TKPEKR &CVCDCUG 5[UV 5GCVVNG 9# #NUQ , %QOR 5[UV 5EKGPEGU =? (% 2GTGKTC ; 5KPIGT CPF 0 6KUJD[ Beyond Word -Grams %QORWVCVKQPCN .KPIWKUVKEU 8QN ,WPG =? .4 4CDKPGT $* ,WCPI CPF %* .GG An Overview of Automatic Speech Recognition %JCRVGT KP #WVQOCVKE 5RGGEJ CPF 5RGCMGT 4GEQIPKVKQP #F XCPEGF 6QRKEU %* .GG (- 5QQPI CPF -- 2CNKYCN 'FU -NWYGT #EC FGOKE 2WDNKUJGTU $QUVQP /# RR Ō =? 4 4QUGPHGNF The CMU Statistical Language Modeling Toolkit and its Use in the 1994 ARPA CSR Evaluation KP 2TQE #42# 5RGGEJ CPF 0CVWTCN .CPIWCIG 9QTMUJQR /QTICP -CWHOCPP 2WDNKUJGTU /CTEJ =? 4 4QUGPHGNF A Maximum Entropy Approach to Adaptive Statistical Language Modeling %QORWVGT 5RGGEJ CPF .CPIWCIG 8QN #ECFGOKE 2TGUU .QPFQP 7- RR Ō ,WN[ =? 4 4QUGPHGNF Two Decades of Statistical Language Modeling: Where Do We Go From Here 2TQE +''' 5RGE +UUWG 5RGGEJ 4GEQI 7PFGTUVCPFKPI $* ,WCPI CPF 5 (WTWK 'FU 8QN 0Q RR Ō #WIWUV =? 4 4QUGPHGNF . 9CUUGTOCP % %CK CPF :, <JW Interactive Feature Induction and Logistic Regression for Whole Sentence Exponential Language Models KP 2TQE #WV 5RGGEJ 4GEQI 7PFGTUVCPFKPI 9QTMUJQR -G[UVQPG %1 RR Ō &GEGODGT =? 5 4QWMQU Language Representation %JCRVGT KP 5WTXG[ QH VJG 5VCVG QH VJG #TV KP *WOCP .CPIWCIG 6GEJPQNQI[ 4 %QNG 'F %CODTKFIG 7PKXGTUKV[ 2TGUU %CODTKFIG /# =? 4 5EJYCTV\ 6 +OCK ( -WDCNC . 0IW[GP CPF , /CMJQWN A Maximum Likelihood Model for Topic Classification of Broadcast News KP 2TQE (KHVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN 4JQFGU )TGGEG 8QN RR Ō 5GRVGODGT =? 4' 5VQT[ An Explanation of the Effectiveness of Latent Semantic Indexing by Means of a Bayesian Regression Model +PHQTO 2TQEGUUKPI /CPCIGOGPV 8QN 0Q RR Ō =? , 9W CPF 5 -JWFCPRWT Combining Nonlocal, Syntactic and N-Gram Dependencies in Language Modeling KP 2TQE 5KZVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN $WFCRGUV *WPICT[ 8QN RR Ō 5GRVGODGT
E\&5&3UHVV//&
=? &* ;QWPIGT Recognition and Parsing of Context–Free Languages in Time ¿ +PHQTO %QPVTQN 8QN RR Ō
=? 4 <JCPI ' $NCEM CPF # (KPEJ Using Detailed Linguistic Structure in Language Modeling KP 2TQE 5KZVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN $W FCRGUV *WPICT[ 8QN RR Ō 5GRVGODGT =? :, <JW 5( %JGP CPF 4 4QUGPHGNF Linguistic Features for Whole Sentence Maximum Entropy Language Models KP 2TQE 5KZVJ 'WTQ %QPH 5RGGEJ %QOO 6GEJPQN $WFCRGUV *WPICT[ 8QN RR Ō 5GRVGODGT =? 8 <WG , )NCUU & )QQFKPG * .GWPI / 2JKNNKRU , 2QNKHTQPK CPF 5 5GPGHH Integration of Speech Recognition and Natural Language Processing in the MIT Voyager System KP 2TQE +''' +PV %QPH #EQWUV 5RGGEJ 5KIPCN 2TQEGUUKPI 6QTQPVQ %CPCFC RR Ō /C[
E\&5&3UHVV//&
10 Semantic Information Processing of Spoken Language – How May I Help You?sm A. L. Gorin, A. Abella, T. Alonso, G. Riccardi, and J. H. Wright, AT&T Laboratories
CONTENTS
+PVTQFWEVKQP %CNN%NCUUKſECVKQP .CPIWCIG /QFGNKPI HQT 4GEQIPKVKQP CPF 7PFGTUVCPFKPI &KCNQI %QPENWUKQPU 4GHGTGPEGU
10.1 Introduction 6JG PGZV IGPGTCVKQP QH XQKEGDCUGF WUGT KPVGTHCEG VGEJPQNQI[ GPCDNGU GCU[VQWUG CWVQOCVKQP QH PGY CPF GZKUVKPI EQOOWPKECVKQP UGTXKEGU CEJKGXKPI C OQTG PCVWTCN JWOCPOCEJKPG KPVGTCEVKQP $[ PCVWTCN YG OGCP VJCV VJG OCEJKPG WPFGTUVCPFU YJCV RGQRNG CEVWCNN[ UC[ KP EQPVTCUV VQ YJCV C U[UVGO FGUKIPGT GZRGEVU VJGO VQ UC[ 6JKU CRRTQCEJ KU KP EQPVTCUV YKVJ OGPWFTKXGP QT UVTQPIN[RTQORVGF U[UVGOU YJGTG OCP[ WUGTU CTG WPCDNG QT WPYKNNKPI VQ PCXKICVG UWEJ JKIJN[ UVTWEVWTGF KPVGTCEVKQPU AT&T’s ‘How May I Help You?’ */+*; VGEJPQNQI[ UJKHVU VJG DWTFGP HTQO JWOCP VQ OCEJKPG YJGTGKP VJG U[UVGO CFCRVU VQ RGQRNGUŏ NCPIWCIG CU EQPVTCUVGF YKVJ HQTEKPI WUGTU VQ NGCTP VJG OCEJKPGŏU LCTIQP 6JG IQCN QH UWEJ U[UVGOU KU VQ GZVTCEV OGCP KPI HTQO WUGTŏU PCVWTCN URQMGP NCPIWCIG +V KU KORQTVCPV VQ SWCPVKH[ VJKU PQVKQP UQ VJCV YG ECP OGCUWTG VJG ŎUGOCPVKE KPHQTOCVKQP EQPVGPVŏ QH C URQMGP WVVGTCPEG CPF HWTVJGTOQTG OGCUWTG QWT UWEEGUU KP GZVTCEVKPI VJCV KPHQTOCVKQP 5WEJ C VJGQT[ KU ETW EKCN VQ DGKPI CDNG VQ GPIKPGGT U[UVGOU VJCV WPFGTUVCPF CPF CEV WRQP URQMGP NCPIWCIG 6JG EQOOWPKECVKQP RCTCFKIO JGTG KPXQNXGU KPFWEKPI VJG OCEJKPG VQ RGTHQTO UQOG CEVKQP QT WPFGTIQ UQOG KPVGTPCN VTCPUHQTOCVKQP # EQOOWPKECVKQP KU FGGOGF UWE EGUUHWN KH VJG OCEJKPG TGURQPFU CRRTQRTKCVGN[ VQ VJG WUGTŏU KPRWV 6JKU KU KP EQPVTCUV VQ VJG VTCFKVKQPCN IQCN QH C EQOOWPKECVKQP U[UVGO YJKEJ YCU FGUETKDGF D[ 5JCPPQP =? CU HQNNQYU “The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another
E\&5&3UHVV//&
point. Frequently the messages have meaning, These semantic aspects of communication are irrelevant to the engineering problem.” +P VJKU YQTM VJG UGOCPVKE CURGEVU QH EQOOWPKECVKQP CTG RTKOCT[ +PHQTOCVKQP VJG QT[ ECP UVKNN DG GZRNQKVGF JQYGXGT VQ RTQXKFG C OGCUWTG QH UGOCPVKE KPHQTOCVKQP ECNNGF UCNKGPEG FGVCKNGF KP =? CPF YJQUG WVKNKV[ KU FGUETKDGF NCVGT KP VJKU CTVKENG 7P FGTUVCPFKPI WPEQPUVTCKPGF URGGEJ KU C FKHſEWNV RTQDNGO DQVJ HTQO VJG RGTURGEVKXG QH URGGEJ TGEQIPKVKQP CPF PCVWTCN NCPIWCIG WPFGTUVCPFKPI $QVJ VGEJPQNQIKGU CTG HCT HTQO RGTHGEV GURGEKCNN[ HQT EQPXGTUCVKQPCNUV[NG URGGEJ QXGT C VGNGRJQPG 6JG KPVWKVKQP WPFGTN[KPI QWT CRRTQCEJ KU VJCV HQT C IKXGP VCUM UQOG NKPIWKUVKE GXGPVU CTG ETWEKCN VQ TGEQIPK\G CPF WPFGTUVCPF QVJGTU PQV UQ 6JKU KPVWKVKQP KU SWCPVKſGF XKC UCNKGPEG CPF YG JCXG FGXGNQRGF CNIQTKVJOU = ? VJCV CWVQOCVKECNN[ NGCTP VJG UCNKGPV YQTFU RJTCUGU CPF ITCOOCT HTCIOGPVU HQT C VCUM +V ECP DG UJQYP GORKTKECNN[ VJCV VJGUG UCNKGPV HTCIOGPVU CTG TGEQIPK\GF HCT OQTG TGNKCDN[ VJCP CXGTCIG YJKEJ CN NQYU QWT */+*; VGEJPQNQI[ VQ YQTM GHHGEVKXGN[ # VQPIWGKPEJGGM EQOOGPV KU VJCV ŎYJQGXGTŏ FGUKIPGF PCVWTCN NCPIWCIG FKF C ŎIQQF LQDŏ CPF OCFG VJG UCNKGPV GXGPVU GCUKGT VQ TGEQIPK\G # FGUETKRVKQP QH GCTN[ NCDQTCVQT[ GZRGTKOGPVU DCUGF QP VJGUG KFGCU KU RTQXKFGF KP VJG VWVQTKCN RCRGT =? +P VJKU CTVKENG YG HQEWU QP VYQ VCUMU KPXQNXKPI NKXG EWUVQOGT VTCHſE KP #66ŏU PGVYQTM +P VJG QRGTCVQT UGTXKEGU FQOCKP VJG VCUM KPXQNXGU RNCEKPI VGNGRJQPG ECNNU URGEKH[KPI DKNNKPI OGVJQFU HQT VJQUG ECNNU
GI EQNNGEV ECNNKPI ECTF GVE CPF TGSWGUVKPI KPHQTOCVKQP CDQWV OCMKPI VJQUG ECNNU
GI TCVG CTGC EQFGU GVE 6JG UGEQPF CRRNKECVKQP KU VQ C EWUVQOGT ECTG VCUM +P VJKU FQOCKP WUGTU CUM SWGUVKQPU CDQWV KVGOU QP VJGKT DKNN VJGKT ECNNKPIRNCPU CEEQWPV DCNCPEGU GVE 6JG EWUVQOGT ECTG FQOCKP KU KPVWKVKXGN[ OQTG EQORNGZ FGVCKNU QH YJKEJ YKNN DG SWCPVKſGF NCVGT KP VJKU CTVKENG 6JG RTKOCT[ HQEWU QH URQMGP NCPIWCIG WPFGTUVCPFKPI 5.7 KP VJGUG FQOCKPU JCU DGGP ECNNV[RG ENCUUKſECVKQP KG FGVGTOKPKPI YJKEJ UGTXKEG V[RG C EWUVQOGT KU TGSWGUVKPI =? %NCUUKſECVKQP KU HQNNQYGF D[ TQWVKPI VJG ECNN VQ CP CRRTQRTKCVG FGU VKPCVKQP GKVJGT CP CWVQOCVGF OQFWNG YJGP CXCKNCDNG QT C JWOCP CIGPV YKVJ UQOG URGEKCNK\GF UMKNN UGV 9G JCXG CNUQ TGRQTVGF QP OGVJQFU HQT GZVTCEVKPI PCOGF GPVK VKGU UWEJ CU RJQPG CPF ETGFKVECTF PWODGTU GODGFFGF KP PCVWTCN URQMGP NCPIWCIG =? CPF HQT VTCPUNCVKQP KPVQ 5RCPKUJ CPF ,CRCPGUG =? *WOCPOCEJKPG KPVGTCEVKQPU TCTGN[ EQPUKUV QH C UKPING VWTP &KCNQI KU PGEGUUCT[ VQ confirm VJG OCEJKPGŏU WPFGT UVCPFKPI YJGP KVU EQPſFGPEG KU NQY VQ clarify CODKIWKVKGU KP C EWUVQOGTŏU TGSWGUV CPF VQ ICVJGT CFFKVKQPCN KPHQTOCVKQP PGEGUUCT[ VQ EQORNGVG VJG VCUM (QT GZCORNG KH UQOGQPG CUMU VQ OCMG C EQNNGEV ECNN HTQO C VTCKP UVCVKQP VJG #54 EQPſFGPEG YQWNF DG NQY DGECWUG QH VJG PQKU[ DCEMITQWPF UQ VJG OCEJKPG UJQWNF EQPſTO KVU TGEQI PKVKQP CPF WPFGTUVCPFKPI XKC “Do you want to make a collect call?” +H C WUGT CUMU “Charge this call please” VJGTG KU CP CODKIWKV[ YJKEJ PGGFU VQ DG ENCTKſGF HQT GZCORNG XKC “How do you want to charge this call, to a credit card or to a third number?” #P GZCORNG QH EQORNGVKQP QEEWTU YJGP UQOGQPG YCPVU VJGKT CEEQWPV DCNCPEG KP EWUVQOGT ECTG YJGPEG VJG OCEJKPG PGGFU VQ MPQY “What is your home phone number?” 6TCFKVKQPCNN[ HQT OGPWU[UVGOU CPF UVTQPIN[RTQORVGF FKCNQIU VJG JWOCPOCEJKPG KPVGTCEVKQP KU FGſPGF D[ C ŎECNNƀQYŏ GUUGPVKCNN[ C NQPI ŎKHVJGPGNUGŏ URGEKſECVKQP
E\&5&3UHVV//&
6JCV CRRTQCEJ FQGU PQV UECNG YGNN HQT EQORNGZ PCVWTCN URQMGP FKCNQIU [KGNFKPI UQHV YCTG VJCV KU FKHſEWNV VQ FGUKIP OCKPVCKP CPF UWRRQTV 1WT CRRTQCEJ VQ FKCNQI OCP CIGOGPV KU DCUGF QP C HTCOGYQTM ECNNGF VJG Construct Algebra =? 6JKU VJGQT[ RTQXKFGU VJG DWKNFKPI DNQEMU HQT VJG FKCNQI RTQEGUU EQORTKUKPI VJG TGNCVKQPU CPF QRGTCVKQPU QH VJCV CNIGDTC 6JG TGUWNV KU C EQNNGEVKQP QH TGWUCDNG dialog motivators IGPGTKE TWNGU VJCV FGVGTOKPG YJCV CEVKQP VJG FKCNQI OCPCIGT VCMGU KP KVU PGZV KPVGTCE VKQP YKVJ C WUGT CPF YJKEJ CTG RQTVCDNG QXGT C TCPIG QH VCUMU (QT GZCORNG VJGTG CTG TGWUCDNG FKCNQI OQVKXCVQTU HQT EQPſTOCVKQP ENCTKſECVKQP CPF OKUUKPI KPHQTOCVKQP =? +P CP[ FQOCKP VJGTG KU VCUM MPQYNGFIG VJCV OWUV DG GPEQFGF CPF RTQXKFGF VQ VJG &KCNQI /CPCIGT CPF 5.7 OQFWNGU +P */+*; VJKU VCUM MPQYNGFIG KU DCUGF QP CP QDLGEVQTKGPVGF inheritance hierarchy =? 6JKU KPJGTKVCPEG JKGTCTEJ[ FGſPGU VJG TGNCVKQPUJKRU COQPIUV VJG ECNNV[RGU CPF PCOGF GPVKVKGU (QT GZCORNG C EWUVQOGTŏU SWGT[ CDQWV CP WPTGEQIPK\GF EJCTIG QP VJGKT DKNN ‘is a’ MKPF QH SWGT[ CDQWV C EJCTIG QP VJGKT DKNN 6JCV WPTGEQIPK\GF PWODGT SWGT[ ‘has a’ FQNNCT COQWPV KVGO PWODGT FKCNGF PWODGT GVE +P EQORWVGT UEKGPEG KV KU YGNN MPQYP JQY VQ TGRTGUGPV VJGUG ŎKU Cŏ CPF ŎJCU Cŏ TGNCVKQPU XKC CP QDLGEVQTKGPVGF KPJGTKVCPEG JKGTCTEJ[ KP RTQITCOOKPI NCPIWCIGU UWEJ CU %
QT ,CXC 6JG FKCNQI OCPCIGT GZRNQKVU VJKU VCUM MPQYNGFIG CPF VJG FKCNQI OQVKXCVQTU VQ IQXGTP YJCV CEVKQP VQ RGTHQTO CV GCEJ VWTP KP VJG FKCNQI 6JKU EJCRVGT RTQEGGFU CU HQNNQYU (KTUV YG YKNN OQVKXCVG CPF FGUETKDG VJG ECNN ENCUUKſECVKQP RTQDNGO HQT CWVQOCVGF XQKEG UGTXKEGU 9G VJGP FGUETKDG VJG QRGTCVQT UGTXKEGU CPF EWUVQOGT ECTG VCUMU KP OQTG FGVCKN IKXKPI GZCORNGU QH YJCV RGQRNG UC[ CPF QH YJCV VJG[ YCPV 8CTKQWU FKOGPUKQPU QH VJG NKPIWKUVKE CPF UGOCPVKE EQORNGZ KV[ QH VJGUG VCUMU YKNN DG FGUETKDGF CPF SWCPVKſGF KPENWFKPI VJG KPJGTKVCPEG JKGTCT EJ[9G VJGP OQXG QP VQ FKUEWUU NCPIWCIG OQFGNKPI HQT DQVJ URGGEJ TGEQIPKVKQP CPF URQMGP NCPIWCIG WPFGTUVCPFKPI KPENWFKPI VJG KFGCU WPFGTN[KPI CWVQOCVGF CESWKUK VKQP CPF GZRNQKVCVKQP QH UCNKGPV YQTFU RJTCUGU CPF ITCOOCT HTCIOGPVU 6JG TQNG QH FKCNQI VQ IWKFG VJG JWOCPOCEJKPG KPVGTCEVKQP YKNN DG TGXKGYGF KPENWFKPI VJG EQP EGRV QH C FKCNQI VTCLGEVQT[ CPCN[UKU (KPCNN[ YG YKNN EQPENWFG CPF RTQXKFG RQKPVGTU VQ TGNCVGF TGUGCTEJ
10.2 Call-Classification +P VTCFKVKQPCN VGNGRJQP[ CWVQOCVKQP C WUGT KU QHHGTGF C NKUV QH OGPW QRVKQPU HTQO YJKEJ VQ UGNGEV +P UQOG ECUGU VJG FGUKTGF UGTXKEG ECP DG RTQXKFGF D[ UKORNG CW VQOCVKQP GI RTQXKFKPI CP CEEQWPV DCNCPEG QT DKNNKPI C ECNN VQ C ETGFKV ECTF +P QVJGT ECUGU VJG TGSWGUVGF UGTXKEG ECP QPN[ DG RTQXKFGF D[ C JWOCP CIGPV YKVJ UQOG URGEKCNK\GF UMKNN UGV +P GKVJGT ECUG VJG OGPW U[UVGO RTQXKFGU VJG WUGT VJG QRRQT VWPKV[ VQ PCXKICVG VQ VJG CRRTQRTKCVG FGUVKPCVKQP YJGTG JG QT UJG ECP QDVCKP UGTXKEG QT JCXG VJGKT RTQDNGO TGUQNXGF 6JGUG OGPWU JCXG DGGP KORNGOGPVGF WUKPI GKVJGT VQWEJVQPG ‘press one if you want x, press two if you want y’ QT XKC XQKEG NCDGNU
E\&5&3UHVV//&
‘please say collect, calling card’ 6JGTG KU CNUQ VJG HCOKNKCT J[DTKF ‘press or say one if you want x’ 'CEJ QH VJGUG JCXG VJGKT RNCEG CPF RWTRQUG CPF JCXG RTQXGF WUGHWN YJGP VJG NKUV QH QRVKQPU KU UJQTV CPF YGNNWPFGTUVQQF D[ EWUVQOGTU 9JGP VJG NKUV DGEQOGU NQPI JQYGXGT VJGP U[UVGO FGUKIPGTU TGUQTV VQ JKGTCTEJKECN OGPWU YJKEJ OCP[ WUGTU CTG WPCDNG QT WPYKNNKPI VQ PCXKICVG +P VJG ECUG QH UWEEKPEV OGPWQRVKQPU VJCV CTG NGUU URGEKſE KV KU QHVGP FKHſEWNV HQT C WUGT VQ FGEKFG YJKEJ QH VJG RTQHHGTGF ECVGIQTKGU OCVEJGU YJCV VJG[ YCPV 6JGTG KU CNYC[U VJG VTCFGQHH KP UWEJ OGPWU QH GZRNCKPKPI GCEJ QRVKQP KP ITGCV FGVCKN YJGPEG VJG WUGT DGEQOGU DQTGF CPF UVQRU NKUVGPKPI QT QH DGKPI UWEEKPEV KP VJG FGUETKRVKQP YJGPEG VJG WUGT ECPPQV ſIWTG QWV YJKEJ QRVKQP VQ UGNGEV +V KU C HCOKNKCT UEGPCTKQ HQT WUGTU VQ DGEQOG HTWUVTCVGF CPF GKVJGT RTGUU \GTQ VQ ŎDCKN QWVŏ QH UWEJ U[UVGOU QT VQ ŎRNC[ RQUUWOŏ CPF FQ PQVJKPI KP VJG JQRG QH DGKPI EQPPGEVGF VQ C JWOCP CIGPV +P EQPVTCUV EQPUKFGT JQY C JWOCP TGEGRVKQPKUV YQWNF JCPFNG VJKU UCOG TQWVKPI VCUM *G QT UJG YQWNF CUM ‘How may I help you?’ CPF VJGP VJG WUGT YQWNF FGUETKDG KP VJGKT QYP YQTFU YJCV KU VJGKT TGSWGUV QT RTQDNGO 6JG TGEGRVKQPKUVŏU LQD KU VQ MPQY GPQWIJ CDQWV VJG FQOCKP VQ VTCPUHGT VJG ECNNGT VQ UQOGDQF[ QT UQOGVJKPI VJCV ECP RTQXKFG VJG TGSWGUVGF UGTXKEG 6JWU YG UGV VJG IQCN QH IQKPI ŎDCEM VQ VJG HWVWTGŏ CPF GPIKPGGTKPI C U[UVGO YKVJ VJKU UCOG PCVWTCN HWPEVKQPCNKV[ # WUGT KU ITGGVGF CPF OCMGU C TGSWGUV CU KH VCNMKPI VQ C RGTUQP 6JG U[UVGOŏU LQD KU VQ TGEQIPK\G CPF WPFGTUVCPF YJCV VJG WUGT YCPVU UWH ſEKGPVN[ VQ TQWVG VJGKT ECNN VQ CP CWVQOCVGF OQFWNG QT JWOCP CIGPV VJCV ECP RTQXKFG VJG TGSWGUVGF UGTXKEG (QT GZCORNG YJGP YG EQNNGEVGF C FCVCDCUG QH YJCV EWUVQOGTU UC[ VQ QRGTCVQTU YG QDUGTXGF VJCV CNVJQWIJ VJG XCTKCVKQP KP XQECDWNCT[ CPF NCPIWCIG KU NCTIG OQUV QH VJG VKOG VJG[ CUM HQT QPG QH UGTXKEG V[RGU =? (QT GZCORNG “I want to reverse the charges on this call.” “Can you tell me what time it is in Tokyo?” “I was trying to call my sister and dialed a wrong number.” “I’ve been trying to dial this number all day and can’t get through.” 6JG ſTUV TGSWGUV KU HQT C %1..'%6 ECNN VJG UGEQPF HQT 6+/' KPHQTOCVKQP CPF VJG VJKTF HQT C $+..+0) %4'&+6 #WVQOCVKQP HQT VJGUG VJTGG ECNNV[RGU KU UVTCKIJVHQT YCTF 6JG ſPCN SWGT[ KU OQTG EQORNGZ CPF EWTTGPVN[ TGSWKTGU C RGTUQP VQ CFFTGUU VJG RTQDNGO +P CP[ ECNNENCUUKſECVKQP VCUM VJGTG KU CNYC[U C ŎVCKN QH VJG FKUVTKDWVKQPŏ YJKEJ FQGU PQV ſV KPVQ CP[ QH VJG RTGFGſPGF ECVGIQTKGU 9G ECNN VJKU ŎPQPG QH VJG CDQXGŏ ENCUU 16*'4 CPF UWEJ ECNNU CTG FKTGEVGF VQ C JWOCP CIGPV 6JG V[RGU QH SWGUVKQPU CUMGF KP EWUVQOGT ECTG CTG SWKVG FKHHGTGPV =? YJGTG RGQRNG CTG CUMKPI CDQWV VJGKT DKNNU ECNNKPIRNCPU GVE (QT GZCORNG “How much money do I owe you?” “I don’t recognize this phone call to Tallahassee on October 4.” “What’s this charge for one dollar and fifty cents?” “I have a question about my bill.”
E\&5&3UHVV//&
FIGURE 10.1 Call classification and routing in HMIHY. 6JG ſTUV SWGT[ KU HQT CP #%%1706 $#.#0%' CPF VJG UGEQPF HQT CP 704'% 1)0+<'& 07/$'4 CWVQOCVKQP HQT DQVJ QH YJKEJ KU UVTCKIJVHQTYCTF CPF GZKUVU VQFC[ 6JG VJKTF KU UQOGYJCV XCIWG CUMKPI CDQWV C %*#4)' 10 $+.. CU KU VJG HQWTVJ YJKEJ KU OGTGN[ C $+..+0) 37'4; (QT VJGUG NCUV VYQ GZCORNGU VJG FKCNQI OCPCIGT OWUV CUM C ENCTKH[KPI SWGUVKQP DGHQTG VJG ECNN ECP DG ENCUUKſGF CPF TQWVGF (KIWTG KNNWUVTCVGU VJG ECNNƀQY HQT ECNNTQWVKPI KP EWUVQOGT ECTG YJGTG VJG WUGT TGURQPFU VQ VJG QRGPGPFGF RTQORVU ‘How may I help you?’ CPF KU VJGP ENCUUKſGF CPF TQWVGF VQ CP CRRTQRTKCVG CWVQOCVGF OQFWNG QT JWOCP CIGPV VJCV ECP RTQXKFG VJG TGSWGUVGF UGTXKEG #U QDUGTXGF GCTNKGT VJG UGV QH UGOCPVKE NCDGNU KP UWEJ VCUMU KU PQV C UKORNG WPUVTWE VWTGF NKUV +P VJG GZCORNGU HTQO QRGTCVQT UGTXKEGU %1..'%6 CPF %4'&+6 %#4& CTG C MKPF QH $+..+0) OGVJQF CPF CP[ ECNN JCU C (149#4& 07/$'4 VJG PWO DGT DGKPI ECNNGF 5KOKNCTN[ TGSWGUVU HQT 4#6' 6+/' QT #4'# %1&' CTG CNN C MKPF QH TGSWGUV HQT +0(14/#6+10 6JGUG ŎKU Cŏ CPF ŎJCU Cŏ TGNCVKQPUJKRU CTG GPEQFGF KP CP QDLGEV QTKGPVGF KPJGTKVCPEG JKGTCTEJ[ RCTVKCNN[ KNNWUVTCVGF KP (KIWTG HQT VJG QRGTCVQT UGTXKEGU FQOCKP 6JG VGTOKPCN PQFGU KP VJG JKGTCTEJ[ RTQXKFG UWHſEKGPV URGEKſEKV[ HQT VJG OCEJKPG VQ ŎVCMG CEVKQPŏ YJKNG VJG PQPVGTOKPCN PQFGU TGSWKTG ENCTKH[KPI SWGTKGU HTQO VJG FKCNQI OCPCIGT +P EQORCTKPI VJG QRGTCVQT UGTXKEGU CPF EWUVQOGT ECTG FQOCKPU QWT KPVWKVKQP VGNNU WU VJCV VJG NCVVGT KU ŎOQTG EQORNGZŏ .GVŏU GZRNQTG JQY VQ SWCPVKH[ VJKU KPVWKVKQP # ſTUV QDUGTXCVKQP KU VJCV EWUVQOGT WVVGTCPEGU CTG UKIPKſECPVN[ NQPIGT KP VJG EWUVQOGT ECTG FQOCKP 9JGP TGURQPFKPI VQ ‘How may I help you?’ VJG CXGTCIG PWODGT QH YQTFU KP VJG ſTUV FQOCKP KU YJKNG KV KU YQTFU KP VJG UGEQPF # JKUVQITCO QH WVVGTCPEG NGPIVJ KU UJQYP HQT DQVJ FQOCKPU =? KP (KIWTG 1DUGTXG VJCV VJG ŎUJCRGŏ QH VJG VYQ FKUVTKDWVKQPU CTG UKOKNCT UMGYGF WPKOQFCN YKVJ C NQPI VCKN # UGEQPF FKOGPUKQP QH NKPIWKUVKE EQORNGZKV[ KU XQECDWNCT[ (QT C TCPFQO UCORNG QH
E\&5&3UHVV//&
*/+*;
+PHQTOCVKQP
&KCN(QT/G
$KNNKPI%TGFKV
#TGC%QFG
1VJGT
$KNNKPI/GVJQF
4CVG %QNNGEV
%CNNKPI%CTF
6JKTFA0WODGT
%CTFA0WO
$KNNGFA0WO
KUŦC JCUŦC
(QTYCTFA0WO
FIGURE 10.2 Inheritance hierarchy of task knowledge in operator services. - WVVGTCPEGU VJG CEEWOWNCVGF XQECDWNCTKGU KP VJG VYQ FQOCKPU CTG - CPF TGURGEVKXGN[ 6JG 118 QWV QH XQECDWNCT[ TCVG QH QDUGTXGF PGY YQTFU KP DQVJ ECUGU KU CRRTQZKOCVGN[ QPG PGY YQTF GXGT[ VJKTF WVVGTCPEG 9G QDUGTXGF VJCV VJGUG 118 YQTFU CTG NGUU VJCP JCNH RTQRGT PQWPU GORJCUK\KPI VJG JKIJ XCTKCVKQP KP EWUVQOGTUŏ NCPIWCIG HQT VJGUG FQOCKPU # ſPCN CPF VTCFKVKQPCN OGCUWTG QH NKPIWKUVKE EQORNGZKV[ KU perplexity =? YJKEJ ECP DG NQQUGN[ KPVGTRTGVGF CU VJG ŎCXGTCIG DTCPEJKPI HCEVQTŏ QH VJG NCPIWCIG 6JGUG CTG CPF TGURGEVKXGN[ CICKP KNNWUVTCVKPI VJG ITGCVGT EQORNGZKV[ QH VJG EWUVQOGT ECTG FQOCKP 1PG ECP CNUQ OGCUWTG VJG UGOCPVKE EQORNGZKV[ QH VJG ŎENCUUKſECVKQP VCUMŏ 6JG GPVTQR[ QH VJKU ECNNV[RG FKUVTKDWVKQP ECP DG EQORWVGF CU DKVU RGT UGOCPVKE NC DGN 6JKU RTQXKFGU KPUKIJV KPVQ YJ[ VJG ŎENCUUKſECVKQPŏ RTQDNGO KU VTCEVCDNG HTQO EQPXGTUCVKQPCNUV[NG URGGEJ QXGT VJG VGNGRJQPG YJKNG #54 TGOCKPU FKHſEWNV # RGT RNGZKV[ QH KU GSWKXCNGPV VQ CP GPVTQR[ QH DKVU RGT YQTF QT DKVU KP C YQTF WVVGTCPEG VJCV YG CTG CVVGORVKPI VQ FGEQFG XKC #54 (TQO VJG ENCUUKſECVKQP RGTURGEVKXG JQYGXGT YG CTG QPN[ UGGMKPI VQ TGNKCDN[ FGEQFG DKVU RGT WVVGTCPEG #NVJQWIJ VJKU KU PQV C TKIQTQWU CTIWOGPV KV KPFKECVGU YJ[ ECNNENCUUKſECVKQP KU RQUUK DNG YKVJ JKIJ CEEWTCE[ YJKNG #54 KU HCT OQTG FKHſEWNV (QT GZCORNG YQTF CEEWTCE[ HQT VJGUG VCUMU KU CEEWTCE[ QP UCNKGPV RJTCUGU KU OWEJ JKIJGT CV CPF ENCUUKſECVKQP CEEWTCE[ YGNN GZEGGFU Evaluating Call Classification. %CNN ENCUUKſECVKQP ECP DG XKGYGF CU C OWNVKENCUU ENCUUKſECVKQP VCUM YKVJ TGLGEVKQP 6JGTG CTG VJTGG VTCFKVKQPCN OGCUWTGU HQT UWEJ VCUMU
E\&5&3UHVV//&
TGNCVKXGHTGSWGPE[
%WUVQOGT%CTG
OGCP
YQTFURGTWVVGTCPEG
TGNCVKXGHTGSWGPE[
1RGTCVQT5GTXKEGU
OGCP
YQTFURGTWVVGTCPEG
FIGURE 10.3 Histogram of utterance lengths. =? (KTUV KU VJG RTQDCDKNKV[ QH false rejection YJKEJ OGCUWTGU JQY QHVGP C TGSWGUV HQT UQOG UGTXKEG KU TGLGEVGF QT ENCUUKſGF CU 16*'4 # UGEQPF KU VJG RTQDCDKNKV[ QH correct classification YJKEJ OGCUWTGU JQY QHVGP C ENCUUKſECVKQP CU UQOG ECNNV[RG KU EQTTGEV 6JG VJKTF KU VJG true rejection rate YJKEJ OGCUWTGU VJG RTQDCDKNKV[ VJCV C TGSWGUV YJKEJ UJQWNF DG ENCUUKſGF CU 16*'4 KU KPFGGF TGLGEVGF CPF VJWU TQWVGF VQ C JWOCP &KCNQI RTQXKFGU VJG QRRQTVWPKV[ VQ CUM EQPſTOKPI CPF ENCTKH[KPI SWGUVKQPU VJWU RTQXKFKPI KORTQXGF ECNNENCUUKſECVKQP QXGT VJG ECUG QH C UKPING WVVGTCPEG 9JKNG HCT HTQO RGTHGEV RGTHQTOCPEG GZEGGFKPI JCU DGGP TGRQTVGF KP QWT RWDNKUJGF RCRGTU YJKEJ KU HCT UWRGTKQT VQ EWUVQOGTUŏ CDKNKV[ VQ UGNHUGNGEV CPF PCXKICVG JKGTCT EJKECN OGPWU 6JWU */+*; RTQXKFGU DQVJ CP KORTQXGF WUGT GZRGTKGPEG RNWU OQTG CEEWTCVG TQWVKPI CPF VJWU KPETGCUGF CWVQOCVKQP Remark: 6JGTG KU XCUV NKVGTCVWTG KP VGZV ECVGIQTK\CVKQP HQT VJG RWTRQUG QH KPHQTOC VKQP CPF FQEWOGPV TGVTKGXCN 6JGTG KU CNUQ C NKVGTCVWTG QP VQRKE ENCUUKſECVKQP HTQO URGGEJ HQT UKOKNCT RWTRQUGU %CNNENCUUKſECVKQP CU FGUETKDGF KP VJKU CPF TGNCVGF YQTM JCU UGXGTCN FKUVKPIWKUJKPI CVVTKDWVGU (KTUV KV KPXQNXGU URGGEJ TCVJGT VJCP VGZV YKVJ VJG KPJGTGPV FKHſEWNVKGU QH URGGEJ TGEQIPKVKQP CPF VJG FKUƀWGPEKGU QH EQPXGTUCVKQPCN UV[NG NCPIWCIG 5GEQPF VJG KPRWV KU HTQO EQQRGTCVKXG WUGTU YJQ CTG VT[KPI VQ EQOOW PKECVG VJGKT PGGF CPF OCMG VJGOUGNXGU WPFGTUVQQF 6JKTF VJG U[UVGO JCU VJG QRRQT VWPKV[ VQ CUM EQPſTOKPI QT ENCTKH[KPI SWGUVKQPU QH VJCV EQQRGTCVKXG WUGT (QWTVJ VJGTG
E\&5&3UHVV//&
KU QHVGP EQNNCVGTCN EWUVQOGT RTQſNG KPHQTOCVKQP CXCKNCDNG YJKEJ ECP DG GZRNQKVGF KP WPFGTUVCPFKPI C TGSWGUV (QT GZCORNG C EWUVQOGT YJQ UC[U “I want to know how to pay my bill” YQWNF DG TQWVGF FKHHGTGPVN[ KH VJG[ JCXG DGGP FGNKPSWGPV KP RC[OGPV XGTUWU C TQWVKPG TGSWGUV
10.3 Language Modeling for Recognition and Understanding (QT TGEQIPK\KPI WPEQPUVTCKPGF URQMGP NCPIWCIG VJG UVCVGQHVJGCTV KPXQNXGU VTCKPKPI C UVQEJCUVKE NCPIWCIG OQFGN YJKEJ RTGFKEVU VJG RTQDCDKNKV[ QH C UGSWGPEG QH YQTFU (QT GZCORNG IKXGP C UGPVGPEG ½ ¾ YG YCPV VQ GUVKOCVG VJG RTQDCDKN KV[ QH VJG YQTF IKXGP VJG JKUVQT[ QH RTGEGFKPI YQTFU KG ½ ¾ ½ +V KU PQV VTCEVCDNG VQ GUVKOCVG VJGUG RTQDCDKNKVKGU HQT CNN RQUUKDNG JKUVQTKGU FWG VQ FCVC URCTUGPGUU 5Q VJG OQUV HCOKNKCT OGVJQF KU VJG PITCO OQFGN YJKEJ GUVK OCVGU VJG RTQDCDKNKV[ QH C YQTF IKXGP QPN[ VJG RTGEGFKPI P YQTFU YJGTG V[RKECNN[ P FGPQVGF C VTKITCO NCPIWCIG OQFGN #U P KPETGCUGU VJG OGOQT[ CPF EQORW VCVKQP TGSWKTGOGPVU QH VJG NCPIWCIG OQFGN KPETGCUG CU FQGU FCVC URCTUGPGUU #P CNVGTPCVKXG KU VQ UGNGEVKXGN[ KPVTQFWEG NQPIGTTCPIG ŎJKUVQT[ŏ KP VJG HQTO QH XCTKCDNG NGPIVJ WPKVU =? (QT #54 NCPIWCIG OQFGNKPI VJGUG WPKVU CTG UGNGEVGF QP VJG DC UKU QH GPVTQR[ OKPKOK\CVKQP NGCFKPI VQ ŎUWRGTYQTFUŏ UWEJ CU ‘I want to make a’ ‘collect call’ CPF ‘card call’ KP VJG QRGTCVQT UGTXKEGU FQOCKP +P VJCV ECUG C DK ITCO NCPIWCIG OQFGN QP VJGUG XCTKCDNG NGPIVJ WPKVU YQWNF NGCF VQ GUVKOCVGU UWEJ CU ŎEQNNGEV ECNNŏ Ŏ+ YCPV VQ OCMG Cŏ GHHGEVKXGN[ RTQXKFKPI C ITCO OQFGN DWV QPN[ UGNGEVKXGN[ +P =? KV YCU UJQYP JQY VQ GODGF VJGUG CESWKTGF RJTCUGU KPVQ C UVQEJCUVKE ſPKVG UVCVG CWVQOCVQP VQ RTQXKFG NCPIWCIG OQFGNU HQT #54 +V YCU UJQYP VJCV VJGUG OQFGNU JCF C UKOKNCT CEEWTCE[ VQ JKIJQTFGT PITCO OQFGNU DWV YKVJ VJG EQORWVCVKQP CPF OGOQT[ TGSWKTGOGPVU QH NQYQTFGT OQFGNU #HVGT TGEQIPK\KPI VJG YQTFU URQMGP D[ C WUGT VJG PGZV UVGR KU VQ ŎWPFGTUVCPFŏ YJCV VJG[ UCKF 1WT GCTN[ GZRGTKOGPVU HQEWUGF QP OGVJQFU DCUGF QP C ŎDCI QH YQTFUŏ OQFGN =? 9G FKUEQXGTGF JQYGXGT VJCV KIPQTKPI VJG VGORQTCN QTFGT QH VJG YQTFU KP CP WVVGTCPEG YCU PQV QRVKOCN .CPIWCIG FQGU KPFGGF JCXG UVTWEVWTG YJKEJ ECP DG GZRNQKVGF VQ GPCDNG OQTG TGNKCDNG WPFGTUVCPFKPI 6JG ſTUV UVCIG YCU VQ KPXGUVKICVG CPF FGXGNQR CNIQTKVJOU VQ CWVQOCVKECNN[ CESWKTG UCNKGPV RJTCUGU HQT C VCUM KG YJKNG ‘wrong’ KU C UCNKGPV YQTF KP VJG QRGTCVQT UGTXKEGU VCUM CUUQEKCVGF VQ PGGFKPI C DKNNKPI ETGFKV ‘wrong number’ KU GXGP OQTG UCNKGPV CPF ‘dialed a wrong number’ OQTG UCNKGPV UVKNN =? 5CNKGPV RJTCUGU CTG RTGHGTCDNG VQ YQTFU DGECWUG VJG[ JCXG UJCTRGT UGOCPVKEU CPF DGECWUG NQPIGT GXGPVU CTG OQTG TGNKCDN[ TGEQIPK\GF KP URGGEJ 6JG UGEQPF UVCIG EQOOGPEGF YKVJ QDUGTXKPI VJCV OCP[ QH VJGUG UCNKGPV RJTCUGU YGTG UKOKNCT UWEJ CU ‘dialed a wrong number’ CPF ‘dialed the wrong number’ NGCFKPI VQ VJG FGXGNQROGPV QH ENWUVGTKPI CNIQTKVJOU GZRNQKVKPI C EQODKPCVKQP QH UVTKPIGFKV FKU VCPEGU CPF UGOCPVKE FKUVQTVKQPU =? 6JGUG ENWUVGTU QH UCNKGPV RJTCUGU CTG EQORCEVN[
E\&5&3UHVV//&
PQV
FKF ECNNU
+
FKFPŏV
YG
FQPŏV
OCMG
TGEQIPK\G
WPFGTUVCPF
FIGURE 10.4 A salient grammar fragment.
TGRTGUGPVGF CU ſPKVG UVCVG OCEJKPGU CPF FGPQVGF salient grammar fragments %NWU VGTU QH RJTCUGU CTG RTGHGTCDNG DGECWUG QH RCTUKOQP[ GPCDNKPI RQQNKPI QH UVCVKUVKEU CETQUU OWNVKRNG NQY HTGSWGPE[ RJTCUGU 6JGUG ITCOOCT HTCIOGPVU CTG CNUQ CFXCPVC IGQWU DGECWUG VJG[ CTG TQDWUV VQ #54 GTTQTU ŎYKVJKP VJG HTCIOGPVŏ +P (KIWTG C UCNKGPV ITCOOCT HTCIOGPV HTQO VJG EWUVQOGT ECTG VCUM KU UJQYP YJKEJ KU UVTQPIN[ CUUQEKCVGF YKVJ SWGTKGU CDQWV CP 704'%1)0+<'& 07/$'4 QP C DKNN 6JGUG OGVJQFU YGTG GZVGPFGF VQ KPENWFG JKGTCTEJKECN ENWUVGTKPI D[ CNUQ GZRNQKVKPI U[PVCEVKE FKUVQTVKQPU =? (KPCNN[ YG QDUGTXG VJCV GODGFFKPI UCNKGPV RJTCUGU KP VJG #54 NCPIWCIG OQFGN JCU DGGP UJQYP VQ KORTQXG ENCUUKſ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
¯ User QMC[ + IQV #66 9KTGNGUU RJQPGU CPF YJGP + IQV VJGO JG VQNF OG VJCV + YQWNF DG UYKVEJGF VQ EGPVU C OKPWVG HQT CNN O[ #66 NQPI FKUVCPEG ECNNKPI DGECWUG + YCU QP EGPVU 1PG 4CVG RNCP ¯
yeah I’m not AT&T WIRELESS PHONE and when I got and she told
me that I would be switched to 7 CENTS A MINUTES FOR ALL my AT&T long distance on that I was on 10 10 cents ONE RATE PLAN
¯ SLU %CNNKPI 2NCPU
E\&5&3UHVV//&
2NC[ RTQORV
7UGT URGGEJ
5RQMGP .CPIWCIG 7PFGTUVCPFKPI
#54
#EQWUVKE /QFGNU
.CPIWCIG /QFGNU
5CNKGPV )TCOOCT (TCIOGPVU
&KCNQI /CPCIGT
+PJGTKVCPEG *KGTCTEJ[
FIGURE 10.5 Natural spoken dialog in HMIHY.
10.4 Dialog #P KPVGTCEVKQP YKVJ */+*; QHVGP EQORTKUGU OQTG VJCP QPG VWTP YKVJ FKCNQI PGEGU UCT[ VQ EQPſTO NQYEQPſ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ſTUV KNNWUVTCVKPI VYQ KPXQECVKQPU QH VJG ŎOKUUKPI KPHQTOCVKQPŏ OQVKXCVQT VQ QDVCKP VJG DKNNKPI OGVJQF CPF ECTF PWODGT VJG UGEQPF KNNWUVTCVKPI VJG ŎENCTKſECVKQPŏ OQVKXCVQT
E\&5&3UHVV//&
Machine: #66 *QY OC[ + JGNR [QW! User: + PGGF C ETGFKV + IQV VJG YTQPI PWODGT Machine: 9JCV YCU VJG PWODGT VJCV [QW FKCNGF! User: + FKCNGF HQWT QPG ſXG VYQ GKIJV Machine: 9CU VJG ECNN DKNNGF VQ VJG RJQPG VJCV [QWŏTG ECNNKPI HTQO PQY! User: 0Q VQ C ETGFKV ECTF Machine: 9JCV YCU VJG ECTF PWODGT! User: HQWT ſXG UKZ UGXGP Machine: +ŏXG IKXGP [QW ETGFKV HQT VJCV ECNN 6JCPM [QW HQT WUKPI #66
Machine: #66 *QY OC[ + JGNR [QW! User: + JCXG C SWGUVKQP CDQWV O[ ,WPG DKNN Machine: 1MC[ YJCV KU [QWT SWGUVKQP! User: 6JGTGŏU C ECNN VQ 6CNNCJCUUGG DWV + FQPŏV MPQY CP[QPG VJGTG Machine: + ECP JGNR [QW YKVJ VJCV #TG [QW ECNNKPI HTQO [QWT JQOG RJQPG! User: +ŏO ECNNKPI HTQO O[ JQWUG Machine: 2NGCUG JQNF QP YJKNG + VTCPUHGT [QWT ECNN
10.5 Conclusions +P UWOOCT[ YG JCXG FGUETKDGF #66ŏU */+*; VGEJPQNQI[ YJKEJ GODQFKGU VJG PGZV IGPGTCVKQP QH XQKEGGPCDNGF UGTXKEGU 6JGUG CFXCPEGU KP URGGEJ NCPIWCIG CPF FKCNQI VGEJPQNQI[ UJKHV VJG DWTFGP HTQO WUGT VQ OCEJKPG YJGTG VJG OCEJKPG CFCRVU VQ EWUVQOGTUŏ URQMGP NCPIWCIG KP EQPVTCUV VQ HQTEKPI RGQRNG VQ NGCTP VJG OCEJKPGŏU LCT IQP +P FGXGNQRKPI UWEJ C ŎſTUV QH KVU MKPFŏ U[UVGO OCP[ TGUGCTEJ KUUWGU JCXG CTKUGP 6JG KPVGTGUVGF TGCFGT ECP CEEGUU QWT YGD UKVG www.research.att.com/algor/hmihy YJKEJ EQPVCKPU C NKPM VQ CP QPNKPG EQNNGEVKQP QH OCP[ QH QWT TGUGCTEJ RCRGTU
References =? # #DGNNC CPF #. )QTKP ő)GPGTCVKPI 5GOCPVKECNN[ %QPUKUVCPV +PRWVU VQ C &KCNQI /CPCIGTŒ 2TQE 'WTQURGGEJ )TGGEG RR 5GRV =? # #DGNNC CPF # )QTKP ő%QPUVTWEV #NIGDTC #PCN[VKECN &KCNQI /CPCIGOGPVŒ 2TQE #%. 9CUJKPIVQP &% ,WPG
E\&5&3UHVV//&
=? - #TCK ,* 9TKIJV ) 4KEECTFK CPF # )QTKP ő)TCOOCT (TCIOGPV #ESWKUK VKQP 7UKPI 5[PVCEVKE CPF 5GOCPVKE %NWUVGTKPIŒ 5RGGEJ %QOOWPKECVKQP XQN ,CP =? 5 $CPICNQTG CPF ) 4KEECTFK ő5VQEJCUVKE (KPKVG5VCVG /QFGNU HQT 5RQMGP .CP IWCIG /CEJKPG 6TCPUNCVKQPŒ 2TQE 0##%. 5GCVVNG 9# /C[ =? #. )QTKP ő1P #WVQOCVGF .CPIWCIG #ESWKUKVKQPŒ RR ,QWTPCN QH VJG #EQWUVKECN 5QEKGV[ QH #OGTKEC ,WPG =? #. )QTKP ) 4KEECTFK CPF ,* 9TKIJV ő*QY /C[ + *GNR ;QW!Œ 5RGGEJ %QOOWPKECVKQP XQN RR =? #. )QTKP , * 9TKIJV ) 4KEECTFK # #DGNNC CPF 6 #NQPUQ ő5GOCPVKE +PHQTOCVKQP 2TQEGUUKPI QH 5RQMGP .CPIWCIGŒ 2TQE #64 9QTMUJQR QP /WN VKNKPIWCN 5RGGEJ %QOOWPKECVKQP -[QVQ ,CRCP 1EV =? / 4CJKO ) 4KEECTFK . 5CWN , 9TKIJV $ $WPVUEJWJ CPF # )QTKP ő4Q DWUV 0WOGTKE 4GEQIPKVKQP KP 5RQMGP .CPIWCIG &KCNQIŒ 5RGGEJ %QOOWPKEC VKQP XQN RR =? ) 4KEECTFK 4 2KGTCEEKPK CPF ' $QEEJKGTK ő5VQEJCUVKE #WVQOCVC HQT .CP IWCIG /QFGNKPIŒ %QORWVGT 5RGGEJ CPF .CPIWCIG RR =? ) 4KEECTFK CPF #. )QTKP ő5RQMGP .CPIWCIG #FCRVCVKQP QXGT 6KOG CPF 5VCVG KP C 0CVWTCN 5RQMGP &KCNQI 5[UVGOŒ +''' 6TCPU QP 5RGGEJ CPF #WFKQ XQN RR ,CP =? % 5JCPPQP ő# /CVJGOCVKECN 6JGQT[ QH %QOOWPKECVKQPŒ $GNN 5[UVGO 6GEJ PKECN ,QWTPCN 8QN ::8++ 0Q RR ,WN[ =? ,* 9TKIJV #. )QTKP CPF ) 4KEECTFK ő#WVQOCVKE #ESWKUKVKQP QH 5CNKGPV )TCOOCT (TCIOGPVU HQT %CNN6[RG %NCUUKſECVKQPŒ 2TQE 'WTQURGGEJ )TGGEG
E\&5&3UHVV//&
11 Machine Translation Using Statistical Modeling Herman Ney, and F. J. Och Aachen University of Technology, Germany
CONTENTS
+PVTQFWEVKQP 5VCVKUVKECN &GEKUKQP 6JGQT[ CPF .KPIWKUVKEU #NKIPOGPV CPF .GZKEQP /QFGNU #NKIPOGPV 6GORNCVGU (TQO 5KPING 9QTFU VQ 9QTF )TQWRU 'ZRGTKOGPVCN 4GUWNVU 5RGGEJ 6TCPUNCVKQP 6JG +PVGITCVGF #RRTQCEJ 5WOOCT[ 4GHGTGPEGU
Abstract. 6JKU EJCRVGT IKXGU CP QXGTXKGY QH VJG UVCVKUVKECN CRRTQCEJ VQ OCEJKPG VTCPUNCVKQP KP RCTVKEWNCT VJG VTCPUNCVKQP QH URQMGP FKCNQIWGU KP VJG HTCOGYQTM QH VJG 8 '4$/1$+. RTQLGEV 5VCTVKPI YKVJ VJG $C[GU FGEKUKQP TWNG CU KP URGGEJ TGEQIPK VKQP YG UJQY JQY VJG TGSWKTGF RTQDCDKNKV[ FKUVTKDWVKQPU ECP DG UVTWEVWTGF KPVQ VJTGG RCTVU VJG NCPIWCIG OQFGN VJG CNKIPOGPV OQFGN CPF VJG NGZKEQP OQFGN 9G FGUETKDG VJG EQORQPGPVU QH VJG U[UVGO CPF TGRQTV TGUWNVU QP VJG 8 '4$/1$+. VCUM 6JG GZ RGTKGPEG QDVCKPGF KP VJG 8 '4$/1$+. RTQLGEV KP RCTVKEWNCT KP VJG ſPCN GXCNWCVKQP UJQYGF VJCV VJG UVCVKUVKECN CRRTQCEJ TGUWNVGF KP UKIPKſECPVN[ NQYGT GTTQT TCVGU VJCP VJTGG EQORGVKPI VTCPUNCVKQP CRRTQCEJGU VJG UGPVGPEG GTTQT TCVG YCU KP EQORCT KUQP YKVJ VQ HQT VJG QVJGT VTCPUNCVKQP CRRTQCEJGU (KPCNN[ YG FKUEWUU VJG KPVGITCVGF CRRTQCEJ VQ URGGEJ VTCPUNCVKQP CU QRRQUGF VQ VJG UGTKCN CRRTQCEJ CU KV KU YKFGN[ WUGF PQYCFC[U
11.1 Introduction 6JG CWVQOCVKE VTCPUNCVKQP QH NCPIWCIG KU IGPGTCNN[ TGHGTTGF VQ CU machine translation 6[RKECNN[ VJKU VGTO KU WUGF HQT written language QT text KPRWV YJGTG VJG KORNKEKV CUUWORVKQP KU VJCV VJG KPRWV KU WPEQTTWRVGF KG YKVJQWV GTTQTU 6JKU VCUM KU XGT[ OWEJ FKHHGTGPV HTQO spoken speech KPRWV YJGTG VJG U[UVGO OWUV EQRG YKVJ URGGEJ
E\&5&3UHVV//&
TGEQIPKVKQP GTTQTU CPF CNUQ VJG WPITCOOCVKECN UVTWEVWTG QH URQMGP NCPIWCIG 6JG VTCPUNCVKQP QH spontaneous speech RQUGU CFFKVKQPCN FKHſEWNVKGU HQT VJG VCUM QH CWVQOCVKE VTCPUNCVKQP 6[RKECNN[ VJGUG FKHſ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
¯ 5GEVKQP Statistical Decision Theory and Linguistics. 9G YKNN RTGUGPV VJG $C[GU FGEKUKQP TWNG CPF VJG TGUWNVKPI CTEJKVGEVWTG HQT VJG VTCPUNCVKQP QH YTKVVGP NCPIWCIG ¯ 5GEVKQP Alignment and Lexicon Models. # MG[ EQORQPGPV KP VJG UVCVKUVKECN CRRTQCEJ KU VJG UQECNNGF CNKIPOGPV EQPEGRV YJKEJ KU UKOKNCT VQ JKFFGP /CTMQX OQFGNU WUGF KP URGGEJ TGEQIPKVKQP CPF YJKEJ YKNN DG EQPUKFGTGF KP OQTG FGVCKN ¯ 5GEVKQP Alignment Templates: From Single Word to Word Groups. 6Q KPVTQFWEG OQTG EQPVGZV KPVQ VJG VTCPUNCVKQP RTQEGUU YG YKNN EQPUKFGT VJG OGVJQF QH CNKIPOGPV VGORNCVGU VJCV CNNQYU WU VQ VTCPUNCVG YQTF ITQWRU QT RJTCUGU CU C YJQNG ¯ 5GEVKQP Experimental Results #NVJQWIJ VJG OGVJQFU RTGUGPVGF CRRN[ DQVJ VQ written CPF spoken NCPIWCIG YG YKNN NKOKV QWTUGNXGU JGTG VQ URQMGP NCPIWCIG CPF TGRQTV QP VJG ſPCN GZRGTK OGPVCN GXCNWCVKQP VJCV YGTG ECTTKGF QWV KP VJG 8 '4$/1$+. RTQLGEV ¯ 5GEVKQP Speech Translation: The Integrated Approach. #U CP CNVGTPCVKXG VQ VJG serial EQWRNKPI QH TGEQIPKVKQP CPF VTCPUNCVKQP VJCV KU WUGF KP QWT CPF QVJGT U[UVGOU CU YGNN YG YKNN EQPUKFGT VJG integrated CR RTQCEJ VQ TGEQIPKVKQP CPF VTCPUNCVKQP CPF VJG EQTTGURQPFKPI HQTO QH VJG $C[GU FGEKUKQP TWNG =? 9JGTGCU UVCVKUVKECN OQFGNNKPI KU YKFGN[ WUGF KP URGGEJ TGEQIPKVKQP CPF KV KU KORQU UKDNG VQ GPWOGTCVG CNN U[UVGOU VJGTG UGGO VQ DG QPN[ C UOCNN PWODGT QH TGUGCTEJ ITQWRU VJCV JCXG CRRNKGF UVCVKUVKECN OQFGNNKPI VQ VJG VTCPUNCVKQP QH YTKVVGP QT URQMGP NCPIWCIG = ? 6JG RTGUGPVCVKQP JGTG KU DCUGF QP YQTM ECTTKGF QWV KP VJG HTCOGYQTM QH VJG ' 7 6 4#05 RTQLGEV =? CPF VJG 8 '4$/1$+. RTQLGEV =?
E\&5&3UHVV//&
11.2 Statistical Decision Theory and Linguistics 11.2.1 The Statistical Approach 6JG WUG QH UVCVKUVKEU KP EQORWVCVKQPCN NKPIWKUVKEU JCU DGGP GZVTGOGN[ EQPVTQXGTUKCN HQT OQTG VJCP VJTGG FGECFGU 6JG EQPVTQXGTU[ KU XGT[ YGNN UWOOCTK\GF D[ VJG UVCVGOGPV QH %JQOUM[ KP =? ő KV OWUV DG TGEQIPK\GF VJCV VJG PQVKQP ŎRTQDCDKNKV[ QH C UGPVGPEGŏ KU CP GPVKTGN[ WUGNGUU QPG WPFGT CP[ MPQYP KPVGTRTGVCVKQP QH VJKU VGTOŒ 6JKU UVCVGOGPV YCU EQPUKFGTGF VQ DG VTWG D[ VJG OCLQTKV[ QH GZRGTVU HTQO CTVKſEKCN KPVGNNKIGPEG CPF EQORWVCVKQPCN NKPIWKUVKEU CPF VJG EQPEGRV QH UVCVKUVKEU YCU DCPPGF HTQO EQORWVCVKQPCN NKPIWKUVKEU HQT OCP[ [GCTU 9JCV KU QXGTNQQMGF KP VJKU UVCVGOGPV KU VJG HCEV VJCV KP CP CWVQOCVKE U[UVGO HQT URGGEJ TGEQIPKVKQP QT VGZV VTCPUNCVKQP YG CTG HCEGF YKVJ VJG RTQDNGO QH OCMKPI FGEKUKQPU +V KU GZCEVN[ JGTG YJGTG UVCVKUVKECN FGEKUKQP VJGQT[ EQOGU KP +P URGGEJ TGEQIPKVKQP VJG UWEEGUU QH VJG UVCVKUVKECN CRRTQCEJ KU DCUGF QP VJG GSWCVKQP 5RGGEJ 4GEQIPKVKQP
#EQWUVKEŌ.KPIWKUVKE /QFGNNKPI
5VCVKUVKECN &GEKUKQP 6JGQT[
5KOKNCTN[ HQT OCEJKPG VTCPUNCVKQP VJG UVCVKUVKECN CRRTQCEJ KU GZRTGUUGF D[ VJG GSWC VKQP /CEJKPG 6TCPUNCVKQP
.KPIWKUVKE /QFGNNKPI 5VCVKUVKECN &GEKUKQP 6JGQT[
(QT VJG ŎNQYNGXGNŏ FGUETKRVKQP QH URGGEJ CPF KOCIG UKIPCNU KV KU YKFGN[ CEEGRVGF VJCV VJG UVCVKUVKECN HTCOGYQTM CNNQYU CP GHſEKGPV EQWRNKPI DGVYGGP VJG QDUGTXCVKQPU CPF VJG OQFGNU YJKEJ KU QHVGP FGUETKDGF D[ VJG DW\\ YQTF ŎUWDU[ODQNKE RTQEGUUKPIŏ $WV VJGTG KU CPQVJGT CFXCPVCIG KP WUKPI RTQDCDKNKV[ FKUVTKDWVKQPU KP VJCV VJG[ QHHGT CP GZRNKEKV HQTOCNKUO HQT GZRTGUUKPI CPF EQODKPKPI J[RQVJGUKU UEQTGU
¯ 6JG RTQDCDKNKVKGU CTG FKTGEVN[ WUGF CU UEQTGU 6JGUG UEQTGU CTG PQTOCNK\GF YJKEJ KU C FGUKTCDNG RTQRGTV[ YJGP KPETGCUKPI VJG UEQTG HQT C EGTVCKP GNGOGPV KP VJG UGV QH CNN J[RQVJGUGU VJGTG OWUV DG QPG QT UGXGTCN QVJGT GNGOGPVU YJQUG UEQTGU CTG TGFWEGF CV VJG UCOG VKOG ¯ +V KU UVTCKIJVHQTYCTF VQ EQODKPG UEQTGU &GRGPFKPI QP VJG VCUM VJG RTQDCDKNK VKGU CTG GKVJGT OWNVKRNKGF QT CFFGF ¯ 9GCM CPF XCIWG FGRGPFGPEGU ECP DG OQFGNNGF GCUKN[ 'URGEKCNN[ KP URQMGP CPF YTKVVGP PCVWTCN NCPIWCIG VJGTG CTG PWCPEGU CPF UJCFGU VJCV TGSWKTG ŎITG[ NGXGNUŏ DGVYGGP CPF
E\&5&3UHVV//&
11.2.2 Bayes Decision Rule for Written Language Translation +P OCEJKPG VTCPUNCVKQP HQT YTKVVGP NCPIWCIG VJG IQCN KU VJG VTCPUNCVKQP QH C VGZV IKXGP KP C UQWTEG NCPIWCIG KPVQ C VCTIGV NCPIWCIG 9G CTG IKXGP C UQWTEG UVTKPI ½ ½ YJKEJ KU VQ DG VTCPUNCVGF KPVQ C VCTIGV UVTKPI ½ ½ (QT JKUVQTKECN TGCUQPU =? YG WUG VJG U[ODQNU NKMG (TGPEJ HQT UQWTEG YQTFU CPF VJG U[ODQN NKMG 'PINKUJ HQT VCTIGV YQTFU +P VJKU EJCRVGT VJG VGTO word CNYC[U TGHGTU VQ C full-form YQTF #OQPI CNN RQUUKDNG VCTIGV UVTKPIU YG YKNN EJQQUG VJG UVTKPI YKVJ VJG JKIJGUV RTQDCDKNKV[ YJKEJ KU IKXGP D[ $C[GU FGEKUKQP TWNG =? ½
½ ½
½
½ ½ ½
½
*GTG ½ KU VJG NCPIWCIG OQFGN QH VJG VCTIGV NCPIWCIG CPF ½ ½ KU VJG UVTKPI VTCPUNCVKQP OQFGN YJKEJ YKNN DG FGEQORQUGF KPVQ NGZKEQP CPF CNKIPOGPV OQF GNU 6JG CTIOCZ QRGTCVKQP FGPQVGU VJG UGCTEJ RTQDNGO KG VJG IGPGTCVKQP QH VJG QWVRWV UGPVGPEG KP VJG VCTIGV NCPIWCIG 6JG QXGTCNN CTEJKVGEVWTG QH VJG UVCVKUVKECN VTCPUNCVKQP CRRTQCEJ KU UWOOCTK\GF KP (KIWTG +P IGPGTCN CU UJQYP KP VJKU ſIWTG VJGTG OC[ DG CFFKVKQPCN VTCPUHQTOCVKQPU VQ OCMG VJG VTCPUNCVKQP VCUM UKORNGT HQT VJG CNIQTKVJO 6JG VTCPUHQTOCVKQPU OC[ TCPIG HTQO VJG ECVGIQTK\CVKQP QH UKPING YQTFU CPF YQTF ITQWRU VQ OQTG EQORNGZ RTGRTQEGUUKPI UVGRU VJCV TGSWKTG UQOG RCTUKPI QH VJG UQWTEG UVTKPI 9G JCXG VQ MGGR KP OKPF VJCV KP VJG UGCTEJ RTQEGFWTG DQVJ VJG NCPIWCIG CPF VJG VTCPUNCVKQP OQFGN CTG CRRNKGF after VJG VGZV VTCPUHQTOCVKQP UVGRU *QYGXGT VQ MGGR VJG PQVCVKQP UKORNG YG YKNN PQV OCMG VJKU GZRNKEKV FKUVKPEVKQP KP VJG UWDUGSWGPV GZRQUKVKQP
11.2.3 Related Approaches 6JGTG CTG C PWODGT QH TGNCVGF CRRTQCEJGU VJCV CTG CNUQ EQTRWUDCUGF CPF VJGTGHQTG ENQUGN[ TGNCVGF VQ VJG UVCVKUVKECN CRRTQCEJ
ſPKVGUVCVG CRRTQCEJGU = ? *GTG VJG RTQDCDKNKUVKE FGRGPFGPEGU CTG TGRTGUGPVGF D[ ſPKVGUVCVG UVTWEVWTGU VJCV ECP DG NGCTPGF CWVQOCVKECNN[ HTQO VTCKPKPI FCVC GZCORNGDCUGF CRRTQCEJGU = ? +P GZCORNGDCUGF CRRTQCEJGU NCTIG DKNKPIWCN EJWPMU CTG GZEKUGF HTQO VJG UGV QH DKNKPIWCN UGPVGPEG RCKTU +P VJG VTCPUNCVKQP RTQEGUU VJG OQUV UKOKNCT EJWPM KP VJG UGV QH UQWTEGNCPIWCIG EJWPMU KU FGVGTOKPGF CPF KVU EQTTGURQPF KPI VCTIGVNCPIWCIG EJWPM KU WUGF CU VTCPUNCVKQP 6JKU DCUGNKPG XCTKCPV OC[ DG TGſPGF KP XCTKQWU YC[U VQ KPVTQFWEG IGPGTCNK\CVKQP ECRCDKNKVKGU UGG CNUQ 5GEVKQP U[PVCZDCUGF UVCVKUVKECN CRRTQCEJGU = ? 6JGUG CRRTQCEJGU CTG QDVCKPGF CU CP GZVGPUKQP QH VJG UVCVKUVKECN CRRTQCEJ
E\&5&3UHVV//&
Source Language Text
Transformation ,
H
Global Search: G + maximize 2T
, ^ G + 2T H
Alignment Model
, ^ G + 2T H
G + over
Lexicon Model
G + 2T
Language Model
Transformation
Target Language Text
FIGURE 11.1 Architecture of the translation approach based on Bayes decision rule. YJGTG U[PVCEVKE UVTWEVWTGU CTG KPEQTRQTCVGF KPVQ VJG DCUGNKPG UVCVKUVKECN CR RTQCEJ KP RCTVKEWNCT VJG UQECNNGF CNKIPOGPV OQFGNU UGG NCVGT 6JG U[PVCEVKE UVTWEVWTG OC[ DG OQFGNNGF KP VJG VCTIGV NCPIWCIG QPN[ QT KP DQVJ VCTIGV CPF UQWTEG NCPIWCIG
11.3 Alignment and Lexicon Models 11.3.1 Concept of Alignment Modelling # MG[ KUUWG KP OQFGNNKPI VJG UVTKPI VTCPUNCVKQP RTQDCDKNKV[ ½ ½ KU VJG SWGUVKQP QH JQY YG FGſPG VJG EQTTGURQPFGPEG DGVYGGP VJG YQTFU QH VJG VCTIGV UGPVGPEG CPF VJG YQTFU QH VJG UQWTEG UGPVGPEG +P V[RKECN ECUGU YG ECP CUUWOG C UQTV QH RCKTYKUG FGRGPFGPEG D[ EQPUKFGTKPI CNN YQTF RCKTU HQT C IKXGP UGPVGPEG RCKT ½ ½
E\&5&3UHVV//&
ja ich denke wenn wir das hinkriegen an beiden Tagen acht Uhr
days both on eight at it make can we if think I well
FIGURE 11.2 Example of an alignment for a German-English sentence pair.
*GTG YG YKNN HWTVJGT EQPUVTCKP VJKU OQFGN D[ CUUKIPKPI GCEJ UQWTEG YQTF VQ exactly one VCTIGV YQTF /QFGNU FGUETKDKPI VJGUG V[RGU QH FGRGPFGPEGU CTG TGHGTTGF VQ CU alignment models = ? 9JGP CNKIPKPI VJG YQTFU KP RCTCNNGN VGZVU YG V[RKECNN[ QDUGTXG C UVTQPI NQECNK\CVKQP GHHGEV (KIWTG KNNWUVTCVGU VJKU GHHGEV C )GTOCPŌ'PINKUJ UGPVGPEG RCKT HTQO VJG 8 '4$/1$+. EQTRWU +P OCP[ ECUGU CNVJQWIJ PQV CNYC[U VJGTG KU CP CFFKVKQPCN RTQRGTV[ QXGT NCTIG RQTVKQPU QH VJG UQWTEG UVTKPI VJG CNKIPOGPV KU OQPQVQPG +P VJG HQNNQYKPI YG YKNN EQPUKFGT VYQ CRRTQCEJGU VQ CNKIPOGPV OQFGNNKPI KP OQTG FGVCKN PCOGN[ JKFFGP /CTMQX OQFGNU CPF OQFGNU +$/ Ō
11.3.2 Hidden Markov Models 6JG ſTUV CRRTQCEJ VQ CNKIPOGPV OQFGNNKPI YKNN DG DCUGF QP JKFFGP /CTMQX OQFGNU
*// CU VJG[ JCXG DGGP WUGF UWEEGUUHWNN[ KP URGGEJ TGEQIPKVKQP HQT C NQPI VKOG = EJCRVGT ? = EJCRVGT ? 6JWU VJG CNKIOGPV OCRRKPI KP VTCPUNCVKQP KU UKOKNCT VQ VJG VKOG CNKIPOGPV RCVJ QT UVCVG UGSWGPEG KP URGGEJ TGEQIPKVKQP
E\&5&3UHVV//&
6Q CTTKXG CV C SWCPVKVCVKXG URGEKſECVKQP YG ſTUV FGſPG VJG CNKIPOGPV OCRRKPI
YJKEJ CUUKIPU C YQTF KP RQUKVKQP VQ C YQTF KP RQUKVKQP 6JG IGPGTCN EQPEGRV QH YQTF CNKIPOGPVU YCU KPVTQFWEGF KP =? 7UKPI VJG UCOG DCUKE RTKPEKRNGU CU KP *//U HQT URGGEJ TGEQIPKVKQP YG ECP TGYTKVG VJG RTQDCDKNKV[ D[ KPVTQFWEKPI VJG ŎJKFFGPŏ CNKIPOGPVU ½ ½ HQT GCEJ UGPVGPEG RCKT ½ ½
½ ½
½ ½ ½
½
6Q ENCTKH[ VJG OGCPKPI QH VJG VGTO ‘hidden’ KP EQORCTKUQP YKVJ URGGEJ TGEQIPKVKQP YG PQVG VJCV VJG OQFGN UVCVGU CU UWEJ TGRTGUGPVKPI YQTFU CTG not JKFFGP DWV VJG CEVWCN CNKIPOGPVU KG VJG sequence QH RQUKVKQP KPFGZ RCKTU YKVJ 6Q FTCY VJG CPCNQI[ YKVJ URGGEJ TGEQIPKVKQP YG JCXG VQ KFGPVKH[ VJG UVCVGU CNQPI VJG XGTVKECN CZKU YKVJ VJG RQUKVKQPU QH VJG VCTIGV YQTFU CPF VJG VKOG CNQPI VJG JQTK\QPVCN CZKU YKVJ VJG RQUKVKQPU QH VJG UQWTEG YQTFU 9G ECP FGEQORQUG VJG RTQDCDKNKV[ FKUVTKDWVKQP ½ ½ ½ CU HQNNQYU
½ ½ ½ ½
½ ½ ½
½
6JG CDQXG HQTOWNCVKQP FQGU PQV OCMG CP[ CUUWORVKQPU CDQWV VJG FGRGPFGPEGU KP VJG RTQDCDKNKV[ FKUVTKDWVKQP CPF TGUWNVU KP VJTGG FKUVTKDWVKQPU YJKEJ PGGF HWTVJGT URGEKſ ECVKQPU VJG NGPIVJ OQFGN VJG CNKIPOGPV OQFGN CPF VJG NGZKEQP OQFGN 6JGUG OQFGNU CTG VQQ IGPGTCN VQ DG WUGF FKTGEVN[ CPF KP VJG HQNNQYKPI YG YKNN NKOKV VJG FGRGPFGPEGU KP VJGUG OQFGNU GI YG YKNN CUUWOG ſTUVQTFGT QXGT GXGP \GTQQTFGT FGRGPFGPEGU HQT VJG EQPFKVKQPCN RTQDCDKNKV[ 9G EQPUKFGT VJTGG ECUGU
baseline HMM
$[ NQQMKPI CV TGCN CNKIPOGPVU HQT UGPVGPEG RCKTU KV KU GXKFGPV VJCV VJG OCVJG OCVKECN OQFGN UJQWNF VT[ VQ ECRVWTG VJG UVTQPI FGRGPFGPEG QH QP VJG RTGEGF KPI CNKIPOGPV 6JGTGHQTG YJGP UKORNKH[KPI VJG FGRGPFGPEGU KP VJG CNKIPOGPV OQFGN YG YQWNF NKMG VQ TGVCKP VJG FGRGPFGPEG QH QP VJG RQUKVKQP QH VJG KOOGFKCVG RTGFGEGUUQT 6JWU YG QDVCKP VJG CNKIPOGPV OQFGN
E\&5&3UHVV//&
YJGTG YG CNUQ JCXG TGVCKPGF VJG FGRGPFGPEG QP VJG NGPIVJ QH VJG QDUGTXGF UQWTEG UGPVGPEG CPF VJG NGPIVJ QH VJG J[RQVJGUK\GF VCTIGV UGPVGPEG (QT VJG NGZKEQP OQFGN YG OCMG VJG CUUWORVKQP VJCV VJG FGRGPFGPEG KU NKOKVGF VQ VJG VCTIGV YQTF YKVJ KG CPF PQVJKPI GNUG
½
½
½ ½
(KPCNN[ HQT VJG NGPIVJ OQFGN YG CUUWOG C FGRGPFGPEG QP VJG NGPIVJ QH VJG UQWTEG UGPVGPEG ½ QPN[
½ 9G OGPVKQP VJCV VJG NGPIVJ OQFGN JCU DGGP KPENWFGF HQT VJG UCMG QH EQO RNGVGPGUU CPF KU PQV XGT[ KORQTVCPV KP RTCEVKEG +P URGGEJ TGEQIPKVKQP VJGTG KU V[RKECNN[ PQ NGPIVJ OQFGN +PUVGCF C URGEKCN U[ODQN HQT UGPVGPEG GPF KU CFFGF VQ VJG XQECDWNCT[
homogeneous HMM 6Q TGPFGT VJG CNKIPOGPV RTQDCDKNKV[ KPFGRGPFGPV QH CDUQNWVG RQUKVKQPU CPF CNUQ VQ TGFWEG VJG PWODGT QH CNKIPOGPV RCTCOGVGTU = ? YG CUUWOG VJCV VJG CNKIPOGPV RTQDCDKNKVKGU ½ FGRGPF QPN[ QP VJG LWOR YKFVJ ½ CPF PQVJKPI GNUG 7UKPI CPF ¼ ½ YG JCXG
½ ½ ½ ½ ½
½ ½
YKVJ C PQPPGICVKXG VCDNG
¼
YJKEJ JCU VQ DG GUVKOCVGF HTQO VJG DKNKPIWCN VTCKPKPI EQTRWU NKMG VJG HTGG RCTCOGVGTU QH VJG QVJGT FKUVTKDWVKQPU KPVTQFWEGF
context dependent HMM +V ECP DG CTIWGF VJCV HQT IQQF OQFGNU OQTG EQPVGZV UJQWNF DG ECRVWTGF KP VJG FGRGPFGPEGU 6JWU YG GZVGPF VJG CNKIPOGPV OQFGN
5Q KP EQORCTKUQP YKVJ VJG DCUGNKPG OQFGN VJGTG KU CP CFFKVKQPCN FGRGPFGPEG QP VJG UQWTEG YQTF KP RQUKVKQP (KTUV GZRGTKOGPVU YKVJ UWEJ C V[RG QH OQFGN CTG TGRQTVGF KP =?
E\&5&3UHVV//&
6JG NGZKEQP OQFGN ECP DG GZVGPFGF KP C UKOKNCT YC[
½ ½ ½ ½ ½ ½ *GTG VJG FGRGPFGPEGU JCXG DGGP GZVGPFGF VQ ½ CPF ½ +PUVGCF QH VJGUG CFFKVKQPCN YQTFU VJGOUGNXGU RCTVUQHURGGEJ ENCUUGU QT CWVQOCVKECNN[ VTCKPGF YQTF ENCUUGU EQWNF DG WUGF UGG NCVGT = ? 5WEJ C V[RG QH GZVGPFGF NGZKEQP OQFGN FQGU PQV UGGO VQ JCXG DGGP VGUVGF [GV GZRGTKOGPVCNN[
11.3.3 Models IBM 1–5 6JG JKUVQTKECN FGXGNQROGPV QH UVCVKUVKECN OCEJKPG VTCPUNCVKQP YCU UNKIJVN[ FKHHGTGPV HTQO VJKU RTGUGPVCVKQP KP VJCV VJG OQFGNU +$/ Ō YGTG KPVTQFWEGF HQT CNKIPOGPV OQFGNNKPI before *//U YGTG WUGF 6JG OQFGNU +$/ Ō YGTG KPVTQFWEGF KP =? CU C UGTKGU QH CNKIPOGPV OQFGNU YKVJ KPETGCUKPI EQORNGZKV[
¯ models IBM-1 and IBM-2: zero-oder dependence. 4CVJGT VJCP C first-order FGRGPFGPEG YG ECP CNUQ WUG C zero-order OQFGN HQT VJG CNKIPOGPV OQFGN YJGTG VJGTG KU QPN[ C FGRGPFGPEG QP VJG absolute RQUK VKQP KPFGZ QH VJG UQWTEG UVTKPI
½ ½ ½ ½ ½ $QVJ VJG NGPIVJ OQFGN CPF VJG NGZKEQP OQFGN CTG VJG UCOG CU HQT VJG *// (QT UWEJ C \GTQQTFGT OQFGN KV ECP DG UJQYP =? VJCV YG JCXG VJG HQNNQYKPI KFGPVKV[
½ ½
½
6JG UWO KP VJG NCUV GSWCVKQP ECP DG KPVGTRTGVGF CU C OKZVWTGV[RG FKUVTKDWVKQP YKVJ OKZVWTG YGKIJVU CU CNKIPOGPV RTQDCDKNKVKGU CPF YKVJ EQORQ PGPV FKUVTKDWVKQPU CU NGZKEQP RTQDCDKNKVKGU 6JG OQFGN +$/ KU C URGEKCN ECUG YKVJ C WPKHQTO CNKIPOGPV RTQDCDKNKV[
6JG RTGUGPVCVKQP UQ HCT JCU PQV WUGF VJG UQECNNGF ŎGORV[ YQTFŏ =? 6JG GORV[ YQTF KU CFFGF VQ VJG VCTIGV UGPVGPEG VQ CNNQY HQT UQWTEG YQTFU YJKEJ JCXG PQ FKTGEV EQWPVGTRCTV KP VJG VCTIGV UGPVGPEG (QTOCNN[ VJG EQPEGRV QH VJG GORV[ YQTF KU KPEQTRQTCVGF KPVQ VJG CNKIPOGPV OQFGNU D[ CFFKPI VJG GORV[ YQTF CV RQUKVKQP VQ VJG VCTIGV UGPVGPEG CPF CNKIPKPI CNN UQWTEG YQTFU YKVJQWV C FKTGEV VTCPUNCVKQP VQ VJKU GORV[ YQTF
E\&5&3UHVV//&
¯ model IBM-3: fertility concept. #U KPVTQFWEGF KP =? VJG CNKIPOGPV OQFGN ECP DG GZVGPFGF D[ VJG EQPEGRV QH HGTVKNKV[ 6JG KFGC KU VJCV QHVGP C YQTF KP VJG VCTIGV NCPIWCIG OC[ DG CNKIPGF VQ UGXGTCN YQTFU KP VJG UQWTEG NCPIWCIG 6JKU GZVGPUKQP TGUWNVU KP VJG UQECNNGF OQFGN +$/ (QT GCEJ VCTIGV YQTF VJGTG KU C RTQDCDKNKV[ FKUVTKDWVKQP QXGT KVU RQUUKDNG HGTVKNKVKGU
'ZRGTKOGPVCNN[ YG QDUGTXG VJCV VJG HGTVKNKVKGU VCMG QP XCNWGU HTQO VQ (QT C IKXGP CNKIPOGPV ½ YG EQORWVG VJG HGTVKNKV[ QH C VCTIGV YQTF KP RQUKVKQP CU VJG PWODGT QH CNKIPGF UQWTEG YQTFU YKVJ HGTVKNKV[
Æ
7UKPI VJKU GSWCVKQP YG ECP UVCTV YKVJ CP *// QT OQFGN +$/ CPF VJGP EQORWVG KPKVKCN XCNWGU HQT VJG HGTVKNKVKGU +P RCTVKEWNCT VJG HGTVKNKV[ EQPEGRV ECP DG WUGF VQ DGVVGT OQFGN VCTIGV YQTFU JCXKPI PQ EQWPVGTRCTV KP VJG UQWTEG UGPVGPEG KG VCTIGV YQTFU YKVJ HGTVKNKV[
models IBM-4 and IBM-5: inverted alignments with first-order dependence. (QT URCEG NKOKVCVKQPU YG ECP IKXG QPN[ C UKORNKſGF FGUETKRVKQP QH VJGUG OQF GNU 6Q QDVCKP VJGUG OQFGNU YG CUUWOG VJCV VJG RTQDCDKNKV[ FKUVTKDWVKQP ½ ½ ½ KU VJG TGUWNV QH C RTQEGUU EQPUKUVKPI QH VJTGG UVGRU GCEJ QH YJKEJ KPXQNXGU C UKORNG RTQDCDKNKV[ FKUVTKDWVKQP 6JG ſTUV UVGR KU VJG UGNGEVKQP QH C HGTVKNKV[ HQT GCEJ J[RQVJGUK\GF VCTIGV YQTF
+P VJG PGZV UVGR HQT GCEJ VCTIGV YQTF YG IGPGTCVG VJG UGV QH CUUQEKCVGF UQWTEG YQTFU CEEQTFKPI VQ VJG HGTVKNKV[ YJGTG VJG ſPCN RQUKVKQPU CTG PQV URGEKſGF [GV +P VJG VJKTF UVGR VJG UQWTEG YQTFU CTG RGTOWVGF UQ VJCV VJG QDUGTXGF UGSWGPEG ½ KU RTQFWEGF 6JG OCKP CFXCPVCIG QH VJG CDQXG KPVGTRTGVCVKQP KU VJCV CU YG YKNN UGG NCVGT KV KU DGVVGT UWKVGF HQT C UGCTEJ UVTCVGI[ VJCV DWKNFU WR RCTVKCN UVTKPI J[RQVJGUGU ½ QXGT VCTIGV RQUKVKQPU
#U C TGUWNV YG JCXG C UQTV QH KPXGTVGF CNKIPOGPVU KG C OCRRKPI HTQO VJG VCTIGV RQUKVKQPU VQ VJG UQWTEG RQUKVKQPU KPXGTVGF CNKIPOGPV OCRRKPI
YJKEJ KP =? KU TGHGTTGF VQ CU FKUVQTVKQP OQFGN (QT VJGUG KPXGTVGF CNKIPOGPVU
½ ½
YG CUUWOG C ſTUVQTFGT FGRGPFGPEG CU HQT VJG *//
½ ½
*GTG VJGTG KU CP CFFKVKQPCN FGRGPFGPEG QP VJG YQTF EQPVGZV VJCV KU ECRVWTGF D[ VJG UQWTEG YQTF KP RQUKVKQP CPF VJG VCTIGV YQTF ½ 6Q TGCNN[ CRRN[
E\&5&3UHVV//&
VJG CDQXG RTQDCDKNKV[ OQFGN UGXGTCN TGſPGOGPVU CTG PGGFGF (KTUV YG OWUV VCMG KPVQ CEEQWPV VJCV VJG HGTVKNKV[ QH YQTF KP RQUKVKQP OC[ DG FKHHGTGPV HTQO GI HQT C HGTVKNKV[ NCTIGT VJCP UGXGTCN RQUKVKQPU QP VJG VCTIGV CZKU JCXG VQ DG RTQFWEGF 5GEQPF VJG FGRGPFGPEG QP ½ FQGU PQV WUG VJG absolute RQUKVKQPU DWV QPN[ relative RQUKVKQPU 6JWU YG JCXG C FGRGPFGPEG QP VJG ŏLWOR YKFVJŏ ½ CNQPI VJG UQWTEG CZKU CU HQT VJG JQOQIGPGQWU *// CNQPI VJG VCTIGV CZKU 6JKTF VQ TGFWEG VJG PWODGT QH HTGG RCTCOGVGTU VJG FGRGPFGPEG QP VJG YQTFU CPF ½ KU TGRNCEGF D[ C FGRGPFGPEG QP VJG EQTTGURQPFKPI RCTVUQHURGGEJ QT YQTF ENCUUGU = ? CPF ½
½ ½
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½ ½ (QT C XCECPV RQUKVKQP YG JCXG
½ ½
6Q DG OQTG GZCEV YG PQVG VJCV VJG FGRGPFGPEG QP ½ ½ KU OCKPN[ NKOKVGF VQ VJG PWODGT QH HTGG UQWTEG RQUKVKQPU ½ ½ CPF VQ VJG PWODGT QH HTGG UQWTEG RQUK VKQPU DGVYGGP ½ CPF +P EQORCTKUQP YKVJ OQFGN +$/ VJG FGRGPFGPEG QP VJG RTGEGFKPI VCTIGV YQTF ½ JCU DGGP FTQRRGF VQ TGFWEG VJG PWODGT QH HTGG RCTCOGVGTU
#NVJQWIJ UQOG QH VJG CDQXG OQFGNU VCMG QPGVQOCP[ CNKIPOGPVU GZRNKEKVN[ KPVQ CEEQWPV VJG NGZKEQP RTQDCDKNKVKGU CTG UVKNN DCUGF QP UKPING YQTFU KP GCEJ QH VJG VYQ NCPIWCIGU 6JG NGZKEQP OQFGN RTGUGPVGF UQ HCT KU XGT[ UKORNG +P TGCNKV[ VJG VTCPUNCVKQP QH C YQTF OC[ FGRGPF QP VJG FGVCKNU QH VJG word context 6Q ECRVWTG VJGUG V[RGU QH FGRGPFGPEGU OCZKOWO GPVTQR[ OQFGNU YGTG RTQRQUGF = ?
11.3.4 Training 6JG HTGG RCTCOGVGTU QH VJG RTQDCDKNKV[ FKUVTKDWVKQPU KPVTQFWEGF CTG GUVKOCVGF HTQO C EQTRWU QH DKNKPIWCN UGPVGPEG RCKTU 6JG VTCKPKPI ETKVGTKQP KU VJG OCZKOWO NKMGNKJQQF ETKVGTKQP 5KPEG VJG OQFGNU VJCV JCXG DGGP KPVTQFWEGF CTG EQORNGZ VJG VTCKPKPI CNIQ TKVJOU ECP IWCTCPVGG QPN[ NQECN EQPXGTIGPEG +P QTFGT VQ OKVKICVG VJG RTQDNGOU YKVJ
E\&5&3UHVV//&
RQQT NQECN QRVKOC YG CRRN[ VJG EQPEGRV RTGUGPVGF KP =? 6JG VTCKPKPI RTQEGFWTG KU UVCTVGF YKVJ C UKORNG OQFGN HQT YJKEJ VJG RTQDNGO QH NQECN QRVKOC FQGU PQV QEEWT QT KU PQV ETKVKECN +P RCTVKEWNCT VJG OQFGN +$/ JCU VJG CFXCPVCIG VJCV KV JCU QPN[ C UKPING QRVKOWO CPF VJWU EQPXGTIGPEG RTQDNGOU ECPPQV GZKUV =? 6JG RCTCOGVGTU QH VJG UKORNG OQFGN CTG VJGP WUGF VQ KPKVKCNK\G VJG VTCKPKPI RTQEGFWTG QH C OQTG EQORNGZ OQFGN +P UWEJ C YC[ C UGTKGU QH OQFGNU YKVJ KPETGCUKPI EQORNGZKV[ ECP DG VTCKPGF 6[RKECN UGSWGPEGU CTG =+$/? QT =+$/ *//? 6JG VTCKPKPI RTQEGFWTG KU DCUGF QP VJG OCZKOWO NKMGNKJQQF ETKVGTKQP YJKEJ JQYGXGT ECP DG WUGF QPN[ KP CP KVGTCVKXG YC[ (QT VJG OQFGNU +$/ +$/ CPF *// VJKU KU VJG UQECNNGF GZRGEVCVKQPOCZKOK\CVKQP CNIQTKVJO HQT YJKEJ C ENQUGFHQTO UQNWVKQP KU CXCKNCDNG YKVJKP GCEJ KVGTCVKQP (QT VJG QVJGT OQFGNU PCOGN[ +$/ +$/ CPF +$/ VJKU KU PQV VJG ECUG CP[OQTG CPF GXGP YKVJKP GCEJ KVGTCVKQP PWOGTKECN CRRTQZKOCVKQPU JCXG VQ DG WUGF = ? 9JCV JCU DGGP UCKF UQ HCT IQGU HQT VJG exact NKMGNKJQQF ETKVGTKQP YJGTG YG UWO QXGT all RQUUKDNG CNKIPOGPVU 9JGP KPUVGCF YG WUG VJG maximum approximation YJGTG QPN[ VJG DGUV CNKIPOGPV KU EQPUKFGTGF VJG UKVWCVKQP OKIJV DG XGT[ OWEJ FKHHGTGPV KP VJCV UQOG QH VJG RTQDNGOU IQ CYC[ *QYGXGT VJGTG JCXG PQV [GV DGGP OCP[ U[UVGOCVKE UVWFKGU QP JQY OWEJ YG NQUG D[ VJG OCZKOWO CRRTQZKOCVKQP =? +P U[UVGOCVKE GZRGTKOGPVU KV YCU HQWPF VJCV VJG SWCNKV[ QH VJG CNKIPOGPVU FGVGTOKPGF HTQO VJG DKNKPIWCN VTCKPKPI EQTRWU JCU C FKTGEV GHHGEV QP VJG VTCPUNCVKQP SWCNKV[ =? $[ GZEJCPIKPI VJG TQNG QH VCTIGV CPF UQWTEG NCPIWCIG KP VJG VTCKPKPI RTQEGFWTG YG HQWPF VJCV VJG SWCNKV[ QH VJG CNKIPOGPVU EQWNF DG UKIPKſECPVN[ KORTQXGF (TQO C IGPGTCN RQKPV QH XKGY VJG CNKIPOGPVU ECP DG KPVGTRTGVGF CU C OGVJQF HQT ſPF KPI YQTFU QT YQTF ITQWRU VJCV CTG GSWKXCNGPV KP UQWTEG NCPIWCIG CPF VCTIGV NCPIWCIG #HVGT VJGUG GSWKXCNGPEGU JCXG DGGP HQWPF VJG[ OC[ DG OQFGNNGF KP XCTKQWU FCVC FTKXGP CRRTQCEJGU VQ DWKNF C VTCPUNCVKQP U[UVGO +P VJKU EJCRVGT YG YKNN EQPUKFGT VJG UQECNNGF CNKIPOGPV VGORNCVGU UGG NCVGT DWV VJGUG GSWKXCNGPEGU OC[ CU YGNN DG WUGF KP ſPKVGUVCVG VTCPUFWEGTU =?
11.3.5 Search 6JG VCUM QH VJG UGCTEJ CNIQTKVJO KU VQ IGPGTCVG VJG OQUV NKMGN[ VCTIGV UGPVGPEG ½ QH WPMPQYP NGPIVJ HQT CP QDUGTXGF UQWTEG UGPVGPEG ½ 6JG UGCTEJ OWUV OCMG WUG QH CNN VJTGG MPQYNGFIG UQWTEGU CU KNNWUVTCVGF D[ (KIWTG VJG CNKIPOGPV OQFGN VJG
DKNKPIWCN NGZKEQP OQFGN CPF VJG NCPIWCIG OQFGN #NN VJTGG QH VJGO OWUV EQPVTKDWVG KP VJG ſPCN FGEKUKQP CDQWV VJG YQTFU KP VJG VCTIGV NCPIWCIG 6Q KNNWUVTCVG VJG URGEKſE FGVCKNU QH VJG UGCTEJ RTQDNGO YG URGEKH[ VJG CNKIPOGPV OQFGN KP OQTG FGVCKN
¯ YG WUG inverted CNKIPOGPVU CU KP VJG OQFGN +$/ =? YJKEJ FGſPG C OCR RKPI HTQO target VQ source RQUKVKQPU TCVJGT VJCP VJG QVJGT YC[ TQWPF ¯ YG CNNQY several RQUKVKQPU KP VJG UQWTEG NCPIWCIG VQ DG EQXGTGF KG YG EQP UKFGT OCRRKPIU QH VJG HQTO
E\&5&3UHVV//&
SENTENCE IN SOURCE LANGUAGE
TRANSFORMATION
ALIGNMENT MODEL
WORD RE-ORDERING
ALIGNMENT HYPOTHESES
BILINGUAL LEXICON
LEXICAL CHOICE
WORD + POSITION HYPOTHESES
SYNTACTIC AND SEMANTIC ANALYSIS
LANGUAGE MODEL
SENTENCE HYPOTHESES
SEARCH: INTERACTION OF KNOWLEDGE SOURCES
KNOWLEDGE SOURCES
TRANSFORMATION
SENTENCE GENERATED IN TARGET LANGUAGE
FIGURE 11.3 Illustration of search in statistical translation. (QT VJKU KPXGTVGF CNKIPOGPV OCRRKPI YKVJ sets UWOG C UQTV QH ſTUVQTFGT OQFGN
QH UQWTEG RQUKVKQPU YG CICKP CU
½ ½ YJGTG YG JCXG FTQRRGF VJG FGRGPFGPEG QP CPF
9G TGRNCEG VJG UWO QXGT CNN CNKIPOGPVU D[ VJG DGUV CNKIPOGPV YJKEJ KU TGHGTTGF VQ CU OCZKOWO CRRTQZKOCVKQP KP URGGEJ TGEQIPKVKQP 7UKPI C VTKITCO NCPIWCIG OQFGN YG QDVCKP VJG HQNNQYKPI UGCTEJ ETKVGTKQP
½ ½
E\&5&3UHVV//&
¾
6#4)'6215+6+10
K K
L 5174%'215+6+10
FIGURE 11.4 Illustration of bottom-to-top search.
%QPUKFGTKPI VJKU ETKVGTKQP YG ECP UGG VJCV YG ECP DWKNF WR J[RQVJGUGU QH RCTVKCN VCTIGV UGPVGPEGU KP C bottom-to-top UVTCVGI[ QXGT VJG RQUKVKQPU QH VJG VCTIGV UGPVGPEG ½ CU KNNWUVTCVGF KP (KIWTG #P KORQTVCPV EQPUVTCKPV HQT VJG CNKIPOGPV KU VJCV all RQUKVKQPU QH VJG UQWTEG UGPVGPEG UJQWNF DG EQXGTGF GZCEVN[ once 6JKU EQPUVTCKPV KU UKOKNCT VQ VJCV QH VJG VTCXGNKPI UCNGUOCP RTQDNGO YJGTG GCEJ EKV[ JCU VQ DG XKUKVGF GZCEVN[ QPEG &GVCKNU QP XCTKQWU UGCTEJ UVTCVGIKGU ECP DG HQWPF KP = ? 6JG V[RG QH NCPIWCIG OQFGN YG WUG TCPIGU HTQO C VTKITCO VQ C ITCO YJKEJ ECP DG GKVJGT YQTF QT ENCUUDCUGF $GCO UGCTEJ KU WUGF VQ JCPFNG VJG JWIG UGCTEJ URCEG 6Q PQTOCNK\G VJG EQUVU QH RCTVKCN J[RQVJGUGU EQXGTKPI FKHHGTGPV RCTVU QH VJG KPRWV UGPVGPEG CP QRVKOKUVKE GUVKOCVKQP QH VJG TGOCKPKPI EQUV KU CFFGF VQ VJG EWTTGPV CEEWOWNCVGF EQUV CU HQNNQYU (QT GCEJ YQTF KP VJG UQWTEG UGPVGPEG C NQYGT DQWPF QP KVU VTCPUNCVKQP EQUV KU FGVGTOKPGF DGHQTGJCPF 7UKPI VJKU NQYGT DQWPF KV KU RQUUKDNG VQ CEJKGXG CP GHſEKGPV GUVKOCVKQP QH VJG TGOCKPKPI EQUV = ? (QT QVJGT RCRGTU QP VJG UGCTEJ RTQEGUU KP VTCPUNCVKQP VJG TGCFGT KU TGHGTTGF VQ = ?
11.3.6 Algorithmic Differences between Speech Recognition and Language Translation +V KU KPVGTGUVKPI VQ EQPUKFGT VJG FKHHGTGPEGU DGVYGGP VJG CNIQTKVJOU HQT URGGEJ TGEQI PKVKQP CPF VJQUG HQT OCEJKPG VTCPUNCVKQP
E\&5&3UHVV//&
¯ OQPQVQPKEKV[ +P URGGEJ TGEQIPKVKQP VJGTG KU C UVTKEV OQPQVQPKEKV[ DGVYGGP VJG UGSWGPEG QH CEQWUVKE XGEVQTU CPF VJG UGSWGPEG QH TGEQIPK\GF YQTFU QT RJQPGOGU 6JKU KU PQV VJG ECUG HQT OCEJKPG VTCPUNCVKQP CPF VJGTGHQTG VJG UGCTEJ RTQDNGO DGEQOGU OQTG EQORNKECVGF ¯ HGTVKNKV[ +P OCEJKPG VTCPUNCVKQP YG JCXG VQ FGEKFG YJGVJGT C YQTF KU RTGUGPV KP VJG VCTIGV UVTKPI QT PQV 6JGTGHQTG KV KU KORQTVCPV VQ CUUKIP C HGTVKNKV[ VQ GCEJ YQTF QH VJG VCTIGV XQECDWNCT[ +P URGGEJ TGEQIPKVKQP VJG EQWPVGTRCTV QH C YQTF KU CP *// UVCVG *QYGXGT YG PGXGT VCMG FGEKUKQPU CDQWV UVCVGU DWV CDQWV YJQNG RJQPGOG OQFGNU GKVJGT YKVJ QT YKVJQWV EQPVGZV 6JGTGHQTG VJG EQPEGRV QH HGTVKNKV[ KU PQV TGCNN[ PGGFGF KP URGGEJ TGEQIPKVKQP
11.4 Alignment Templates: From Single Words to Word Groups 11.4.1 Concept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ſEKGPE[ QH VJG DCUGNKPG CNKIPOGPV OQFGNU KU VJCV VJG[ CTG CDNG VQ OQFGN EQTTGURQPFGPEGU QPN[ DGVYGGP UKPING YQTFU # ſTUV EQWPVGTOGCUWTG YCU VJG TGſPGF CNKIPOGPV OQFGN WUGF KP VJG SWCUKOQPQVQPG UGCTEJ # OQTG U[UVGOCVKE CRRTQCEJ KU VQ EQPUKFGT YQTF ITQWRU TCVJGT VJCP UKPING YQTFU CU VJG DCUKU HQT VJG CNKIPOGPV OQFGNU +P QVJGT YQTFU C YJQNG ITQWR QH CFLCEGPV YQTFU KP VJG UQWTEG UGPVGPEG OC[ DG CNKIPGF YKVJ C YJQNG ITQWR QH CFLCEGPV YQTFU KP VJG VCTIGV NCPIWCIG =? 5WEJ C OCRRKPI YKNN DG TGHGTTGF VQ CU alignment template KP VJG HQNNQYKPI 'ZCORNG QH UWEJ CNKIPOGPV VGORNCVGU CTG UJQYP KP (KIWTG VJGUG GZCORNGU YGTG TGCN GZRGTKOGPVCN TGUWNVU QDVCKPGF D[ VJG OGVJQF VQ DG RTGUGPVGF #U ECP DG UGGP HTQO
E\&5&3UHVV//&
okay , wie sieht es am neunzehnten aus , vielleicht um zwei Uhr nachmittags ?
? afternoon the in o’clock two , maybe at nineteenth the about how , okay
FIGURE 11.5 Example of alignment templates for a German-English sentence pair.
VJGUG GZCORNGU VJG CFXCPVCIG QH VJG CNKIPOGPV VGORNCVG KU VJCV DQVJ VJG YQTF EQPVGZV CPF VJG NQECN TGQTFGTKPI QH YQTFU ECP DG VCMGP KPVQ CEEQWPV 6Q FGUETKDG VJG CNKIPOGPV VGORNCVG CRRTQCEJ KP C HQTOCN YC[ YG ſTUV FGEQORQUG DQVJ VJG UQWTEG UGPVGPEG ½ CPF VJG VCTIGV UGPVGPEG ½ KPVQ C UGSWGPEG QH YQTF ITQWRU
½ ½ ½
½
·½ ½ ·½
½
6Q UKORNKH[ VJG PQVCVKQP CPF VJG RTGUGPVCVKQP YG KIPQTG VJG HCEV VJCV VJGTG ECP DG C NCTIG PWODGT QH RQUUKDNG UGIOGPVCVKQPU CPF CUUWOG VJCV VJGTG KU QPN[ QPG UGIOGP VCVKQP 9G FKUVKPIWKUJ VYQ NGXGNU QH CNKIPOGPVU CNKIPOGPV within VJG YQTF ITQWRU ½ between VJG UQWTEG YQTF CPF CNKIPOGPV between YQTF ITQWRU (QT VJG CNKIPOGPV
E\&5&3UHVV//&
ITQWRU ½ CPF VJG VCTIGV YQTF ITQWRU ½ YG JCXG VJG HQNNQYKPI GSWCVKQP
½ ½ ½ ½
½
½
½
YJGTG YG JCXG WUGF C ſTUVQTFGT CNKIPOGPV OQFGN (QT VJG CNKIPOGPV within VJG YQTF ITQWR YG KPVTQFWEG C PGY JKFFGP XCTKCDNG YJKEJ YKNN DG TGHGTTGF VQ CU CNKIPOGPV VGORNCVG &GPQVKPI VJG UQWTEG YQTF ITQWR D[ ¼ ¼ CPF VJG VCTIGV YQTF ITQWR D[ YG JCXG VJG OQFGN
'CEJ CNKIPOGPV VGORNCVG ECP DG TGRTGUGPVGF CU C DKPCT[ OCVTKZ YKVJ ¼ TQYU CPF ¼ EQNWOPU YJGTG QPN[ VJG JKIJRTQDCDKNKV[ RCKTU JCXG C XCNWG QH
FGPQVGF D[ C HWNN USWCTG KP (KIWTG 6JG RTQDCDKNKVKGU CPF CTG FGVGTOKPGF WUKPI VJG CNKIPGF VTCKPKPI EQTRWU CPF CTG UGV VQ \GTQ KH VJG VTKRNG FKF PQV QEEWT KP VJG VTCKPKPI EQTRWU +H VJG VTKRNG FKF QEEWT KP VJG VTCKPKPI EQTRWU YG WUG VJG HQNNQYKPI OQFGN HQT
¼
¼
¼
¼
6[RKECNN[ VJGTG KU CP CFFKVKQPCN TGſPGOGPV UVGR D[ KPVTQFWEKPI C UGV QH DKNKPIWCN YQTF ENCUUGU VJCV CTG FGVGTOKPGF CWVQOCVKECNN[ =? 6JG CNKIPOGPV VGORNCVGU CTG VJGP FGſPGF CV VJKU NGXGN QH DKNKPIWCN YQTF ENCUUGU TCVJGT VJCP QP VJG NGXGN QH VJG YQTFU VJGOUGNXGU 6JKU UVGR UNKIJVN[ KORTQXGU VJG IGPGTCNK\CVKQP ECRCDKNKV[ HQT WPUGGP VGUV FCVC
11.4.2 Training 6JG VTCKPKPI QH CNKIPOGPV VGORNCVGU UVCTVU YKVJ VJG VTCKPKPI QH VYQ *// CNKIPOGPV OQFGNU HQT each QH VJG VYQ VTCPUNCVKQP FKTGEVKQPU UQWTEG VCTIGV CPF VCTIGV UQWTEG #U C TGUWNV YG QDVCKP CP CNKIPOGPV OCVTKZ HQT GCEJ VTCKPKPI UGPVGPEG RCKT D[ OGTIKPI VJG CNKIPOGPV RCVJU QH DQVJ VTCPUNCVKQP FKTGEVKQPU +P UWEJ CP CNKIPOGPV OCVTKZ KV KU RQUUKDNG VJCV QPG UQWTEG YQTF KU CNKIPGF VQ OQTG VJCP QPG VCTIGV YQTF
E\&5&3UHVV//&
7UKPI VJG YJQNG UGV QH CNKIPOGPV OCVTKEGU HQT VJG VTCKPKPI EQTRWU YG VJGP GZVTCEV VJG CNKIPOGPV VGORNCVGU D[ EQPUKFGTKPI CNN RQUUKDNG UQWTEGVCTIGV YQTF ITQWRU WP FGT VJG EQPUVTCKPV VJCV VJG YQTFU YKVJKP VJG UQWTEGVCTIGV RJTCUG CTG QPN[ CNKIPGF VQ YQTFU YKVJKP VJG VCTIGVUQWTEG RJTCUG 6JG RTQDCDKNKV[ KU VJGP GUVKOCVGF CU VJG TGNCVKXG HTGSWGPE[ HQT VJG GXGPV RCKTU
11.4.3 Search 6Q RGTHQTO VJG UGCTEJ YG WUG VJG HQNNQYKPI OQFGNU
¯ #U NCPIWCIG OQFGN YG WUG C ENCUUDCUGF ŌITCO GI VTKITCO QT ITCO NCPIWCIG OQFGN YKVJ DCEMKPIQHH 6[RKECNN[ VJKU KU UNKIJVN[ DGVVGT VJCP VJG UVCPFCTF DKITCO NCPIWCIG OQFGN ¯ 9G CUUWOG VJCV CNN RQUUKDNG UGIOGPVCVKQPU JCXG VJG UCOG RTQDCDKNKV[ ¯ 6JG CNKIPOGPV OQFGN CV VJG VGORNCVG NGXGN KU CP *//V[RG CNKIPOGPV OQFGN 1DXKQWUN[ CU WUWCN CNN YQTFU KP VJG UQWTEG UVTKPI OWUV DG EQXGTGF 6Q IGPGTCVG VJG WPMPQYP VCTIGV UGPVGPEG KP VJG UGCTEJ RTQEGFWTG YG JCXG VQ CNNQY HQT CNN RQUUKDNG UGIOGPVCVKQPU QH VJG UQWTEG UGPVGPEG KPVQ YQTF ITQWRU HQT CNN RQUUK DNG CNKIPOGPVU between the word groups CPF HQT RQUUKDNG CNKIPOGPVU within VJG YQTF ITQWRU 6JGTG CTG C EQWRNG QH UKORNKſECVKQPU CPF CRRTQZKOCVKQPU VQ TGFWEG VJG EQO RWVCVKQPCN EQUV QH VJG UGCTEJ YJKEJ ECPPQV DG FGUETKDGF JGTG HQT URCEG NKOKVCVKQPU +P RTKPEKRNG YG WUG C DGCO UGCTEJ UVTCVGI[ 6JG UGCTEJ CNIQTKVJO DWKNFU WR J[RQVJG UGU QH KPETGCUKPI NGPIVJ CNQPI VJG RQUKVKQPU QH VJG VCTIGV UVTKPI &WTKPI VJG UGCTEJ RTQEGUU YG EQORWVG CP GUVKOCVG HQT VJG TGOCKPKPI RQTVKQP QH VJG UQWTEG UVTKPI VJCV JCU PQV DGGP EQXGTGF 6JKU GUVKOCVG HQT VJG TGOCKPKPI RQTVKQP KU EQODKPGF YKVJ VJG RTQDCDKNKV[ UEQTG HQT VJG CNTGCF[ EQXGTGF RQTVKQP QH VJG UQWTEG UVTKPI VQ PCTTQY FQYP VJG UGCTEJ VQ VJG OQUV RTQOKUKPI UGCTEJ J[RQVJGUGU
11.5 Experimental Results 11.5.1 The Task and the Corpus 6JG IQCN QH VJG 8 '4$/1$+. =? KU VJG VTCPUNCVKQP QH URQMGP FKCNQIWGU KP VJG FQ OCKPU QH CRRQKPVOGPV UEJGFWNKPI CPF VTCXGN RNCPPKPI +P C V[RKECN UKVWCVKQP C PCVKXG )GTOCP URGCMGT CPF C PCVKXG 'PINKUJ URGCMGT EQPFWEV C FKCNQIWG YJGTG VJG[ ECP QPN[ KPVGTCEV D[ URGCMKPI CPF NKUVGPKPI VQ VJG 8 '4$/1$+. U[UVGO 9KVJKP VJG 8 '4$/1$+. RTQLGEV URQMGP FKCNQIWGU YGTG TGEQTFGF 6JGUG FKCNQIWGU YGTG OCPWCNN[ VTCPUETKDGF CPF NCVGT OCPWCNN[ VTCPUNCVGF D[ 8 '4$/1$+. RCTVPGTU
*KNFGUJGKO HQT 2JCUG + CPF 6iWDKPIGP HQT 2JCUG ++ 5KPEG FKHHGTGPV JWOCP VTCPUNC VQTU YGTG KPXQNXGF VJGTG KU ITGCV XCTKCDKNKV[ KP VJG VTCPUNCVKQPU
E\&5&3UHVV//&
'CEJ QH VJGUG UQECNNGF FKCNQIWG VWTPU OC[ EQPUKUV QH UGXGTCN UGPVGPEGU URQMGP D[ VJG UCOG URGCMGT CPF KU UQOGVKOGU TCVJGT NQPI #U C TGUWNV VJGTG KU PQ QPGVQ QPG EQTTGURQPFGPEG DGVYGGP UQWTEG CPF VCTIGV UGPVGPEGU 6Q CEJKGXG C QPGVQQPG EQTTGURQPFGPEG VJG FKCNQIWG VWTPU CTG URNKV KPVQ UJQTVGT UGIOGPVU WUKPI RWPEVWCVKQP OCTMU CU RQVGPVKCN URNKV RQKPVU 5KPEG VJG RWPEVWCVKQP OCTMU KP UQWTEG CPF VCTIGV UGPVGPEGU CTG PQV PGEGUUCTKN[ KFGPVKECN C F[PCOKE RTQITCOOKPI CRRTQCEJ KU WUGF VQ ſPF VJG QRVKOCN UGIOGPVCVKQP RQKPVU 6JG PWODGT QH UGIOGPVU KP VJG UQWTEG UGPVGPEG CPF KP VJG VGUV UGPVGPEG ECP DG FKHHGTGPV 6JG UGIOGPVCVKQP KU UEQTGF WUKPI C YQTF DCUGF CNKIPOGPV OQFGN CPF VJG UGIOGPVCVKQP YKVJ VJG DGUV UEQTG KU UGNGEVGF 6JKU UGIOGPVGF EQTRWU KU VJG UVCTVKPI RQKPV HQT VJG VTCKPKPI QH VTCPUNCVKQP CPF NCPIWCIG OQFGNU #NKIPOGPV OQFGNU QH KPETGCUKPI EQORNGZKV[ CTG VTCKPGF QP VJKU DKNKPIWCN EQTRWU = ? # UVCPFCTF XQECDWNCT[ JCF DGGP FGſPGF HQT VJG XCTKQWU URGGEJ TGEQIPK\GTU WUGF KP 8 '4$/1$+. *QYGXGT PQV CNN YQTFU QH VJKU XQECDWNCT[ YGTG QDUGTXGF KP VJG VTCKP KPI EQTRWU 6JGTGHQTG VJG VTCPUNCVKQP XQECDWNCT[ YCU GZVGPFGF UGOKCWVQOCVKECNN[ D[ CFFKPI CDQWV )GTOCPŌ'PINKUJ YQTF RCKTU HTQO CP QPNKPG DKNKPIWCN NGZKEQP CXCKNCDNG QP VJG YGD 6JG TGUWNVKPI NGZKEQP EQPVCKPGF PQV QPN[ YQTFYQTF GPVTKGU DWV CNUQ OWNVKYQTF VTCPUNCVKQPU GURGEKCNN[ HQT VJG NCTIG PWODGT QH )GTOCP EQORQWPF YQTFU 6Q EQWPVGTCEV VJG URCTUGPGUU QH VJG VTCKPKPI FCVC C EQWRNG QH UVTCKIJVHQTYCTF TWNGDCUGF RTGRTQEGUUKPI UVGRU YGTG CRRNKGF before CP[ QVJGT V[RG QH RTQEGUUKPI
¯ ECVGIQTK\CVKQP QH RTQRGT PCOGU HQT RGTUQPU CPF EKVKGU ¯ PQTOCNK\CVKQP QH Ō PWODGTU Ō VKOG CPF FCVG RJTCUGU do not Ō URGNNKPI don’t URNKVVKPI QH )GTOCP EQORQWPF YQTFU
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
11.5.2 Offline Results &WTKPI VJG RTQITGUU QH VJG 8 '4$/1$+. RTQLGEV FKHHGTGPV XCTKCPVU QH UVCVKUVKECN VTCPU NCVKQP YGTG KORNGOGPVGF CPF GZRGTKOGPVCN VGUVU YGTG RGTHQTOGF HQT DQVJ VGZV CPF URGGEJ KPRWV 6Q UWOOCTK\G VJGUG GZRGTKOGPVCN VGUVU YG DTKGƀ[ TGRQTV GZRGTKOGPVCN QHƀKPG TGUWNVU HQT VJG HQNNQYKPI VTCPUNCVKQP CRRTQCEJGU
UKPINGYQTF DCUGF CRRTQCEJ =? CNKIPOGPV VGORNCVG CRRTQCEJ =?
E\&5&3UHVV//&
¯ ECUECFGF VTCPUFWEGT CRRTQCEJ =? 7PNKMG VJG QVJGT VYQCRRTQCEJGU VJKU CRRTQCEJ TGSWKTGU C UGOKCWVQOCVKE VTCKPKPI RTQEGFWTG KP YJKEJ VJG UVTWEVWTG QH VJG ſPKVG UVCVG VTCPUFWEGTU KU FG UKIPGF OCPWCNN[ (QT OQTG FGVCKNU UGG =? 6JG QHƀKPG VGUVU YGTG RGTHQTOGF QP VGZV KPRWV HQT VJG VTCPUNCVKQP FKTGEVKQP HTQO )GT OCP VQ 'PINKUJ 6JG VGUV UGV EQPUKUVGF QH UGPVGPEGU YJKEJ EQORTKUGF YQTFU CPF RWPEVWCVKQP OCTMU 6JG TGUWNVU CTG UJQYP KP 6CDNG 6Q LWFIG CPF EQORCTG VJG SWCNKV[ QH FKHHGTGPV VTCPUNCVKQP CRRTQCEJGU KP QHƀKPG VGUVU YG V[RK ECNN[ WUG VJG HQNNQYKPI GTTQT OGCUWTGU =?
¯ O9'4 OWNVKTGHGTGPEG YQTF GTTQT TCVG (QT GCEJ VGUV UGPVGPEG KP VJG UQWTEG NCPIWCIG VJGTG CTG several TGHGTGPEG VTCPU NCVKQPU KP VJG VCTIGV NCPIWCIG (QT GCEJ VTCPUNCVKQP QH VJG VGUV UGPVGPEG VJG GFKV FKUVCPEGU PWODGT QH UWDUVKVWVKQPU FGNGVKQPU CPF KPUGTVKQPU CU KP URGGEJ TGEQIPKVKQP VQ CNN TGHGTGPEG VTCPUNCVKQPU CTG ECNEWNCVGF CPF VJG UOCNNGUV FKU VCPEG KU UGNGEVGF CPF WUGF CU GTTQT OGCUWTG (QT CP GZVGPUKQP QH O9'4DCUGF OGCUWTGU UGG CNUQ =? ¯ 55'4 UWDLGEVKXG UGPVGPEG GTTQT TCVG =? 'CEJ VTCPUNCVGF UGPVGPEG KU LWFIGF D[ C JWOCP GZCOKPGT CEEQTFKPI VQ CP GT TQT UECNG HTQO UGOCPVKECNN[ CPF U[PVCEVKECNN[ EQTTGEV VQ EQORNGVGN[ YTQPI $QVJ GTTQT OGCUWTGU CTG TGRQTVGF KP 6CDNG #NVJQWIJ VJG GZRGTKOGPVU YKVJ VJG ECUECFGF VTCPUFWEGTU =? YGTG PQV HWNN[ QRVKOK\GF [GV VJG RTGNKOKPCT[ TGUWNVU KPFK ECVGF VJCV VJKU UGOKCWVQOCVKE CRRTQCEJ FQGU PQV IGPGTCNK\G CU YGNN CU VJG QVJGT VYQ HWNN[ CWVQOCVKE CRRTQCEJGU #OQPI VJGUG VYQ VJG CNKIPOGPV VGORNCVG CRRTQCEJ YCU HQWPF VQ YQTM EQPUKUVGPVN[ DGVVGT CETQUU FKHHGTGPV VGUV UGVU CPF CNUQ VCUMU FKHHGTGPV HTQO 8 '4$/1$+. 6JGTGHQTG VJG CNKIPOGPV VGORNCVG CRRTQCEJ YCU WUGF KP VJG ſPCN 8 '4$/1$+. RTQVQV[RG U[UVGO TABLE 11.1
$KNKPIWCN VTCKPKPI EQTRWU TGEQIPKVKQP NGZKEQP CPF VTCPUNCVKQP NGZKEQP 6TCKPKPI 6GZV 5GPVGPEG 2CKTU 9QTFU 9QTFU 2WPEV/CTMU 8QECDWNCT[ 5KPINGVQPU 4GEQIPKVKQP 8QECDWNCT[ 6TCPUNCVKQP #FFGF 9QTF 2CKTU 8QECDWNCT[
E\&5&3UHVV//&
)GTOCP 'PINKUJ
TABLE 11.2
%QORCTKUQP QH VJTGG UVCVKUVKECN VTCPUNCVKQP CRRTQCEJGU VGUV QP VGZV KPRWV UGPVGPEGU YQTFU RWPEVWCVKQP OCTMU 6TCPUNCVKQP #RRTQCEJ 5KPING9QTF $CUGF #NKIPOGPV 6GORNCVG %CUECFGF 6TCPUFWEGTU
O9'4 =?
55'4 =?
11.5.3 Integration into the 8 '4$/1$+. Prototype System 6JG UVCVKUVKECN CRRTQCEJ VQ OCEJKPG VTCPUNCVKQP KU GODQFKGF KP VJG stattrans OQFWNG YJKEJ KU KPVGITCVGF KPVQ VJG 8 '4$/1$+. RTQVQV[RG U[UVGO 9G DTKGƀ[ TGXKGY VJQUG CURGEVU QH KV VJCV CTG TGNGXCPV HQT VJG UVCVKUVKECN VTCPUNCVKQP CRRTQCEJ 6JG KORNGOGP VCVKQP UWRRQTVU VJG VTCPUNCVKQP FKTGEVKQPU HTQO )GTOCP VQ 'PINKUJ CPF HTQO 'PINKUJ VQ )GTOCP +P TGIWNCT RTQEGUUKPI OQFG VJG stattrans OQFWNG TGEGKXGU KVU KPRWV HTQO VJG repair OQFWNG =? #V VJCV VKOG VJG YQTF NCVVKEGU CPF DGUV J[RQVJGUGU HTQO VJG URGGEJ TGEQIPKVKQP U[UVGOU JCXG CNTGCF[ DGGP RTQUQFKECNN[ CPPQVCVGF KG KPHQTOC VKQP CDQWV RTQUQFKE UGIOGPV DQWPFCTKGU UGPVGPEG OQFG CPF CEEGPVWCVGF U[NNCDNGU CTG CFFGF VQ GCEJ GFIG KP VJG YQTF NCVVKEG =? 6JG VTCPUNCVKQP KU RGTHQTOGF QP VJG UKPING DGUV UGPVGPEG J[RQVJGUKU QH VJG TGEQIPK\GT 6JG RTQUQFKE DQWPFCTKGU CPF VJG UGPVGPEG OQFG KPHQTOCVKQP CTG WVKNK\GF D[ VJG stattrans OQFWNG CU HQNNQYU +H VJGTG KU C OCLQT RJTCUG DQWPFCT[ C HWNN UVQR QT SWGU VKQP OCTM KU KPUGTVGF KPVQ VJG YQTF UGSWGPEG FGRGPFKPI QP VJG UGPVGPEG OQFG CU KPFKECVGF D[ VJG prosody OQFWNG #FFKVKQPCN EQOOCU CTG KPUGTVGF HQT QVJGT V[RGU QH UGIOGPV DQWPFCTKGU 6JG prosody OQFWNG ECNEWNCVGU RTQDCDKNKVKGU HQT UGIOGPV DQWPFCTKGU CPF VJTGUJQNFU CTG WUGF VQ FGEKFG KH VJG UGPVGPEG OCTMU CTG VQ DG KP UGTVGF 6JGUG VJTGUJQNFU JCXG DGGP UGNGEVGF KP UWEJ C YC[ VJCV QP VJG CXGTCIG HQT GCEJ FKCNQIWG VWTP C IQQF UGIOGPVCVKQP KU QDVCKPGF 6JG UGIOGPV DQWPFCTKGU TG UVTKEV RQUUKDNG YQTF TGQTFGTKPI DGVYGGP UQWTEG CPF VCTIGV NCPIWCIG 6JKU PQV QPN[ KORTQXGU VTCPUNCVKQP SWCNKV[ DWV CNUQ TGUVTKEVU VJG UGCTEJ URCEG CPF VJGTGD[ URGGFU WR VJG VTCPUNCVKQP RTQEGUU
11.5.4 Final Evaluation 9JGTGCU VJG QHƀKPG VGUVU TGRQTVGF CDQXG YGTG KORQTVCPV HQT VJG QRVKOK\CVKQP CPF VWPKPI QH VJG U[UVGO VJG OQUV KORQTVCPV GXCNWCVKQP YCU VJG ſPCN GXCNWCVKQP QH VJG 8 '4$/1$+. RTQVQV[RG KP URTKPI 6JKU ſPCN GXCNWCVKQP QH VJG 8 '4$/1$+. U[UVGO YCU RGTHQTOGF CV VJG 7PKXGTUKV[ QH *CODWTI =? 6JTGG QVJGT VTCPUNCVKQP CRRTQCEJGU JCF DGGP KPVGITCVGF KPVQ VJG 8 '4$/1$+. RTQVQ V[RG U[UVGO
¯ C ENCUUKECN VTCPUHGT CRRTQCEJ = ? YJKEJ KU DCUGF QP C OCPWCNN[ FGUKIPGF CPCN[UKU ITCOOCT C UGV QH VTCPUHGT
E\&5&3UHVV//&
TWNGU CPF C IGPGTCVKQP ITCOOCT
¯ C FKCNQIWG CEV DCUGF CRRTQCEJ =? YJKEJ COQWPVU VQ C UQTV QH slot filling D[ ENCUUKH[KPI GCEJ UGPVGPEG KPVQ QPG QWV QH C UOCNN PWODGT QH RQUUKDNG UGPVGPEG RCVVGTPU CPF ſNNKPI KP VJG UNQV XCNWGU CPF ¯ CP GZCORNGDCUGF CRRTQCEJ =? YJGTG C UQTV QH PGCTGUV PGKIJDQT EQPEGRV KU CRRNKGF VQ VJG UGV QH DKNKPIWCN VTCKP KPI UGPVGPEG RCKTU CHVGT UWKVCDNG RTGRTQEGUUKPI +P VJG ſPCN GXCNWCVKQP JWOCP GXCNWCVQTU LWFIGF VJG VTCPUNCVKQP SWCNKV[ HQT GCEJ QH VJG HQWT VTCPUNCVKQP TGUWNVU WUKPI VJG HQNNQYKPI ETKVGTKQP +U VJG UGPVGPEG CRRTQZKOCVKXGN[ EQTTGEV
[GUPQ!
6JG GXCNWCVQTU YGTG CUMGF VQ RC[ RCTVKEWNCT CVVGPVKQP VQ VJG UGOCPVKE KPHQTOCVKQP
GI FCVG CPF RNCEG QH OGGVKPI RCTVKEKRCPVU GVE EQPVCKPGF KP VJG VTCPUNCVKQP # OKUUKPI VTCPUNCVKQP CU KV OC[ JCRRGP HQT VJG VTCPUHGT CRRTQCEJ QT QVJGT CRRTQCEJGU YCU EQWPVGF CU YTQPI VTCPUNCVKQP 6JG GXCNWCVKQP YCU DCUGF QP FKCNQIWG VWTPU HQT VJG VTCPUNCVKQP HTQO )GTOCP VQ 'PINKUJ CPF QP FKCNQIWG VWTPU HQT VJG VTCPU NCVKQP HTQO 'PINKUJ VQ )GTOCP 6JG URGGEJ TGEQIPK\GTU WUGF JCF C YQTF GTTQT TCVG QH CDQWV 6JG QXGTCNN UGPVGPEG GTTQT TCVGU KG TGUWNVKPI HTQO TGEQIPKVKQP and VTCPUNCVKQP CTG UWOOCTK\GF KP 6CDNG #U YG ECP UGG VJG GTTQT TCVGU HQT VJG UVCVKUVKECN CRRTQCEJ CTG UOCNNGT D[ C HCEVQT QH CDQWV KP EQORCTKUQP YKVJ VJG QVJGT CRRTQCEJGU #NVJQWIJ VJG CDUQNWVG XCNWGU QH VJG GTTQT TCVGU UJQYP KP 6CDNG OC[ FGRGPF JGCXKN[ QP VJG URGEKſE VGUV EQPFKVKQPU WUGF KP =? VJGTG KU PQ TGCUQP VQ CUUWOG VJCV VJG relative RGTHQTOCPEG QH VJG HQWT CRRTQCEJGU YKNN DG VJGTGD[ EJCPIGF +P CFFKVKQP VQ VJG HQWT OGVJQFU UJQYP KP 6CDNG VJGTG YCU C ſHVJ OGVJQF ECNNGF UWDUVTKPIDCUGF VTCPUNCVKQP =? 6JKU OGVJQF KU DCUGF QP DKNKPIWCN YQTF UVTKPIU VJCV CTG UKOKNCT VQ CNKIPOGPV VGORNCVGU CPF CTG GZVTCEVGF HTQO UVCVKUVKECN CNKIPOGPVU +VU RGTHQTOCPEG YCU UNKIJVN[ KPHGTKQT VQ VJG UVCVKUVKECN CRRTQCEJ *QYGXGT VJKU OGVJQF YCU PQV RCTV QH VJG QTKIKPCN RTQVQV[RG U[UVGO CPF YCU PQV GXCNWCVGF QP GZCEVN[ VJG UCOG EQTRWU =? TABLE 11.3
5GPVGPEG GTTQT TCVGU QH ſPCN GXCNWCVKQP URGGEJ TGEQIPK\GT YKVJ 9'4 EQTRWU QH CPF FKCNQIWG VWTPU HQT VTCPUNCVKQP )GTOCP VQ 'PINKUJ CPF 'PINKUJ VQ )GTOCP TGURGEVKXGN[ 6TCPUNCVKQP /GVJQF 'TTQT =? 5GOCPVKE 6TCPUHGT &KCNQIWG #EV $CUGF 'ZCORNG $CUGF 5VCVKUVKECN
E\&5&3UHVV//&
+P CITGGOGPV YKVJ QVJGT GXCNWCVKQP GZRGTKOGPVU VJGUG GZRGTKOGPVU UJQY VJCV VJG UVCVKUVKECN OQFGNNKPI CRRTQCEJ OC[ DG EQORCTCDNG VQ QT DGVVGT VJCP VJG EQPXGPVKQPCN TWNGDCUGF CRRTQCEJ +P RCTVKEWNCT VJG UVCVKUVKECN CRRTQCEJ UGGOU VQ JCXG VJG CF XCPVCIG KH TQDWUVPGUU KU KORQTVCPV GI YJGP VJG KPRWV UVTKPI KU PQV ITCOOCVKECNN[ EQTTGEV QT YJGP KV KU EQTTWRVGF D[ TGEQIPKVKQP GTTQTU #NVJQWIJ DQVJ VGZV CPF URGGEJ KPRWV CTG VTCPUNCVGF YKVJ IQQF SWCNKV[ QP VJG CXGTCIG D[ VJG UVCVKUVKECN CRRTQCEJ VJGTG CTG GZCORNGU YJGTG VJG U[PVCEVKE UVTWEVWTG QH VJG RTQFWEGF UGPVGPEG KU PQV EQTTGEV 5QOG QH VJGUG U[PVCEVKE GTTQTU CTG TGNCVGF VQ NQPI TCPIG FGRGPFGPEGU CPF U[PVCEVKE UVTWEVWTGU VJCV CTG PQV ECRVWTGF D[ VJG ITCO NCP IWCIG OQFGN WUGF 6Q EQRG YKVJ VJGUG RTQDNGOU OQTRJQU[PVCEVKE CPCN[UKU =? CPF ITCOOCTDCUGF NCPIWCIG OQFGNU =? CTG EWTTGPVN[ DGKPI UVWFKGF
11.6 Speech Translation: The Integrated Approach 11.6.1 Principle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̽ ½ Ø Ì QXGT VKOG VJG UQWTEG YQTFU ½Â CPF VJG VCTIGV YQTFU Á½
̽
 Á ½
½
(TQO C UVTKEV RQKPV QH XKGY VJG UQWTEG YQTFU ½Â CTG PQV QH FKTGEV KPVGTGUV HQT VJG URGGEJ VTCPUNCVKQP VCUM /CVJGOCVKECNN[ VJKU KU ECRVWTGF D[ KPVTQFWEKPI VJG RQUUKDNG UQWTEG YQTF UVTKPIU ½Â CU JKFFGP XCTKCDNGU KPVQ VJG $C[GU FGEKUKQP TWNG
E\&5&3UHVV//&
½ ½ ½
½
½
½
½
½
½ ½ ½
½ ½
½
½ ½ ½
½ ½ ½ ½ ½
½
½
½
½ ½ ½ ½
½ ½ ½ ½ ½ ½
*GTG YG JCXG OCFG PQ URGEKCN OQFGNNKPI CUUWORVKQP CRCTV HTQO VJG TGCUQPCDNG CUUWORVKQP VJCV
½ ½ ½ ½ ½ KG VJG VCTIGV UVTKPI ½ FQGU PQV JGNR VQ RTGFKEV VJG CEQWUVKE XGEVQTU KP VJG UQWTEG NCP IWCIG if VJG UQWTEG UVTKPI ½ KU IKXGP +P CFFKVKQP KP VJG NCUV GSWCVKQP YG JCXG WUGF VJG OCZKOWO CRRTQZKOCVKQP 1PN[ KP VJCV URGEKCN ECUG QH URGGEJ VTCPUNCVKQP CV NGCUV HTQO C UVTKEV RQKPV QH XKGY VJGTG KU VJG PQVKQP QH C ŏTGEQIPK\GFŏ UQWTEG YQTF UGSWGPEG ½ *QYGXGT VJKU YQTF UGSWGPEG KU XGT[ OWEJ FGVGTOKPGF D[ VJG EQODKPCVKQP QH VJG NCPIWCIG OQFGN ½ QH VJG VCTIGV NCPIWCIG CPF VJG VTCPUNCVKQP OQFGN ½ ½ +P EQPVTCUV KP TGEQIPKVKQP VJGTG YQWNF DG QPN[ VJG NCPIWCIG OQFGN ½
11.6.2 Practical Implementation 9JGP RTGUGPVKPI VJG UVCVKUVKECN CRRTQCEJ VQ YTKVVGP NCPIWCIG VTCPUNCVKQP VJG VCEKV CUUWORVKQP JCF DGGP VJCV VJG UQWTEG UGPVGPEG ½ YCU YGNN HQTOGF *QYGXGT HQT URGGEJ KPRWV VJKU CUUWORVKQP KU PQ OQTG XCNKF 6JGTGHQTG VQ VCMG KPVQ CEEQWPV VJG TGSWKTGOGPV QH ŏYGNNHQTOGFPGUUŏ YG WUG C OQTG EQORNGZ VTCPUNCVKQP OQFGN D[ KP ENWFKPI VJG FGRGPFGPEG QP VJG RTGFGEGUUQT YQTF
½ ½ ½
½
E\&5&3UHVV//&
KP NKGW QH
½
½
Speech Input in Source Language
Acoustic Analysis 6
Z
Z6 ^ H , 2T
Global Search: maximize 2T G +
6 2T Z H ,
6 ^ H ,
, ^ G + 2T H
Lexicon Model
,
2T H ^ G +
+ over G
Acoustic Model
Alignment Model G + 2T
Language Model
G + Translated Text in Target Language
FIGURE 11.6 Integrated architecture of speech translation approach based on Bayes decision rule. (QT VJG UCMG QH UKORNKEKV[ JGTG YG JCXG EJQUGP VJG DKITCO FGRGPFGPEG +V KU KPUVTWEVKXG VQ TGKPVGTRTGV CNTGCF[ GZKUVKPI CRRTQCEJGU HQT JCPFNKPI URGGEJ KPRWV KP C VTCPUNCVKQP VCUM KP VJG NKIJV QH VJG $C[GU FGEKUKQP TWNG HQT URGGEJ VTCPUNCVKQP GXGP KH VJGUG CRRTQCEJGU CTG PQV DCUGF QP UVQEJCUVKE OQFGNNKPI 6JG MG[ KUUWG KP CNN VJGUG CRRTQCEJGU KU VJG SWGUVKQP QH JQY VJG TGSWKTGOGPV QH JCXKPI DQVJ C YGNN HQTOGF UQWTEG UGPVGPEG ½ CPF C YGNNHQTOGF VCTIGV UGPVGPEG ½ CV VJG UCOG VKOG KU UCVKUſGF (TQO VJG UVCVKUVKECN RQKPV QH XKGY VJKU SWGUVKQP KU ECRVWTGF D[ ſPFKPI UWKVCDNG OQFGNU HQT VJG joint RTQDCDKNKV[ ½ ½ ½ ½ ½ (TQO VJG FGEKUKQP TWNG KV KU ENGCT VJCV VJG VTCPUNCVKQP RTQEGUU YKNN JCXG CP GHHGEV QP VJG TGEQIPKVKQP RTQEGUU QPN[ KH VJG VCTIGV NCPIWCIG OQFGN ½ KU UWHſEKGPVN[ UVTQPI QT VQ DG OQTG GZCEV KH KVU UVTGPIVJ KU EQORCTCDNG VQ VJCV QH VJG UQWTEG NCPIWCIG OQFGN ½ 9G OGPVKQP VJG HQNNQYKPI CRRTQCEJGU
+P OCP[ U[UVGOU VJG OGVJQF QH PDGUV NKUVU KU WUGF 6JG TGEQIPK\GT RTQFWEGU C NKUV QH P DGUV UQWTEG UGPVGPEGU CPF VJG VTCPUNCVKQP U[UVGO YQTMU CU C ſNVGT VJCV UGNGEVU QPG QWV QH VJG P UGPVGPEGU WUKPI UQOG UWKVCDNG ETKVGTKQP 6JKU LQKPV
E\&5&3UHVV//&
IGPGTCVKQP CPF ſNVGTKPI RTQEGUU ECP DG XKGYGF CU C ETWFG CRRTQZKOCVKQP QH VJG LQKPV RTQDCDKNKV[ ½ ½
¯ 9JGP WUKPI ſPKVGUVCVG OGVJQFQNQI[ TCVJGT VJCP C HWNN[ UVQEJCUVKE CRRTQCEJ VJG RTQDCDKNKV[ ½ ½ KU OQFGNNGF D[ VJG ſPKVGUVCVG PGVYQTM QH VJG EQT TGURQPFKPI VTCPUFWEGT YJKEJ KU V[RKECNN[ TGſPGF D[ FQOCKP CPF TCPIG TGUVTKE VKQPU = ? ¯ +P VJG GZVTGOG ECUG YG OKIJV DG QPN[ KPVGTGUVGF KP VJG meaning QH VJG VCTIGV VTCPUNCVKQP 5WEJ CP CRRTQCEJ YCU WUGF KP =? HQT VJG 8GTDOQDKN VCUM +P $C[GU FGEKUKQP TWNG VJKU ECUG KU ECRVWTGF D[ RWVVKPI OQUV GORJCUKU QP C semantically EQPUVTCKPGF NCPIWCIG OQFGN ½ +P CFFKVKQP EQPſFGPEG OGC UWTGU =? ECP DG WUGF VQ ſNVGT QWV VJQUG YQTFU VJCV CTG OQUV NKMGN[ VQ JCXG DGGP TGEQIPK\GF EQTTGEVN[ *QYGXGT KV KU ENGCT VJCV PQPG QH VJGUG CRRTQCEJGU HWNN[ KORNGOGPVU VJG KPVGITCVGF EQWRNKPI QH TGEQIPKVKQP CPF VTCPUNCVKQP HTQO C UVCVKUVKECN RQKPV QH XKGY 9G EQPUKFGT VJKU KPVGITCVGF CRRTQCEJ CPF KVU UWKVCDNG KORNGOGPVCVKQP VQ DG CP QRGP SWGUVKQP HQT HWVWTG TGUGCTEJ QP URQMGP NCPIWCIG VTCPUNCVKQP 9JCV YG JCXG EQPUKFGTGF JGTG KU URGGEJ KPRWV KP VJG source NCPIWCIG +P OCEJKPG CKFGF VTCPUNCVKQP VJG URGGEJ KPRWV KU KP VJG target NCPIWCIG CPF VJWU VJG UQWTEG UGPVGPEG KU V[RKECNN[ WUGF VQ EJCPIG VJG NCPIWCIG OQFGN KP VJG VCTIGV NCPIWCIG = ?
11.7 Summary +P VJKU EJCRVGT YG JCXG IKXGP CP QXGTXKGY QH VJG UVCVKUVKECN CRRTQCEJ VQ OCEJKPG VTCPUNCVKQP CPF GURGEKCNN[ KVU KORNGOGPVCVKQP KP VJG 8 '4$/1$+. RTQVQV[RG U[UVGO 6JG UVCVKUVKECN U[UVGO JCU DGGP VTCKPGF QP CDQWV TWPPKPI YQTFU HTQO C DKNKP IWCN )GTOCPŌ'PINKUJ EQTRWU 6TCPUNCVKQPU CTG RGTHQTOGF HQT DQVJ FKTGEVKQPU KG HTQO )GTOCP VQ 'PINKUJ CPF HTQO 'PINKUJ VQ )GTOCP %QORCTCVKXG GXCNWCVKQPU YKVJ QVJGT VTCPUNCVKQP CRRTQCEJGU QH VJG 8 '4$/1$+. RTQVQV[RG U[UVGO UJQY VJCV VJG UVCVKUVKECN VTCPUNCVKQP KU UWRGTKQT GURGEKCNN[ KP VJG RTGUGPEG QH URGGEJ KPRWV CPF WPITCOOCVKECN KPRWV +P CFFKVKQP YG JCXG RTGUGPVGF VJG HWNN[ KPVGITCVGF CRRTQCEJ VQ URQMGP NCPIWCIG VTCPUNCVKQP
Acknowledgment 6JG CWVJQTU YQWNF NKMG VQ VJCPM VJG TGUGCTEJGTU KP +6+ 8CNGPEKC CPF #CEJGP 7PKXGT UKV[ QH 6GEJPQNQI[ YJQ RCTVKEKRCVGF KP VJG RTQLGEVU 8 '4$/1$+. CPF ' 7 6 4#05 CPF FGXGNQRGF VJG CRRTQCEJGU RTGUGPVGF KP VJKU EJCRVGT +P RCTVKEWNCT VJG CWVJQTU YQWNF NKMG VQ VJCPM 'PTKSWG 8KFCN CPF (TCPEKUEQ %CUCEWDGTVC HQT OCP[ FKUEWUUKQPU
E\&5&3UHVV//&
6JG YQTM TGRQTVGF JGTG YCU RCTVN[ ECTTKGF QWV KP VJG 8 '4$/1$+. RTQLGEV EQPVTCEV PWODGT +8 6 HWPFGF D[ VJG )GTOCP (GFGTCN /KPKUVT[ QH 'FWECVKQP 5EKGPEG 4GUGCTEJ CPF 6GEJPQNQI[ CPF KP VJG ' 7 6 4#05 RTQLGEV EQPVTCEV PWODGT +6.6415 HWPFGF D[ VJG 'WTQRGCP 7PKQP
11.8 References =? * #NUJCYK ( :KCPI 'PINKUJVQ/CPFCTKP URGGEJ VTCPUNCVKQP YKVJ JGCF VTCPU FWEGTU Spoken Language Translation Workshop, 35th Annual Conf. of the Assoc. for Computational Linguistics RR Ō /CFTKF 5RCKP ,WN[ =? * #NUJCYK 5 $CPICNQTG 5 &QWINCU .GCTPKPI FGRGPFGPE[ VTCPUNCVKQP OQF GNU CU EQNNGEVKQP QH ſPKVGUVCVG JGCF VTCPUFWEGTU Computational Linguistics 8QN 0Q RR Ō =? / #WGTUYCNF 'ZCORNGDCUGF OCEJKPG VTCPUNCVKQP YKVJ VGORNCVGU +P =? RR Ō =? 5 $CPICNQTG ) 4KEECTFK (KPKVGUVCVG OQFGNU HQT NGZKECN TGQTFGTKPI KP URQ MGP NCPIWCIG VTCPUNCVKQP Int. Conf. on Spoken Language Processing, 8QN +8 RR Ō $GLKPI %JKPC 1EV =? # $CVNKPGT , $WEMQY * 0KGOCPP ' 0iQVJ 8 9CTPMG 6JG RTQUQF[ OQFWNG +P =? RR Ō =? 6 $GEMGT # -KNIGT 2 .QRG\ 2 2QNNGT 6JG 8GTDOQDKN IGPGTCVKQP EQORQPGPV 8/)'%1 +P =? RR Ō =? # . $GTIGT 2 ( $TQYP , %QEMG 5 # &GNNC 2KGVTC 8 , &GNNC 2KGVTC , 4 )KNNGVV , & .CHHGTV[ 4 . /GTEGT * 2TKPV\ . 7TGU 6JG %CPFKFG U[U VGO HQT OCEJKPG VTCPUNCVKQP ARPA Human Language Technology Workshop 2NCKPUDQTQ 0, /QTICP -CWHOCPP 2WDNKUJGTU RR Ō 5CP /CVGQ %# /CTEJ =? # . $GTIGT 2 ( $TQYP 5 # &GNNC 2KGVTC 8 , &GNNC 2KGVTC , 4 )KNNGVV # 5 -GJNGT .CPIWCIG VTCPUNCVKQP CRRCTCVWU CPF OGVJQF QH WUKPI EQPVGZV DCUGF VTCPUNCVKQP OQFGNU United States Patent 2CVGPV 0WODGT #RTKN =? # . $GTIGT 5 &GNNC 2KGVTC 8 &GNNC 2KGVTC # OCZKOWO GPVTQR[ CRRTQCEJ VQ PCVWTCN NCPIWCIG RTQEGUUKPI Computational Linguistics 8QN 0Q RR Ō =? 7 $NQEM 'ZCORNGDCUGF KPETGOGPVCN U[PEJTQPQWU KPVGTRTGVCVKQP +P =? RR Ō
E\&5&3UHVV//&
=? 2 ( $TQYP 5 ( %JGP 5 # &GNNC 2KGVTC 8 , &GNNC 2KGVTC # 5 -GJNGT 4 . /GTEGT #WVQOCVKE URGGEJ TGEQIPKVKQP KP OCEJKPGCKFGF VTCPUNCVKQP Computer Speech and Language 8QN RR Ō =? 2 ( $TQYP 5 # &GNNC 2KGVTC 8 , &GNNC 2KGVTC 4 . /GTEGT 6JG OCVJG OCVKEU QH UVCVKUVKECN OCEJKPG VTCPUNCVKQP 2CTCOGVGT GUVKOCVKQP Computational Linguistics 8QN 0Q RR Ō =? 2 ( $TQYP 8 , &GNNC 2KGVTC 2 8 FG5QW\C , % .CK 4 . /GTEGT %NCUUŌ DCUGF ŌITCO OQFGNU QH PCVWTCN NCPIWCIG Computational Linguistics 8QN 0Q RR Ō =? ( %CUCEWDGTVC (KPKVGUVCVG VTCPUFWEGTU HQT URGGEJ KPRWV VTCPUNCVKQP IEEE Automatic Speech Recognition and Understanding Workshop /CFQPPC FK %CORKINKQ +VCN[ RCIGU %& 41/ +''' %CVCNQI 0Q ': &GE =? 0 %JQOUM[ 3WKPGŏU GORKTKECN CUUWORVKQPU R KP & &CXKFUQP , *KP VKMMC GFU Words and objections. Essays on the work of W. V. Quine 4GKFGN &QTFTGEJV 6JG 0GVJGTNCPFU =? + &CICP - %JWTEJ 9 # )CNG 4QDWUV DKNKPIWCN YQTF CNKIPOGPV HQT OCEJKPGCKFGF VTCPUNCVKQP Workshop on Very Large Corpora RRŌ %QNWO DWU 1* ,WPG =? / &[OGVOCP , $TQWUUGCW ) (QUVGT 2 +UCDGNNG ; 0QTOCPFKP 2 2NCO QPFQP 6QYCTFU CP CWVQOCVKE FKEVCVKQP U[UVGO HQT VTCPUNCVQTU VJG 6TCPU6CNM RTQLGEV Int. Conf. on Spoken Language Processing, 8QN ++ RR Ō ;QMQ JCOC ,CRCP 5GR =? / % 'OGNG / &QTPC # .iWFGNKPI *
496* #CEJGP )GTOCP[
E\&5&3UHVV//&
=? 4 -PGUGT * 0G[ +ORTQXGF ENWUVGTKPI VGEJPKSWGU HQT ENCUUŌDCUGF UVCVKUVKECN NCPIWCIG OQFGNNKPI Europ. Conf. on Speech Communication and Technology, RR Ō $GTNKP )GTOCP[ 5GRV =? - -PKIJV &GEQFKPI EQORNGZKV[ KP YQTFTGRNCEGOGPV VTCPUNCVKQP OQFGNU Computational Linguistics 0Q 8QN RR Ō =? # .CXKG . .GXKP # 9CKDGN & )CVGU / )CXCNFC . /C[ſGNF ,#075 /WNVKNKPIWCN VTCPUNCVKQP QH URQPVCPGQWU URGGEJ KP C NKOKVGF FQOCKP 2nd Conf. of the Assoc. for Machine Translation in the Americas RR /QPVTGCN %CPCFC 1EV =? * 0G[ 5RGGEJ VTCPUNCVKQP EQWRNKPI QH TGEQIPKVKQP CPF VTCPUNCVKQP IEEE Int. Conf. on Acoustics, Speech and Signal Processing, RR + 2JQGPKZ #4 /CTEJ =? * 0G[ 5 0KG²GP ( , 1EJ * 5CYCH % 6KNNOCPP 5 8QIGN #NIQTKVJOU HQT UVCVKUVKECN VTCPUNCVKQP QH URQMGP NCPIWCIG IEEE Trans. on Speech and Audio Processing 8QN 0Q RR Ō ,CP =? 5 0KG²GP * 0G[ +ORTQXKPI 5/6 SWCNKV[ YKVJ OQTRJQU[PVCEVKE CPCN[UKU 18th Int. Conf. on Computational Linguistics RR Ō 5CCTDTiWEMGP )GTOCP[ ,WN[ =? 5 0KG²GP ( , 1EJ ) .GWUEJ * 0G[ #P GXCNWCVKQP VQQN HQT OCEJKPG VTCPU NCVKQP HCUV GXCNWCVKQP HQT /6 TGUGCTEJ 2nd Int. Conf. on Language Resources and Evaluation RR Ō #VJGPU )TGGEG /C[ =? 5 0KG²GP 5 8QIGN * 0G[ % 6KNNOCPP # &2 DCUGF UGCTEJ CNIQTKVJO HQT UVCVKUVKECN OCEJKPG VTCPUNCVKQP 36th Annual Meeting of the Assoc. for Computational Linguistics and 17th Int. Conf. on Computational Linguistics RR Ō /QPVTGCN %CPCFC #WI =? ( , 1EJ #P GHſEKGPV OGVJQF VQ FGVGTOKPG DKNKPIWCN YQTF ENCUUGU 9th Conf. of the Europ. Chapter of the Assoc. for Computational Linguistics RR Ō $GTIGP 0QTYC[ ,WPG =? ( , 1EJ * 0G[ # EQORCTKUQP QH CNKIPOGPV OQFGNU HQT UVCVKUVKECN OC EJKPG VTCPUNCVKQP 18th Int. Conf. on Computational Linguistics RR Ō 5CCTDTiWEMGP )GTOCP[ ,WN[ =? ( , 1EJ * 0G[ +ORTQXGF UVCVKUVKECN CNKIPOGPV OQFGNU 38th Annual Meeting of the Assoc. for Computational Linguistics RR *QPI -QPI 1EV =? ( , 1EJ % 6KNNOCPP * 0G[ +ORTQXGF CNKIPOGPV OQFGNU HQT UVCVKUVKECN OC EJKPG VTCPUNCVKQP Joint SIGDAT Conf. on Empirical Methods in Natural Language Processing and Very Large Corpora RR Ō 7PKXGTUKV[ QH /CT[NCPF %QNNGIG 2CTM /& ,WPG
E\&5&3UHVV//&
=? ( , 1EJ 0 7GHſPI * 0G[ #P GHſEKGPV # UGCTEJ CNIQTKVJO HQT UVCVKU VKECN OCEJKPG VTCPUNCVKQP Data-Driven Machine Translation Workshop, 39th Annual Meeting of the Assoc. for Computational Linguistics RR Ō 6QWNQWUG (TCPEG ,WN[ =? - 2CRKPGPK 5 4QWMQU 6 9CTF 9, <JW $.'7 C OGVJQF HQT CWVQOCVKE GXCNWCVKQP QH OCEJKPG VTCPUNCVKQP IBM Research Report RCIGU ;QTMVQYP *GKIJVU 0; 5GRV =? . 4 4CDKPGT $ * ,WCPI Fundamentals of speech recognition 2TGPVKEG *CNN 'PINGYQQF %NKHHU 0, =? 0 4GKVJKPIGT 4 'PIGN 4QDWUV EQPVGPV GZVTCEVKQP HQT VTCPUNCVKQP CPF FKCNQI RTQEGUUKPI +P =? RR Ō =? 5 & 4KEJCTFUQP 9 $ &QNCP # /GPG\GU / %QTUVQP1NKXKGT 1XGTEQOKPI VJG EWUVQOK\CVKQP DQVVNGPGEM WUKPI GZCORNGDCUGF /6 Data-Driven Machine Translation Workshop, 39th Annual Meeting of the Assoc. for Computational Linguistics RR Ō 6QWNQWUG (TCPEG ,WN[ =? * 5CYCH - 5EJiWV\ * 0G[ 1P VJG WUG QH ITCOOCT DCUGF NCPIWCIG OQFGNU HQT UVCVKUVKECN OCEJKPG VTCPUNCVKQP 6th Int. Workshop on Parsing Technologies RR Ō 6TGPVQ +VCN[ (GD =? , 5RKNMGT / -NCTPGT ) )iQT\ 2TQEGUUKPI UGNHEQTTGEVKQPU KP C URGGEJVQ URGGEJ U[UVGO +P =? RR Ō =? ' 5WOKVC 'ZCORNGDCUGF OCEJKPG VTCPUNCVKQP WUKPI &2OCVEJKPI DGVYGGP YQTF UGSWGPEGU Data-Driven Machine Translation Workshop, 39th Annual Meeting of the Assoc. for Computational Linguistics RR Ō 6QWNQWUG (TCPEG ,WN[ =? . 6GUUKQTG 9 8 *CJP (WPEVKQPCN XCNKFCVKQP QH C OCEJKPG VTCPUNCVKQP U[U VGO 8GTDOQDKN +P =? RR Ō =? % 6KNNOCPP * 0G[ 9QTF TGQTFGTKPI KP C &2DCUGF CRRTQCEJ VQ UVCVKU VKECN /6 18th Int. Conf. on Computational Linguistics 2000 RR Ō 5CCTDTiWEMGP )GTOCP[ #WI =? * 7U\MQTGKV & (NKEMKPIGT 9 -CURGT + # 5CI &GGR NKPIWKUVKE CPCN[UKU YKVJ *25) +P =? RR Ō =? ' 8KFCN (KPKVGUVCVG URGGEJVQURGGEJ VTCPUNCVKQP IEEE Int. Conf. on Acoustics, Speech and Signal Processing RR Ō /WPKEJ )GTOCP[ #RTKN =? 5 8QIGN * 0G[ 6TCPUNCVKQP YKVJ ECUECFGF ſPKVGUVCVG VTCPUFWEGTU 38th Annual Meeting of the Assoc. for Computational Linguistics RR Ō *QPI -QPI 1EV
E\&5&3UHVV//&
=? 5 8QIGN * 0G[ % 6KNNOCPP *//DCUGF YQTF CNKIPOGPV KP UVCVKUVKECN VTCPUNCVKQP 16th Int. Conf. on Computational Linguistics RR Ō %QRGP JCIGP &GPOCTM #WIWUV =? 9 9CJNUVGT 'F Verbmobil: Foundations of speech-to-speech translation. 5RTKPIGT8GTNCI $GTNKP )GTOCP[ =? ;; 9CPI # 9CKDGN &GEQFKPI CNIQTKVJO KP UVCVKUVKECN VTCPUNCVKQP 35th Annual Conf. of the Assoc. for Computational Linguistics RR Ō /CFTKF 5RCKP ,WN[ =? ( 9GUUGN 4 5EJNiWVGT - /CEJGTG[ * 0G[ %QPſFGPEG OGCUWTGU HQT NCTIG XQECDWNCT[ EQPVKPWQWU URGGEJ TGEQIPKVKQP IEEE Trans. on Speech and Audio Processing, 8QN 0Q RR Ō /CTEJ =? & 9W 5VQEJCUVKE KPXGTUKQP VTCPUFWEVKQP ITCOOCTU CPF DKNKPIWCN RCTUKPI QH RCTCNNGN EQTRQTC Computational Linguistics 8QN 0Q RR Ō =? - ;COCFC - -PKIJV # U[PVCZDCUGF UVCVKUVKECN VTCPUNCVKQP OQFGN 39th Annual Meeting of the Assoc. for Computational Linguistics, RR Ō 6QWNQWUG (TCPEG ,WN[
E\&5&3UHVV//&
12 Modeling Topics for Detection and Tracking James Allan University of Massachusetts Amherst
CONTENTS
6QRKE &GVGEVKQP CPF 6TCEMKPI $CUKE 6QRKE /QFGNU +ORNGOGPVKPI VJG /QFGNU %QORCTKPI /QFGNU /KUEGNNCPGQWU +UUWGU 7UKPI 6&6 +PVGTCEVKXGN[ /QFGNKPI 'XGPVU %QPENWUKQP 4GHGTGPEG
6QRKE FGVGEVKQP CPF VTCEMKPI 6&6 KU C TGUGCTEJ RTQITCO CPF CP GXCNWCVKQP RCTCFKIO VJCV KPXGUVKICVGU VGEJPKSWGU HQT CWVQOCVKECNN[ QTICPK\KPI DTQCFECUV PGYU UVQTKGU D[ VJG GXGPVU VJCV VJG[ FGUETKDG 6&6 KU CP QWVITQYVJ QH KPHQTOCVKQP TGVTKGXCN
+4 VGEJPQNQI[ CPF UJCTGU OCP[ QH KVU VGEJPKSWGU CPF KFGCU 6JG URGEKſE VCUMU YKVJKP GXGPVDCUGF QTICPK\CVKQP CPF VJG PCVWTG QH VJG UVQTKGU VQ YJKEJ VJQUG VCUMU CTG CRRNKGF OGCPU VJCV 6&6 CFOKVU C TCPIG QH CRRTQCEJGU VJCV CTG PQV WPKXGTUCNN[ CRRNKECDNG YKVJKP +4 1PG QH VJG TGUGCTEJ KUUWGU YKVJKP 6&6 KU TGRTGUGPVKPI C PGYU VQRKE DCUGF QP URCTUG KPHQTOCVKQPōHQT GZCORNG DCUGF QP C UKPING UVQT[ 6JKU EJCRVGT FKUEWUUGU UGXGTCN YC[U KP YJKEJ VQRKEU CTG OQFGNGF YKVJKP VJG 6&6 TGUGCTEJ EQOOWPKV[ CPF EQO RCTGU CPF EQPVTCUVU VJGO 9G CTG CIPQUVKE CU VQ YJKEJ KU VJG DGUV OQFGNōTGUGCTEJ UWIIGUVU VJCV OQUV QH VJG CRRTQCEJGU CTG GSWCNN[ GHHGEVKXG *QYGXGT KV KU FKUCRRQKPV KPI VJCV PQ VGEJPKSWGU GZRNKEKVN[ OQFGN VJG GXGPVU QWV QH YJKEJ 6&6 VQRKEU CTKUG 9G YKNN EQPENWFG VJG EJCRVGT YKVJ URGEWNCVKQP CDQWV JQY GXGPVU OKIJV DG OQTG FKTGEVN[ KPEQTRQTCVGF KPVQ VJG VQRKE OQFGNU
12.1 Topic Detection and Tracking 6JG IQCN QH 6&6 TGUGCTEJ KU VQ QTICPK\G PGYU UVQTKGU D[ VJG GXGPVU VJCV VJG[ FGUETKDG CPF VQ FQ VJCV CU UQQP CU VJG UVQTKGU CRRGCT YJGVJGT CU PGYUYKTG VGNGXKUKQP QT TCFKQ
E\&5&3UHVV//&
6JCV KU VJG FGEKUKQP CDQWV JQY VQ JCPFNG C UVQT[ OWUV DG OCFG DGHQTG CP[ CFFKVKQPCN UVQTKGU CTG RTQEGUUGF 6JG 6&6 TGUGCTEJ RTQITCO DGICP KP CU C EQNNCDQTCVKQP DGVYGGP %CTPGIKG /GN NQP 7PKXGTUKV[ &TCIQP 5[UVGOU VJG 7PKXGTUKV[ QH /CUUCEJWUGVVU CPF *# =? 6JCV ITQWR QH TGUGCTEJGTU TCP C RKNQV UVWF[ VJCV FGſPGF VJG DCUKE VCUMU QH 6&6 CPF JQY VJG[ UJQWNF DG GXCNWCVGF 6Q ſPF QWV JQY YGNN ENCUUKE +4 VGEJPQNQIKGU CF FTGUUGF 6&6 VJG[ ETGCVGF C UOCNN EQNNGEVKQP QH PGYU UVQTKGU CPF KFGPVKſGF UQOG VQRKEU YKVJKP VJGO $GECWUG VJG TGUWNVU YGTG GPEQWTCIKPI C NCTIGT CPF OQTG HQTOCN UGTKGU QH GXCNWCVKQPU YGTG JGNF GXGT[ [GCT HTQO VJTQWIJ CPF EQPVKPWGU KP CPF RGTJCRU NQPIGT 6JGUG GXCNWCVKQPU OQTG ECTGHWNN[ FGſPGF VJG PQVKQPU QH VQRKE CPF GXGPV FGXGNQRGF VJG UGV QH VCUMU OQTG HWNN[ CPF EQPUVTWEVGF C NCTIGT CPF TKEJGT EQTRWU QH UVQTKGU CPF VQRKEU
12.1.1 Topic and Events #P event YCU FGſPGF CU UQOGVJKPI VJCV JCRRGPU CV UQOG URGEKſE VKOG CPF RNCEG CNQPI YKVJ CNN PGEGUUCT[ RTGEQPFKVKQPU CPF WPCXQKFCDNG EQPUGSWGPEGU 6JCV KU KV KU UQOGVJKPI VJCV JCRRGPU KP VJG TGCN YQTNF # RCTVKEWNCT GCTVJSWCMG KU CP GXGPV CU KU VJG FKUEQXGT[ QH C PGY EQOGV # topic KU OGCPV VQ ECRVWTG VJG NCTIGT UGV QH JCRRGPKPIU VJCV CTG TGNCVGF VQ UQOG VTKIIGTKPI GXGPV 6JG QHſEKCN FGſPKVKQP QH C VQRKE KU VJCV KV KU C UGOKPCN GXGPV CNQPI YKVJ CNN FKTGEVN[ TGNCVGF GXGPVU CPF CEVKXKVKGU $[ HQTEKPI VJG CFFKVKQPCN GXGPVU VQ DG FKTGEVN[ TGNCVGF VJG VQRKE KU RTGXGPVGF HTQO URTGCFKPI QWV VQ KPENWFG VQQ OWEJ PGYU 1PG YC[ VQ VJKPM QH C VQRKE KU VJCV IKXGP C UVCTVKPI GXGPV KV KPENWFGU VJG CFFKVKQPCN GXGPVU VJCV C V[RKECN TGCFGT YQWNF GZRGEV VQ UGG KP HQNNQYWR PGYU 1T IKXGP CP GXGPV VJCV KU PQV QH KPVGTGUV VJG UGV QH HQNNQYKPI GXGPVU VJCV VJG TGCFGT YQWNF RTGHGT not VQ UGG 6QRKEU VJGP CTG CPEJQTGF KP VKOG CPF URCEG D[ VJG UGOKPCN GXGPV 6JG HQEWU QP GXGPVU CPF VJG VKIJV TGNCVKQP VQ VKOG FKUVKPIWKUJ 6&6 VQRKEU HTQO VJG OQTG IGPGTCN WUG QH VJG YQTF őVQRKEŒ YKVJKP KPHQTOCVKQP TGVTKGXCN +P VJCV UGVVKPI C VQRKE KU WUWCNN[ UWDLGEVDCUGF KV TGRTGUGPVU CP CTGC QH KPVGTGUV VQ VJG UGCTEJGT 5QOG UWDLGEVDCUGF VQRKEU CTG KFGPVKECN VQ 6&6 VQRKEU ő+ CO KPVGTGUVGF KP KPHQTOCVKQP CDQWV VJG -QDG GCTVJSWCMG QH Œ DWV UQOG JCXG PQ RCTCNNGN KP 6&6 ő6GNN OG CDQWV GPFCPIGTGF URGEKGU KP #HTKECŒ 6JG FKHHGTGPEGU DGVYGGP 6&6 VQRKEU CPF +4 VQRKEU OGCPU VJCV FKHHGTGPV VGEJPKSWGU UJQWNF DG WUGHWN VQ CFFTGUU VJGKT TGURGEVKXG VCUMU 6JG UKOKNCTKVKGU CTG UWHſEKGPV GPQWIJ VJQWIJ VJCV OQUV TGUGCTEJ JCU HQEWUGF QP VJG FKTGEV CRRNKECVKQP QH +4 OGVJ QFU VQ 6&6 VCUMU
12.1.2 TDT Tasks 6JG 6&6 GXCNWCVKQP RTQITCO FGſPGU ſXG VCUMU HQT QTICPK\KPI PGYU D[ GXGPVU UGI OGPVCVKQP ENWUVGT FGVGEVKQP VTCEMKPI PGY GXGPV FGVGEVKQP CPF NKPM FGVGEVKQP
E\&5&3UHVV//&
12.1.2.1 Segmentation 0GYU VJCV CTTKXGU XKC PGYUYKTG KU FKXKFGF KPVQ KPFKXKFWCN UVQTKGU YJGTGCU VGNGXKUKQP CPF TCFKQ PGYU KU PQV 6JG VCUM QH UGIOGPVCVKQP KU VQ DTGCM CP CWFKQ VTCEM KPVQ FKUETGVG UVQTKGU GCEJ QP C UKPING VQRKE /QUV TGUGCTEJ QP VJKU RTQDNGO JCU WUGF URGGEJ TGEQIPK\GT QWVRWV CU C UVCTVKPI RQKPV TCVJGT VJCP YQTMKPI QP VJG CWFKQ KVUGNH 6JKU VCUM KU C PGEGUUCT[ RTGEQPFKVKQP VQ CNN QH VJG QVJGT VCUMU UKPEG VJG[ CUUWOG C UGV QH UVQTKGU VJCV PGGF VQ DG QTICPK\GF 9G YKNN PQV VCNM CDQWV UGIOGPVCVKQP KP VJG TGUV QH VJKU EJCRVGT +V JCU DGGP FKUEWUUGF GNUGYJGTG = ? 12.1.2.2 Cluster Detection +P VJG ENWUVGT FGVGEVKQP VCUM CNUQ TGHGTTGF VQ UKORN[ CU őFGVGEVKQPŒ C U[UVGO OWUV RNCEG CNN CTTKXKPI PGYU UVQTKGU KPVQ ITQWRU DCUGF QP VJGKT VQRKEU +H PQ GZKUVKPI ITQWRŏU VQRKE OCVEJGU VJG UVQT[ UWHſEKGPVN[ VJG U[UVGO OWUV FGEKFG YJGVJGT VQ ETGCVG C PGY ITQWR 6JG FGEKUKQP CDQWV JQY VQ RTQEGUU CP KPFKXKFWCN UVQT[ OWUV DG OCFG DGHQTG VJG PGZV UVQT[ KU EQPUKFGTGF 6JG 6&6 GXCNWCVKQP RTQITCO TGSWKTGU VJCV VJG GCEJ UVQT[ DG RNCEGF KP RTGEKUGN[ QPG ENWUVGT KORN[KPI VJCV GCEJ UVQT[ KU CDQWV C UKPING VQRKE 6JKU UKORNKH[KPI CUUWORVKQP YCU WUGHWN KP GCTN[ GXCNWCVKQPU CPF KU DGKPI FTQRRGF CHVGT 6&6 12.1.2.3 Tracking 6JG VCUM QH VTCEMKPI UVCTVU YKVJ C UOCNN UGV QH PGYU UVQTKGU VJCV C WUGT JCU KFGPVKſGF CU DGKPI QP VJG UCOG VQRKE )KXGP VJCV UGV VJG U[UVGO OWUV OQPKVQT VJG UVTGCO QH CTTKXKPI PGYU VQ ſPF CNN CFFKVKQPCN UVQTKGU QP VJG UCOG VQRKE #U YKVJ FGVGEVKQP C FGEKUKQP OWUV DG OCFG CDQWV GCEJ UVQT[ DGHQTG CFFKVKQPCN UVQTKGU ECP DG UGGP (WTVJGT VJG U[UVGO KU PGXGT IKXGP HGGFDCEM QVJGT VJCP C ſPCN GXCNWCVKQP CDQWV YJGVJGT KV JCU OCFG C EQTTGEV FGEKUKQP 6JKU VCUM KU CPCNQIQWU VQ VJG +4 ſNVGTKPI RTQDNGO =? &KHHGTGPEGU DGVYGGP VJG VYQ NKG RTKOCTKN[ KP VJG FGſPKVKQP QH VQRKE CU KP 5GEVKQP CPF KP FKHHGTGPV GXCNWCVKQP RCTCFKIOU = ? 12.1.2.4 New Event Detection 6JG VCUM QH PGY GXGPV FGVGEVKQP HQEWUGU QP VJG ENWUVGT ETGCVKQP CURGEV QH ENWUVGT FGVGEVKQP # U[UVGO KU GXCNWCVGF GPVKTGN[ QP KVU CDKNKV[ VQ FGEKFG YJGP C PGY VQRKE
GXGPV CRRGCTU 9JGVJGT QT PQV VJG TGOCKPKPI UVQTKGU KP VJG VQRKE CTG RTQRGTN[ RNCEGF KP VJGKT VQRKEU KU WPKORQTVCPV #U WUWCN FGEKUKQPU OWUV DG OCFG CU UVQTKGU CTTKXG 6JKU VCUM KU CNUQ TGHGTTGF VQ CU őſTUV UVQT[ FGVGEVKQPŒ 12.1.2.5 Link Detection 6JG ſPCN 6&6 VCUM NKPM FGVGEVKQP YCU ETGCVGF CU C EQTG VGEJPQNQI[ HQT VJG QVJGT VCUMU 6JG KFGC KU VQ FGVGTOKPG YJGVJGT QT PQV VYQ TCPFQON[ RTGUGPVGF UVQTKGU FKUEWUU VJG UCOG VQRKE # UQNWVKQP VQ VJKU VCUM EQWNF DG WUGF VQ UQNXG PGY GXGPV FGVGEVKQP
E\&5&3UHVV//&
HQT GZCORNG D[ EQORCTKPI VJG PGYN[ CTTKXGF UVQT[ VQ GXGT[ UVQT[ KP VJG RCUV +H PQ GCTNKGT UVQT[ JCF VJG UCOG VQRKE VJGP C PGY VQRKE ECP DG FGENCTGF +V KU PQV PGEGUUCT[ VJCV NKPM FGVGEVKQP DG VJG VGEJPQNQI[ VQ CFFTGUU QVJGT VCUMU DWV OQUV CRRTQCEJGU VQ 6&6 RTQDNGOU WUG KFGCU UKOKNCT VQ VJKU VCUM
12.1.3 Corpora 6JG EQTRWU HQT VJG RKNQV UVWF[ KPENWFGF CDQWV UVQTKGU HTQO VYQ UQWTEGU %00 CPF 4GWVGTU ICVJGTGF HTQO VJG NCUV JCNH QH CPF VJG ſTUV JCNH QH +P CFFKVKQP VQRKEU YGTG UGNGEVGF D[ VJG TGUGCTEJGTU VJCV GCEJ UVQT[ YCU LWFIGF CICKPUV 6JG EQPUVTWEVKQP QH VJG EQTRWU CPF VJG HQTOCVKQP QH VJG VQRKEU YCU UWHſ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ſ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ſTUV YGTG FGXGNQRGF HQT VJG GXCNWCVKQP CPF JCXG VJG WPWUWCN TGSWKTGOGPV VJCV VJGTG OWUV DG CV NGCUV HQWT QPVQRKE UVQTKGU KP each QH 'PINKUJ CPF %JKPGUG 6JG QVJGT VQRKEU YGTG FGXGNQRGF HQT VJG GXCNWCVKQP CPF TGOQXG VJCV TGSWKTGOGPVōJQYGXGT VJG[ YGTG UGGFGF GSWCNN[ HTQO 'PINKUJ CPF %JK PGUG UVQTKGU UQ GCEJ NCPIWCIG KU TGRTGUGPVGF 6JKU NCVVGT UGV QH VQRKEU KU DGKPI LWFIGF CICKPUV VJG PGYN[ CFFGF #TCDKE UQWTEGU +P CFFKVKQP VJG LWFIOGPVU HQT VJG UGEQPF UGV QH VQRKEU YGTG PQV FQPG D[ EQORNGVG TGXKGYKPI QH GXGT[ UVQT[ HQT GCEJ VQRKE +PUVGCF CU C EQUV UCXKPI OGCUWTG VJG CPPQVCVKQPU YGTG OCFG WUKPI JWOCPIWKFGF UGCTEJ VGEJPKSWGU GZRGTKOGPVU UJQYGF GSWCN CEEWTCE[ DGVYGGP VJG CRRTQCEJGU +P VJG GXCNWCVKQP VQRKEU HTQO GCEJ UGV YGTG EJQUGP HQT VJG GXCNWCVKQP 6JG 6&6 EQTRWU YKNN DG WUGF CU VTCKPKPI FCVC HQT VJG 6&6 GXCNWCVKQP 6JG NCVGUV 6&6 EQTRWU KU 6&6 DGKPI ETGCVGF KP VJG HCNN QH CPF URTKPI QH +V KPENWFGU CRRTQZKOCVGN[ UVQTKGU EQXGTKPI 1EVQDGT VJTQWIJ ,CP 6JKU KU KP EQPVTCUV VQ V[RKECN +4 LWFIOGPVU VJCV CTG QPN[ LWFIGF CICKPUV UVQTKGU TGVTKGXGF D[ UQOG U[UVGO RCTVKEKRCVKPI KP VJG GXCNWCVKQP 6JG TGCUQP VJCV 6&6 ECP OCPCIG VJG EQORNGVG LWFIOGPV UGV KU DGECWUG VJG EQTRWU KU UWDUVCPVKCNN[ UOCNNGT VJCP C V[RKECN +4 EQNNGEVKQP
E\&5&3UHVV//&
WCT[ 6JG OWNVKNKPIWCN CURGEV QH 6&6 KU DGKPI UVTGUUGF OQTG KP VJKU EQTRWU UQ VJG UVQTKGU EQOG HTQO GKIJV 'PINKUJ UQWTEGU UGXGP %JKPGUG UQWTEGU CPF HQWT #TCDKE UQWTEGU #FFKVKQPCN NCPIWCIGU YGTG EQNNGEVGF KP RCTCNNGN DWV CTG PQV DGKPI KPENWFGF YKVJ VJG 6&6 EQTRWU CV VJKU VKOG 5KZV[ PGY VQRKEU CTG DGKPI FGXGNQRGF HTQO VJKU FCVC WUKPI C OQFGN UKOKNCT VQ VJCV QH VJG UGEQPF UKZV[ VQRKEU KP 6&6 KG UGGFGF GSWCNN[ HTQO GCEJ NCPIWCIG 6JKU EQTRWU YKNN DG WUGF HQT VJG 6&6 GXCNWCVKQP +P CNN VJTGG EQTRQTC CWFKQ UQWTEGU YGTG RCUUGF VJTQWIJ C URGGEJ TGEQIPKVKQP U[UVGO CPF VJG QWVRWV KU KPENWFGF KP VJG EQTRWU +P CFFKVKQP C TGHGTGPEG ENQUGFECRVKQP SWCNKV[ VTCPUETKRV YCU OCFG KH KV YCU PQV CXCKNCDNG GI HQT TCFKQ UQWTEGU 0QP 'PINKUJ UQWTEGU YGTG TGEQIPK\GF HQT CWFKQ CPF VJGP VTCPUNCVGF VQ 'PINKUJ WUKPI VJG 5;564#0Ý U[UVGO 6JG UQWTEG NCPIWCIG CPF VJG VTCPUNCVKQP YGTG OCFG CXCKNCDNG VQ CNN UKVGU
12.1.4 Evaluation #NN 6&6 VCUMU CTG GPXKUKQPGF CU őQPNKPGŒ VCUMU VJCV OWUV EQORNGVGN[ RTQEGUU GCEJ UVQT[ DGHQTG TGEGKXKPI CP[ CFFKVKQPCN UVQTKGU &GEKUKQPU CTG KTTGXQECDNG GXGP KH C OKUVCMG KU FGVGEVGF NCVGT 6JKU CRRTQCEJ OQFGNU C UKVWCVKQP YJGTG VJG QWVRWV KU EQP UWOGF KOOGFKCVGN[ CPF KP C VKOGETKVKECN HCUJKQP +V GZRNQTGU VJG EQTG VGEJPQNQI[ TCVJGT VJCP JQY KV OKIJV DG WUGF KP CP KPVGTCEVKXG UGVVKPI =? 1WVRWV KU EQWEJGF KP VGTOU QH C FGVGEVKQP VCUM YJGTG ő[GUŒ QT őPQŒ FGEKUKQPU OWUV DG OCFG =? 'XCNW CVKQP KU KP VGTOU QH GTTQTU OKUUGU CPF HCNUG CNCTOU CPF VJG VTCFGQHH DGVYGGP VJGO (KIWTG UJQYU C UCORNG FGVGEVKQP GTTQT VTCFGQHH &'6 ITCRJ =? HQT C 6&6 VCUM 6JG HCNUG CNCTO TCVG KU UJQYP QP VJG :CZKU CPF VJG OKUU TCVG KU QP VJG ;CZKU #U YKVJ OQUV NCPIWCIG VCUMU VJG ITCRJ UJQYU VJCV VJG GTTQTU VTCFGQHH CICKPUV GCEJ QVJGT NQYGTKPI QPG VGPFU VQ TCKUG VJG QVJGT 6JG QHſEKCN GXCNWCVKQP OGCUWTG QH 6&6 KU DCUGF QP C EQUV HWPEVKQP C YGKIJVGF EQO DKPCVKQP QH OKUU CPF HCNUG CNCTO TCVGU %QUV OKUU OKUU VCTIGV HC HC QHHVCTIGV YJGTG VCTIGV KU VJG RTKQT RTQDCDKNKV[ VJCV C UVQT[ YKNN DG QP VQRKE Ü CTG WUGT URGEKſGF XCNWGU VJCV TGƀGEV VJG EQUV CUUQEKCVGF YKVJ GCEJ GTTQT CPF OKUU CPF HC CTG VJG CEVWCN U[UVGO GTTQT TCVGU 9KVJKP 6&6 GXCNWCVKQPU OKUU HC CPF VCTIGV QHHVCTIGV FGTKXGF HTQO VTCKPKPI FCVC +P HCEV C PQTOCNK\GF XGTUKQP QH VJG EQUV HWPEVKQP KU WUGF # U[UVGO VJCV CNYC[U CPUYGTU őPQŒ YQWNF JCXG PQ HCNUG CNCTOU VJQWIJ KV YQWNF JCXG C OKUU TCVG 6JCV U[UVGO YQWNF IGV C UEQTG QH 5KOKNCTN[ C U[UVGO VJCV CNYC[U CPUYGTU ő[GUŒ YQWNF IGV C UEQTG QH 6Q GPUWTG VJCV U[UVGOU VJCV WPFGTRGTHQTO UWEJ UKORNG CRRTQCEJGU CTG XKUKDNG VJG EQUV XCNWG KU FKXKFGF D[ VJG OKPKOWO QH VJG őCNYC[U UC[ [GUŒ QT őCNYC[U UC[ PQŒ CRRTQCEJGU KP VJKU ECUG D[ # PQTOCNK\GF FGVGEVKQP EQUV QH OGCPU VJCV VJG U[UVGO RGTHQTOU GZCEVN[ CU YGNN CU C U[UVGO VJCV FQGU PQ YQTM Ý JVVRYYYU[UVTCPUQHVEQO
E\&5&3UHVV//&
0.02
Miss Rate
90
0.10.2 0.5 1
2
5
10
20
40
60
80
80
80
60
60
40
40
20
20
10
10
5
5
2
2
0.02
0.10.2 0.5 1
2
5 10 20 False Alarm Rate
40
60
80
1 90
FIGURE 12.1 A sample detection error tradeoff (DET) curve for the TDT tracking task with one training story ( ).
0QVG VJCV GXGT[ RQKPV QH VJG &'6 EWTXG EQTTGURQPFU VQ C OKUU CPF HCNUG CNCTO TCVG UQ VJGTG KU C EQUV CV GXGT[ RQKPV QH VJG EWTXG 9KVJKP 6&6 UKVGU CTG GZRGEVGF VQ ſPF C OKPKOWO EQUV QP VJG EWTXG DWV FKHHGTGPEGU DGVYGGP VTCKPKPI CPF VGUV FCVC IGPGTCNN[ OGCP VJG[ OKUU KV UNKIJVN[ # EQOOQP GXCNWCVKQP YKVJKP VJG 6&6 EQOOWPKV[ KU VJG minimum EQUV VJCV EQWNF JCXG DGGP CVVCKPGF WUKPI VJCV &'6 EWTXG 6JCV KU CP GXCNWCVKQP VJCV UKFGUVGRU VJG ſPCN UGNGEVKQP QH VJTGUJQNF VQ IGV C UGPUG QH RQVGPVKCN HQT C VGEJPQNQI[ KH VJTGUJQNF UGNGEVKQP ECP DG TGUQNXGF #NN GXCNWCVKQPU KP 6&6 JCXG DGGP ECTTKGF QWV D[ VJG 0CVKQPCN +PUVKVWVG QH 5VCPFCTFU CPF 6GEJPQNQI[ 2CTVKEKRCVKPI UKVGU YGTG RTQXKFGF YKVJ VJG EQTRWU CPF KPHQTOCVKQP VJCV URGEKſGF VJG UVCTVKPI EQPFKVKQP HQT GCEJ VCUM (QT GZCORNG VTCEMKPI TGSWKTGF VJG UGV QH VTCKPKPI UVQTKGU HQT GCEJ VQRKE CPF VJG QVJGT VCUMU TGSWKTGF C NKUV QH YJKEJ UVQTKGU VQ EQPUKFGT KP VJG UVTGCO 'CEJ UKVG IGPGTCVGF KVU FGEKUKQPU QP VJG UVQTKGU KP VJG GXCNWCVKQP UGV CPF UWDOKVVGF VJGO VQ 0+56 +P VWTP 0+56 FKF VJG GXCNWCVKQP CPF IGPGTCVGF EQORCTCVKXG TGUWNVU QH CNN VJG U[UVGOU =?
E\&5&3UHVV//&
12.2 Basic Topic Models 7PFGTN[KPI CNN CRRTQCEJGU VQ CNN QH VJG 6&6 VCUMU KU VJG PQVKQP QH őVQRKEŒ 6Q CF FTGUU VJG VCUMU KV KU PGEGUUCT[ VJCV C UKVG UQOGJQY OQFGN VQRKEU CPF RQUUKDN[ VJG GXGPVU YKVJKP VJGO 6JG OQFGN EQWNF DG XGT[ UKORNG C NKUV QH UKIPKſECPV YQTFU QT GZVTGOGN[ EQORNKECVGF KP VJG URKTKV QH C MPQYNGFIG DCUG QH RCTVKEKRCPVU CPF VJGKT CEVKQPU /QUV YQTM YKVJKP 6&6 VQ FCVG JCU TGRTGUGPVGF VQRKEU GKVJGT CU C XGEVQT QH YGKIJVGF YQTFU QT CU C RTQDCDKNKV[ FKUVTKDWVKQP QH YQTFU 6JG CRRTQCEJGU CTG UKOKNCT KP VJGKT KORNGOGPVCVKQP CPF GHHGEVKXGPGUU DWV SWKVG FKHHGTGPVN[ OQVKXCVGF
12.2.1 Vector Space 6JG XGEVQT URCEG JCU C NQPI JKUVQT[ YKVJKP KPHQTOCVKQP TGVTKGXCN TGUGCTEJ = ? CPF KU RTQDCDN[ VJG OQUV RQRWNCT YC[ QH KORNGOGPVKPI CP +4 U[UVGO 0QV UWTRTKUKPIN[ IKXGP VJG UKOKNCTKV[ KP VJG RTQDNGOU VJG XGEVQT URCEG JCU DGGP WUGF D[ UGXGTCN UKVGU KP 6&6 = ? 6JG DCUKE KFGC QH VJG XGEVQT URCEG CRRTQCEJ KU VQ TGRTGUGPV KVGOU UVQTKGU QT VQRKEU CU XGEVQTU KP C JKIJ FKOGPUKQPCN URCEG 6JG FKOGPUKQPU EQTTGURQPF VQ VJG HGCVWTGU VJCV CTG WUGF VQ TGRTGUGPV VJG KVGOU CPF CTG QTVJQIQPCN +VGOU VJCV CTG UKOKNCT GPQWIJō IGPGTCNN[ CU OGCUWTGF GKVJGT D[ VJG EQUKPG QH VJG CPING DGVYGGP VJGO QT D[ VJGKT UGRCTCVKQP KP 'WENKFGCP URCEGōCTG CUUWOGF VQ DG QP VJG UCOG VQRKE )GPGTCNN[ VJG YQTFU VJCV QEEWT KP UVQTKGU CTG VJG HGCVWTGU QH VJG XGEVQT URCEG KV KU QDXKQWU VJCV VJG YQTFU CTG PQV KPFGRGPFGPV QH GCEJ QVJGT DWV VJG OQFGN JCU DGGP TGRGCVGFN[ UJQYP VQ YQTM GORKTKECNN[ PQPGVJGNGUU 9QTFU CTG IKXGP YGKIJVU CU FKUEWUUGF NCVGT KP VJKU EJCRVGT 6JG OQUV EQOOQP EQORCTKUQP HWPEVKQP KU VJG EQUKPG QH VJG CPING DGVYGGP VJG VYQ XGEVQTU
¾ ¾
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ſPF VJG TKIJV UGV QH HGCVWTGU YGKIJVU CPF EQORCTKUQP OGVJQF 6JG VJGQT[ FQGU PQV KP CPF QH KVUGNH RTQXKFG OWEJ JGNR KP VJQUG GHHQTVU *QYGXGT VJG OQFGN TGOCKPU RQRWNCT RTGEKUGN[ DGECWUG KV KU UQ UKORNG KV ECP DG GCUKN[ WPFGTUVQQF CPF KORNGOGPVGF
E\&5&3UHVV//&
12.2.2 Language Models 5VCVKUVKECN NCPIWCIG OQFGNKPI CRRTQCEJGU ECOG VQ 6&6 XKC VJG URGGEJ TGEQIPKVKQP EQOOWPKV[ =? CPF VJG +4 EQOOWPKV[ = ? YJKEJ CNUQ IQV VJG KFGC HTQO URGGEJ TGEQIPKVKQP +P VJKU CRRTQCEJ C VQRKE KU TGRTGUGPVGF CU C RTQDCDKNKV[ FKUVTKDWVKQP QH YQTFU *KIJGT RTQDCDKNKV[ YQTFU CTG OWEJ OQTG NKMGN[ VQ CRRGCT KP QPVQRKE UVQTKGU VJCP CTG NQYGT RTQDCDKNKV[ YQTFU 6JG FKHſEWNV CURGEV QH NCPIWCIG OQFGNKPI KU EQOKPI WR YKVJ IQQF YC[U HQT GUVKOCV KPI VJG RTQDCDKNKVKGU 1PG QT OQTG UVQTKGU VJCV CTG MPQYP VQ DG QP VJG UCOG VQRKE CTG VJG UVCTVKPI RQKPV HQT DWKNFKPI C VQRKE OQFGN 6JG KPKVKCN RTQDCDKNKV[ GUVKOCVGU EQOG HTQO VJG OCZKOWO NKMGNKJQQF GUVKOCVG DCUGF QP VJCV FQEWOGPV VH ON Û VH
YJGTG VHÛ TGRTGUGPVU VJG EQWPV QH VKOGU VJCV VJG YQTF QEEWTU KP C UVQT[ CPF VH KU VJG VQVCN PWODGT QH YQTFU KP VJG UVQT[ 6JKU GUVKOCVG KU PQV UWHſEKGPV DGECWUG KV YKNN IKXG \GTQ RTQDCDKNKVKGU VQ CP[ YQTF PQV KP VJG UVQT[ (QT VJCV TGCUQP VJG OCZK OWO NKMGNKJQQF GUVKOCVG KU WUWCNN[ UOQQVJGF YKVJ GUVKOCVGU HTQO C NCTIGT EQTRWU QH PGYU UVQTKGU YKVJ UQOG OKZKPI RCTCOGVGT VJCV FGVGTOKPGU YJGVJGT VJG UVQT[ QT VJG DCEMITQWPF EQTRWU EQPVTKDWVGU OQTG QH VJG GUVKOCVG 6JGTG CTG IGPGTCNN[ VYQ YC[U VQ WUG VJGUG VQRKE OQFGNU 6JG ſTUV KU VQ UGG JQY NKMGN[ KV KU VJCV C RCTVKEWNCT UVQT[ EQWNF DG IGPGTCVGF D[ VJG OQFGN UVQT[ # UVCPFCTF KPFGRGPFGPEG CUUWORVKQP KU OCFG CPF VJG RTQDCDKNKV[ KU GUVKOCVGF CU
UVQT[
Û¾UVQT[
5VQTKGU VJCV JCXG JKIJGT RTQDCDKNKV[ CTG OQTG NKMGN[ VQ DG RCTV QH VJG UCOG VQRKE VJCV KU OQFGNGF # UGEQPF YC[ VQ WUG VQRKE OQFGNU KU VQ EQORCTG VJGO FKTGEVN[ 6JCV KU WUWCNN[ CEEQORNKUJGF YKVJ C U[OOGVTKE XGTUKQP QH VJG -WNNDCEM.GKDNGT FKXGTIGPEG =? UWEJ CU ½ ¾ ¾ ½ 1VJGT YC[U QH EQORCTKPI VJG OQFGNU CTG CNUQ WUGF =?
12.3 Implementing the Models +P VJKU UGEVKQP YG FKUEWUU UGXGTCN QH VJG VGEJPKSWGU VJCV JCXG DGGP WUGF KP 6&6 VQ KORTQXG VJG OQFGN /QUV QH VJGUG VGEJPKSWGU CTG CRRNKECDNG KP UQOG VQ GKVJGT QH VJG OCLQT V[RGU QH OQFGN VJQWIJ UQOG OCMG UGPUG KP QPN[ QPG QT UQOG JCXG DGGP VTKGF KP QPN[ QPG 9G FKUEWUU VJG WUG QH PCOGF GPVKVKGU VJG WUG QH SWGT[ GZRCPUKQP KFGCU UVQT[ ENWUVGTKPI CPF VJG KPENWUKQP QH C VKOG FGEC[ HCEVQT
E\&5&3UHVV//&
12.3.1 Named Entities 0GYU KU WUWCNN[ CDQWV RGQRNG UQ KV UGGOU TGCUQPCDNG VJCV VJGKT PCOGU EQWNF DG VTGCVGF URGEKCNN[ KP C YC[ VJCV YQWNF KORTQXG VJG CEEWTCE[ QH 6&6 U[UVGOU 0COGF GPVKV[ GZVTCEVKQP U[UVGOU JCXG CEJKGXGF JKIJ NGXGNU QH CEEWTCE[ DQVJ HQT IQQF SWCNKV[ PGYU VGZV CPF HQT VJG QWVRWV QH C URGGEJ TGEQIPK\GT = ? 6JCV OGCPU VJCV KV KU RQUUKDNG VQ GZVTCEV YKVJ TGCUQPCDNG EQPſ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resident Bush CPF George Bush YKNN OCVEJ CP KORQTVCPV KUUWG YJKNG PCOG EQ TGHGTGPEG =? TGOCKPU C FKHſEWNV RTQDNGO )KXGP VJG KORQTVCPEG QH RGQRNG RNCEGU CPF FCVGU VQ PGYU TGRQTVKPI KV KU FKUCRRQKPV KPI VJCV PCOGF GPVKVKGU JCXG PQV [GV HQWPF C RNCEG QH RTQOKPGPEG KP VQRKE OQFGNU +V OC[ DG VJCV VJG OQFGNU CTG PQV [GV UQRJKUVKECVGF GPQWIJ VQ KORTQXG WRQP UKORNG YQTFDCUGF OQFGNU (QT GZCORNG GTTQTU KP PCOG GZVTCEVKQP CPF EQORCTKUQP OC[ DG UYCORKPI VJG XCNWG VJCV WUKPI PCOGU CFFU #U VJG OQFGNU DGEQOG OQTG UQRJKUVK ECVGF CPF CEEWTCVG VJGTG OC[ DG OQTG XCNWG KP WUKPI PCOGU
12.3.2 Document Expansion 9JGP VQRKE OQFGNU CTG ETGCVGF HTQO C UKPING UVQT[ VJG[ UWHHGT HTQO GZVTGOGN[ NKO KVGF XQECDWNCT[ 6JGTG CTG EQWPVNGUU YQTFU VJCV EQWNF DG WUGF KP VJG VQRKE CPF UQOG QH VJGO VJCV JCXG C JKIJ NKMGNKJQQF QH CRRGCTKPI 5Q DQVJ OQFGNKPI CRRTQCEJGU PGGF C YC[ QH GZRCPFKPI VJG UGV QH YQTFU VJCV CTG KPENWFGF KP VJG VQRKE OQFGN 8GEVQT URCEG U[UVGOU IGPGTCNN[ WUG VGEJPKSWGU DCUGF QP SWGT[ GZRCPUKQP = ? VJCV JCXG DGGP JKIJN[ UWEEGUUHWN KP +4 GXCNWCVKQPU (QT GZCORNG KP VJG UGIOGPVCVKQP VCUM C RQUUKDNG UGIOGPVCVKQP DQWPFCT[ EQWNF DG EJGEMGF D[ EQORCTKPI VJG OQFGNU IGPGTCVGF D[ VGZV QP GKVJGT UKFG 6Q KORTQXG VJG EJCPEG QH XQECDWNCT[ QXGTNCR VJG VGZV EQWNF DG WUGF CU C SWGT[ VQ TGVTKGXG C HGY FQ\GP TGNCVGF UVQTKGU CPF VJGP VJG OQUV HTGSWGPVN[ QEEWTTKPI YQTFU HTQO VJQUG UVQTKGU EQWNF DG WUGF HQT VJG EQORCTKUQP =? 5KOKNCT CRRTQCEJGU EQWNF DG WUGF VQ CFF YQTFU VQ UVQTKGU HQT VJG QVJGT VCUMU =? #FFKPI YQTFU VQ C NCPIWCIG OQFGN KU KP UQOG YC[U UKORNGT CPF KP QVJGT YC[U OQTG EQORNGZ +V KU UKORNGT DGECWUG UOQQVJKPI YKVJ VJG DCEMITQWPF OQFGN CU FGUETKDGF CDQXG ETGCVGU PQP\GTQ RTQDCDKNKVKGU HQT GXGT[ YQTF KP VJG EQTRWU CPF VJGTGHQTG
E\&5&3UHVV//&
DTKPIU VJQUG YQTFU KPVQ VJG VQRKE VQ UQOG FGITGG *QYGXGT KV FQGU PQV KPETGCUG VJG RTQDCDKNKV[ QH YQTFU VJCV CTG TGNCVGF VQ VJG VQRKE 1PG UVCVKUVKECN NCPIWCIG OQFGNKPI CRRTQCEJ VQ ECRVWTKPI VJG TGNCVGF YQTFU KU VQ WUG TGNGXCPEG OQFGNU =? 6JCV VGEJPKSWG CUUWOGU VJCV UVTQPIN[ UKOKNCT PGYU UVQTKGU CTKUG QWV QH VQRKEU VJCV CTG GKVJGT VJG UCOG CU QT UVTQPIN[ TGNCVGF VQ VJG VQRKE VJCV KU DGKPI OQFGNGF 'CEJ TGVTKGXGF UVQT[ IGPGTCVGU C OQFGN CPF CNN QH VJG OQFGNU CTG EQODKPGF VQ ETGCVG C VQRKE OQFGN
È Û
È Å È ÛÅ
¾
6JG ſPCN TGUWNV KU UKOKNCT KP KORNGOGPVCVKQP VQ SWGT[ GZRCPUKQP DWV KU LWUVKſGF RTQD CDKNKUVKECNN[ UQ VJCV KORQTVCPV RTQRGTVKGU QH VJG NCPIWCIG OQFGNU CTG RTGUGTXGF 4GN GXCPEG OQFGNU TGUWNV KP UWDUVCPVKCN KORTQXGOGPVU KP VJG NKPM FGVGEVKQP VCUM =?
12.3.3 Clustering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őMPQYUŒ VJCV VJG UVQTKGU CTG QP VJG UCOG VQRKE KV ECP ſPF HGCVWTGU VJCV CTG EQOOQP KP VJQUG UVQTKGU VJCV CTG PQV EQOOQP GNUGYJGTG #NVJQWIJ VJG UCOG RTQEGUU ECP DG FQPG UVCTVKPI HTQO C UKPING UVQT[ KV KU OWEJ OQTG TGNKCDNG YKVJ UGXGTCN UVQTKGU 6JKU CRRTQCEJ KU EQOOQP KP VJG XGEVQT URCEG OQFGN YJGTG VJG VQRKE OKIJV DG TGRTGUGPVGF D[ VJG CXGTCIG QH VJG QPVQRKE UVQT[ XGEVQTU OKPWU VJG CXGTCIG QHHVQRKE UVQT[ XGEVQT %NWUVGTKPI KU PQV C ENGCTN[ FGſPGF QRGTCVKQP YKVJKP C NCPIWCIG OQFGNKPI EQPVGZV VJQWIJ KH VYQ OQFGNU CTG UKOKNCT GPQWIJ KV OKIJV OCMG UGPUG VQ OGTIG VJGO 6Q FCVG YG MPQY QH PQ UKVGU VJCV JCXG CVVGORVGF VQ OKOKE XGEVQT ENWUVGTKPI KP VJG NCPIWCIG OQFGN EQPVGZV +PUVGCF VJG NCTIGT UGV QH QPVQRKE UVQTKGU KU WUGF VQ RTQXKFG DGVVGT GUVKOCVGU QH YQTF RTQDCDKNKVKGU YKVJKP VJG VQRKE KP OWEJ VJG UCOG YC[ VJCV TGNGXCPEG OQFGNU FQ CU FGUETKDGF CDQXG KP 5GEVKQP #NVJQWIJ ENWUVGTKPI OCMGU UGPUG KP VTCEMKPI KV ECP CNUQ DG FQPG HQT CNN QH VJG QVJGT VCUMU &GVGEVKQP HQT GZCORNG KU C ENWUVGTKPI VCUM CPF TGSWKTGU VJCV VJG U[UVGO IGP GTCVG ITQWRKPIU # U[UVGO ECP CFCRV KVU VQRKE OQFGNU D[ KPEQTRQTCVKPI PGYN[ CTTKXGF UVQTKGU KPVQ VJG ENWUVGT YJGP VJG[ UGGO UWHſEKGPVN[ ENQUG = ? 6JKU PQVKQP QH CFCRVKPI VJG VQRKE OQFGN KU UKOKNCT VQ VJG CFCRVKXG ſNVGTKPI KFGCU VJCV JCXG TGEGPVN[ DGGP CFQRVGF D[ VJG +4 ſNVGTKPI EQOOWPKV[ =? 0QVG VJCV UQOG UKVGU JCXG HQWPF DGVVGT TGUWNVU D[ MGGRKPI VJG UVQTKGU YKVJKP C ENWUVGT GPVKTGN[ FKUVKPEVōKP C UGPUG VJG[ CTG PQV FQKPI VJG ENWUVGTKPI =? +PUVGCF VJG VQRKE KU TGRTGUGPVGF D[ C UGV QH VQRKEU QPG HQT GCEJ UVQT[ VJCV KU DGNKGXGF VQ DG RCTV QH VJG VQRKE
E\&5&3UHVV//&
1VJGT UKVGU JCXG WUGF C ETQUU DGVYGGP VJG VYQ KFGCU ENWUVGTKPI UVQTKGU YKVJKP VJG VQRKE YJGP VJG[ CTG UWHſEKGPVN[ UKOKNCT DWV MGGRKPI VJGO CRCTV YJGP VJG[ CTG PQV =? 6JKU ETGCVGU C PQVKQP QH őOKETQENWUVGTUŒ VJCV JCU VJG RQVGPVKCN VQ RTQXKFG ƀGZKDKNKV[ YJGP C VQRKE KU OWNVKHCEGVGF +V KU UKOKNCT VQ YQTM FQPG KP VJG +4 EQOOWPKV[ QP MGGRKPI VTCEM QH UJKHVKPI WUGT KPVGTGUVU =?
12.3.4 Time Decay 5GXGTCN UKVGU JCXG QDUGTXGF VJCV VJG NKMGNKJQQF VJCV VYQ UVQTKGU FKUEWUU VJG UCOG VQRKE FKOKPKUJGU CU VJG UVQTKGU CTG HWTVJGT UGRCTCVGF KP VKOG +V KU RQUUKDNG VQ NGXGTCIG VJKU QDUGTXCVKQP D[ ETGCVKPI C RTKQT RTQDCDKNKV[ VJCV VYQ UVQTKGU CTG TGNGXCPV CPF VJGP OQFKH[KPI VJCV DCUGF QP EQPVGPV 1T KP C XGEVQT URCEG OQFGN VJG EQUKPG UKOKNCTKV[ HWPEVKQP ECP DG EJCPIGF UQ VJCV KV KPENWFGU C VKOG FGEC[ =? #HVGT C UWTIG QH KPVGTGUV KP VJG WUG QH C VKOG FGEC[ YKVJKP VJG 6&6 RKNQV UVWF[ CPF GXCNWCVKQP KV JCU WUGF D[ UWDUVCPVKCNN[ HGYGT U[UVGOU KP TGEGPV 6&6 GXCNW CVKQPU 9G UWURGEV VJCV VJG SWCNKV[ QH YQTFDCUGF OCVEJKPI JCU KORTQXGF VQ VJG GZVGPV VJCV VJG VKOG FGEC[ PQ NQPIGT JGNRU CU OWEJ CU KV FKF +V OC[ CNUQ DG VJG ECUG VJCV VJG GXCNWCVKQP VQRKEU CTG PQV GPQWIJ VQ GCEJ QVJGT VJCV VJG VKOG VJG[ CTG TGRQTVGF ECP DG CP KORQTVCPV FKUVKPIWKUJKPI EJCTCEVGTKUVKE +OCIKPG JQY OWEJ UKORNGT KV KU VQ UGRCTCVG UKOKNCT VGTTQTKUV GXGPVU GI DQODKPIU IKXGP VJG VKOG VJCV VJG CVVCEM YCU TGRQTVGF
12.4 Comparing Models 1PEG VQRKE OQFGNU CTG DWKNV TGICTFNGUU QH VJG OQFGN VJG[ PGGF VQ DG WUGF 9KVJKP 6&6 VJCV OGCPU EQORCTKPI C UVQT[ VQ C OQFGN VQ UGG KH VJG UVQT[ KU RCTV QH VJG VQRKE QT RQUUKDN[ EQORCTKPI VYQ OQFGNU VQ FGVGTOKPG VJG EJCPEG VJCV VJG[ TGRTGUGPV VJG UCOG VQRKE +P VJKU UGEVKQP YG FKUEWUU C HGY QH VJG EQORCTKUQP HWPEVKQPU VJCV JCXG DGGP WUGF 6JG ſTUV KU URGEKſE VQ VJG XGEVQT URCEG OQFGN PGCTGUV PGKIJDQT FGEKUKQPU 6JG UGEQPF VJG WUG QH FGEKUKQP VTGGU KU KPFGRGPFGPV QH VJG OQFGNU FKUEWUUGF JGTG UQ HCT 9G CNUQ FKUEWUU FKTGEV OQFGN EQORCTKUQPU YKVJKP VJG NCPIWCIG OQFGNKPI HTCOGYQTM
12.4.1 Nearest Neighbors +P VJG XGEVQT URCEG OQFGN C VQRKE OKIJV DG TGRTGUGPVGF CU C UKPING XGEVQT 5Q YJGP C 6&6 U[UVGO KU TWPPKPI KV YQWNF JCXG C NCTIG UGV QH XGEVQTU TGRTGUGPVKPI CNN QH VJG VQRKEU UGGP VQ FCVG # PGYN[ CTTKXGF PGYU UVQT[ ECP CNUQ DG TGRTGUGPVGF D[ C XGEVQT CPF FTQRRGF KPVQ VJG UCOG URCEG 6Q FGVGTOKPG YJGVJGT QT PQV VJCV UVQT[ KU QP CP[ QH VJG GZKUVKPI VQRKEU YG EQPUKFGT VJG FKUVCPEG WUWCNN[ OGCUWTGF D[ VJG EQUKPG QH VJG CPING DGVYGGP XGEVQTU DGVYGGP VJG UVQT[ŏU XGEVQT CPF VJG ENQUGUV VQRKE XGEVQT +H KV KU UWHſEKGPVN[ UOCNN VJG UVQT[ KU CUUWOGF VQ DG RCTV QH VJG VQRKE +H KV HCNNU QWVUKFG C
E\&5&3UHVV//&
URGEKſGF FKUVCPEG VJG UVQT[ KU NKMGN[ VQ DG VJG UGGF QH C PGY VQRKE CPF C PGY XGEVQT ECP DG HQTOGF 6JG CRRTQCEJ NKUVGF CDQXG KU GUUGPVKCNN[ C PGCTGUV PGKIJDQT CRRTQCEJ YJGTG KU QPG # UVQT[ KU CUUKIPGF VJG VQRKE QH KVU UKPING PGCTGUV PGKIJDQT .CTIGT XCNWGU QH OCMG UGPUG YJGP VJG VQRKE KU TGRTGUGPVGF D[ OWNVKRNG XGEVQTU GKVJGT DGECWUG VJG VQRKE KU OWNVKHCEGVGF =? QT DGECWUG VJG VQRKEŏU UVQT[ XGEVQTU CTG PGXGT EQPUQNKFCVGF KPVQ C UKPING XGEVQT =? +P VJQUG ECUGU C U[UVGO OKIJV NQQM CV UGXGTCN PGKIJDQTU VQ GUVKOCVG VJG VQRKE QH C PGY UVQT[ +H KV NQQMGF CV VJTGG XGEVQTU HQT GZCORNG KV YQWNF UGNGEV VJG VQRKE VJCV KU OQUV EQOOQP YKVJKP VJCV UOCNN UGV 9KVJKP OQUV QH VJG 6&6 VCUMU GXGP YJGP VQRKEU CTG TGRTGUGPVGF D[ OWNVKRNG XGEVQTU TGOCKPU QPG WUKPI C ENWUVGTKPI OQFGN VJCV KU UKOKNCT KP URKTKV VQ UKPING NKPM ENWUVGTKPI =? 6JG TGCUQP HQT VJKU KU VJCV VQRKEU VGPF VQ ITQY QXGT VKOG CPF KPENWFG C YKFGT TCPIG QH FKUEWUUKQP +H VJG UVQTKGU KP C VQRKE CTG OGTIGF VQIGVJGT VJG EQTG QH VJG VQRKE KU ENGCT DWV KVU GFIGU IGV NQUV 5VQTKGU QP VJG HTKPIG QH VJG VQRKE YQWNF PQV DG EQPUKFGTGF RCTV QH VJG VQRKE CPF YQWNF KPEQTTGEVN[ ETGCVG PGY VQRKEU $[ MGGRKPI VJG VQRKE CU C UGV QH FKUVKPEV UVQT[ XGEVQTU VJG TCPIG QH KUUWGU CPF GXGPVU FKUEWUUGF YKVJKP VJG VQRKE KU PQV DNWTTGF 6JG FQYPUKFG QH VJKU CRRTQCEJ KU VJCV VJGTG CTG CNUQ HTKPIG UVQTKGU VJCV UJQWNF PQV VTWN[ DG EQPUKFGTGF RCTV QH VJG VQRKE (QT GZCORNG VJG[ OKIJV EQPVCKP C DTKGH OGPVKQP QH VJG VQRKE DWV RTKOCTKN[ FKUEWUU C FKHHGTGPV VQRKE +H PQV VTGCVGF ECTGHWNN[ UWEJ UVQTKGU ECP KPEQTTGEVN[ OGTIG WPTGNCVGF VQRKEU VQIGVJGT 0QVG VJCV ½ CU C PGCTGUV PGKIJDQT UVTCVGI[ OC[ OCMG VJG OQUV UGPUG HQT VJG VTCEMKPI VCUM YJGTG VJGTG CTG VYQ ENCUUGU QP VQRKE CPF QHH VQRKE 5Q EQORCTKPI C UVQT[ XGEVQT VQ C UGV QH CNTGCF[ ENCUUKſGF XGEVQTU CNNQYU C YKFGT TCPIG QH RQUUKDKNKVKGU ;CPI GV CN =? JCXG GZRGTKOGPVGF YKVJ C YKFG TCPIG QH 00 UVTCVGIKGU VQ ſPF QPGU VJCV YQTM DGUV HQT VTCEMKPI #NVJQWIJ VJG[ YGTG CDNG VQ KORTQXG VJG GHHGEVKXGPGUU QH VJGKT U[UVGO WUKPI XCTKCPV 00 OGVJQFU VJG GTTQT VTCFGQHHU QH VJG VGEJPKSWGU YGTG FKHHGTGPV GPQWIJ VJCV KV YCU PQV ENGCT YJKEJ YQWNF DG DGUV 6JG[ HQWPF VJCV EQODKPKPI TGUWNVU HTQO FKHHGTGPV VGEJPKSWGU CFFTGUUGF OWEJ QH VJCV EQPEGTP
12.4.2 Decision Trees #PQVJGT OQFGN EQORCTKUQP CRRTQCEJ KU VJG WUG QH FGEKUKQP VTGGU +P UQOG YC[U C FGEKUKQP VTGG KU TGCNN[ C VJKTF V[RG QH OQFGN +V CPCN[\GU UQOG VTCKPKPI KPUVCPEGU
GI UVQTKGU VJCV CTG MPQYP VQ DG QP QT QHH VQRKE CPF FGXGNQRU C UGV QH TWNGU HQT ENCUUKH[KPI HWVWTG KPUVCPEGU 6JKU CRRTQCEJ ECP DG WUGF YKVJKP VJG VCEMKPI VCUM YJGP VJGTG CTG UWHſEKGPV PWODGT QH QPVQRKE VTCKPKPI KPUVCPEGU +P VJGQT[ VTCEMKPI ECP DG FQPG YKVJ CU HGY CU QPG RQUKVKXG KPUVCPEG QH C UVQT[ DWV VJCV KU XGT[ UECPV VTCKPKPI HQT DWKNFKPI C TGNKCDNG FGEKUKQP VTGG +V JCU DGGP UJQYP VQ JCXG CEEGRVCDNG TGUWNVU EQORCTCDNG VQ 00 OGVJQFU =? DWV KU PQV YKFGN[ CFQRVGF RGTJCRU DGECWUG KV KU PQV CU ƀGZKDNG 6JG DGUV RNCEG HQT FGEKUKQP VTGGU YKVJKP 6&6 OC[ DG VJG UGIOGPVCVKQP VCUM YJGTG VJGTG CTG PWOGTQWU VTCKPKPI KPUVCPEGU KG JCPFUGIOGPVGF UVQTKGU (KPFKPI HGCVWTGU
E\&5&3UHVV//&
VJCV CTG KPFKECVKXG QH C UVQT[ DQWPFCT[ QT VJG CDUGPEG QH C DQWPFCT[ KU RQUUKDNG CPF CEJKGXGU IQQF SWCNKV[ TGUWNVU =?
12.4.3 Model-to-Model #PQVJGT UV[NG QH EQORCTKUQP KU FKTGEV EQORCTKUQP QH UVCVKUVKECN NCPIWCIG OQFGNU VJCV TGRTGUGPV VQRKEU 6JG V[RKECN YC[ VQ EQORCTG VYQ RTQDCDKNKV[ FKUVTKDWVKQPU KU VQ WUG TGNCVKXG GPVTQR[ QT VJG -WNNDCEM.GKDNGT FKXGTIGPEG ½ ¾ 6JG XCNWG TGRTGUGPVU őVJG CXGTCIG PWODGT QH DKVU VJCV CTG YCUVGF D[ GPEQFKPI GXGPVU HTQO C FKUVTKDWVKQP YKVJ C EQFG DCUGF QP C PQVSWKVGTKIJV FKUVTKDWVKQP Œ = R? .CTIGT XCNWGU EQTTGURQPF VQ NGUU UKOKNCT FKUVTKDWVKQPU +H VJG FKUVTKDWVKQPU CTG KFGPVKECN VJG -. FKXGTIGPEG JCU C XCNWG QH \GTQ )GPGTCNN[ KP 6&6 YG CTG KPVGTGUVGF KP C PQVKQP QH UKOKNCTKV[ TCVJGT VJCP FKUVCPEG UQ VJG PWODGT KU PGICVGF 6JGTGHQTG NCTIGT PWODGTU ENQUGT VQ \GTQ KPFKECVG OQTG UKOKNCT FKUVTKDWVKQPU 0QVG VJCV VJG ECNEWNCVKQP QH VJG -. FKXGTIGPEG
Ü
KU PQV U[OOGVTKE *QYGXGT YG IGPGTCNN[ CUUWOG VJCV KH UVQT[ KU QP VJG UCOG KU QP VJG UCOG VQRKE CU UVQT[ 6Q ſPGUUG VJCV VQRKE CU UVQT[ VJGP UVQT[ RTQDNGO YG ECNEWNCVG VJG -. FKXGTIGPEG DQVJ YC[U CPF CFF VJGO VQIGVJGT ½ ¾ ¾ ½ #PF QH EQWTUG PGICVG KV 1PG QH VJG RTQDNGOU YKVJ EQORCTKPI OQFGNU KU VJCV KV KU CNUQ KORQTVCPV VJCV VJG OQFGNU DG OGCPKPIHWN 5WRRQUG VJCV VYQ OQFGNU CTG EQPUVTWEVGF VJCV CTG KPFKUVKP IWKUJCDNG HTQO IGPGTCN PGYUYKTG VGZV 6JQUG OQFGNU OC[ DG PGCTN[ KFGPVKECN D[ VJG -. FKXGTIGPEG DWV KV KU PQV WUGHWN VQ MPQY VJCV VJG[ CTG 1PG CRRTQCEJ VJCV JCU DGGP WUGF VQ KPEQTRQTCVG VJCV PQVKQP RGPCNK\GU VJG EQORCTKUQP KH VJG OQFGNU CTG VQQ OWEJ NKMG DCEMITQWPF PGYU =? 6JCV KU VJG KPKVKCN EQORCTKUQP KU TGRNCEGF D[ ½ ¾ ½ PGYU 6JG UGEQPF XCNWG KU TGHGTTGF VQ CU őSWGT[ ENCTKV[Œ =? UKPEG VJG NCTIGT KV KU VJG OQTG VJG OQFGN ½ FKXGTIGU HTQO DCEMITQWPF PGYU UQ VJG NGUU IGPGTKE KV KU
12.5 Miscellaneous Issues +P VJKU UGEVKQP YG VCNM CDQWV UGXGTCN CFFKVKQPCN KUUWGU VJCV ECP CHHGEV JQY VQRKE OQF GNU CTG EQPUVTWEVGF 6&6 CNNQYU C OQFGUV COQWPV QH őNQQM CJGCFŒ KPVQ VJG HWVWTG Þ #TIWCDN[ XGEVQT EQORCTKUQPU QH VQRKE OQFGNU KU CNUQ C FKTGEV EQORCTKUQP UQ VJKU KU PQV UVTKEVN[ URGEKſE
VQ NCPIWCIG OQFGNU *QYGXGT KV KU EQPXGPKGPV VQ FKUVKPIWKUJ DGVYGGP VJG VYQ HQT ENCTKV[
Ü 1H EQWTUG KV EQWNF UVKNN DG C VGTTKDNG OQFGN QH VJG VQRKE *QYGXGT CV NGCUV YG MPQY KV KU OQTG URGEKſE
VJCP IGPGTCN PGYU VJCV EQWNF DG CDQWV CP[VJKPI
E\&5&3UHVV//&
KV TGSWKTGU UWRRQTVKPI OWNVKRNG NCPIWCIGU 'PINKUJ %JKPGUG CPF OQTG TGEGPVN[ #TC DKE CPF KV GZRGEVU VJCV OWNVKRNG OQFCNKVKGU YKNN CRRGCT XK\ PGYUYKTG CPF URGGEJ TGEQIPK\GT QWVRWV 'CEJ QH VJQUG ECP KORCEV VJG ETGCVKQP QH C VQRKE OQFGN
12.5.1 Deferral #NN QH VJG 6&6 VCUMU CTG GPXKUKQPGF CU őQPNKPGŒ VCUMU VJCV QRGTCVG QP C EQPVKPWQWUN[ CTTKXKPI UVTGCO QH PGYU +P VJG NKOKV VJCV OGCPU VJCV C FGEKUKQP CDQWV C UVQT[ KU GZRGEVGF DGHQTG VJG PGZV UVQT[ KU RTGUGPVGF +P HCEV 6&6 RTQXKFGU C OQFGTCVG COQWPV QH NQQM CJGCF HQT VJG VCUMU (KTUV UVQTKGU CTG CNYC[U RTGUGPVGF VQ VJG U[UVGO ITQWRGF KPVQ őſNGUŒ VJCV EQTTGURQPF VQ CDQWV C JCNH JQWT QH PGYU PGYUYKTG UVQTKGU CTG ITQWRGF VQIGVJGT VQ CRRTQZKOCVG VJG UCOG COQWPV QH PGYU # U[UVGO ECP FQ CP[ RTQEGUUKPI KV NKMGU QP VJCV GPVKTG ſNG DGHQTG RTGUGPVKPI KVU TGUWNVU HQT VJG ſTUV UVQT[ KP VJG ſNG 6JCV OGCPU VJCV VJGTG KU CNYC[U CP GHHGEVKXG NQQM CJGCF QH WR VQ VJKTV[ OKPWVGU QT CP GSWKXCNGPV PWODGT QH UVQTKGU 5GEQPF VJG HQTOCN 6&6 GXCNWCVKQP =? KPEQTRQTCVGU C PQVKQP QH FGHGTTCN VJCV CN NQYU C U[UVGO VQ GZRNQTG VJG CFXCPVCIG QH FGHGTTKPI FGEKUKQPU WPVKN UGXGTCN ſNGU JCXG RCUUGF 6[RKECN XCNWGU CTG PQ FGHGTTCN KG LWUV VJG YKVJKPſNG FGHGTTCN VGP QT CFFKVKQPCN ſNGU 6JG CFXCPVCIG QH FGHGTTKPI C FGEKUKQP CRRGCTU HQT UVQTKGU VJCV CTG JGCXKN[ TGRQTVGF 6JG GZVTC UVQTKGU YQWNF V[RKECNN[ DG WUGF D[ ENWUVGTKPI VJGO VQIGVJGT CPF VJGP WUKPI VJG CIINQOGTCVGF UWRGTUVQT[ VQ ſPF CP CRRTQRTKCVG VQRKE 6JKU CRRTQCEJ JCU VJG CFXCPVCIG VJCV KH C PGY VQRKE CRRGCTU VJCV KU XGT[ UKOKNCT VQ CP GZKUVKPI QPG VJG GZVTC UVQTKGU OKIJV CWIOGPV VJG FKUVKPEVKQP DGVYGGP VJG PGY CPF QNF VQRKEU CPF FGETGCUG VJG EJCPEG QH C HCNUG CNCTO 1DXKQWUN[ VJG FGHGTTCN KU QPN[ WUGHWN HQT VQRKEU YKVJ OWNVKRNG UVQTKGU KP VJG RGTKQF #NVJQWIJ UGXGTCN UKVGU JCXG YQTMGF YKVJ FGHGTTCN RGTKQFU YG MPQY QH PQ GZJCWUVKXG UVWFKGU VQ FGVGTOKPG VJGKT CFXCPVCIG 6JG 6&6 RKNQV UVWF[ KPENWFGF CP KPſPKVG FGHGT TCN RGTKQF YJQUG RWTRQUG YCU VQ GZRNQTG VJG RQUUKDKNKVKGU QH TGVTQURGEVKXG ENWUVGTKPI QH PGYU =?
12.5.2 Multi-modal Issues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Ō TCPIG =? 4GUGCTEJ KP KPHQTOCVKQP TGVTKGXCN JCU UJQYP NKVVNG KORCEV QP GHHGEVKXGPGUU HTQO TGEQIPKVKQP GTTQTU CU JKIJ CU
E\&5&3UHVV//&
=? DWV KV JCU PQV DGGP VQVCNN[ ENGCT VJG GZVGPV VQ YJKEJ VJCV YQWNF ECTT[ QXGT KPVQ 6&6 VGEJPQNQI[ 1PG CTGC YJGTG VJG VYQ OQFGU VGZV CPF CWFKQ JCXG ENGCT FKHHGTGPEGU CTG KP UEQTG PQTOCNK\CVKQP +H RCKTU QH UVQTKGU CTG EQORCTGF KP VJG NKPM FGVGEVKQP VCUM VJCV CTG HTQO VJG UCOG OQFG KG PGYUYKTG UVQT[ CICKPUV PGYUYKTG UVQT[ YG IGV QPG FKUVTK DWVKQP QH UEQTGU +H VJG RCKTU CTG KPUVGCF FTCYP HTQO CWFKQ UQWTEGU VJG FKUVTKDWVKQP KU FKHHGTGPV #PF KH VJG UVQTKGU EQOG HTQO FKHHGTGPV OQFGU C third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
12.5.3 Multi-lingual Issues 5Q HCT VJG FKUEWUUKQP QH VQRKE OQFGNKPI JCU KORNKEKVN[ CUUWOGF VJCV CNN PGYU UVQTKGU CTG KP 'PINKUJ 6JG 6&6 TGUGCTEJ RTQITCO JCU UVTQPI KPVGTGUV KP GXCNWCVKPI VJG VCUMU CETQUU OWNVKRNG NCPIWCIGU (QT 6&6 VJTQWIJ UKVGU YGTG TGSWKTGF VQ JCP FNG 'PINKUJ CPF %JKPGUG PGYU UVQTKGU KPVGTOKZGF VJQWIJ YKVJKP GCEJ ſNG VJG UVQTKGU YGTG KP VJG UCOG NCPIWCIG (QT 6&6 UKVGU YKNN DG KPEQTRQTCVKPI #TCDKE CU C VJKTF NCPIWCIG 6JG CRRTQCEJ VJCV OQUV UKVGU JCXG WUGF KU VQ EQPXGTV VJG %JKPGUG UVQTKGU KPVQ 'PINKUJ 5KPEG C 5;564#0 VTCPUNCVKQP QH GXGT[ %JKPGUG UVQT[ KU UWRRNKGF YKVJ VJG EQTRWU VJCV KU VJG UKORNGUV UQNWVKQP CPF QPG CFQRVGF D[ OCP[ UKVGU *QYGXGT QVJGT ITQWRU CTG CEVKXGN[ TGUGCTEJKPI ETQUUNCPIWCIG KPHQTOCVKQP TGVTKGXCN CPF TGNCVGF RTQDNGOU CPF VJG[ FKF VJGKT QYP KPHQTOCVKQP TGVTKGXCN SWCNKV[ VTCPUNCVKQPU QH VJG UVQTKGU = ? 5KOKNCT RTQDNGOU CTKUG KP RTQEGUUKPI ETQUUNCPIWCIG UVQTKGU CU KP ETQUUOQFG UVQ TKGU 6JG FKUVTKDWVKQPU QH UEQTGU CTG FKHHGTGPV FGRGPFKPI QP VJG OQFGUōRGTJCRU GXGP OQTG UQ KP VJG ETQUUNCPIWCIG ECUG 6JG 5;564#0 UVQTKGU HQT GZCORNG WUG 'PINKUJ KP C YC[ VJCV KU RGEWNKCT VQ VJG U[UVGOŏU QWVRWV CPF SWKVG FKUVKPEV HTQO JWOCP IGPGTCVGF VGZV 6JCV OGCPU VJCV 5;564#0 UVQTKGU CTG OWEJ OQTG NKMGN[ VQ DG UKOKNCT VQ GCEJ QVJGT VJCP VQ UVQTKGU VJCV YGTG QTKIKPCNN[ YTKVVGP KP 'PINKUJ 6JG WRUJQV QH VJKU KU VJCV PQTOCNK\KPI VJG FKUVTKDWVKQPU ECP JCXG CP GXGP NCTIGT KORCEV VJCP KP VJG ETQUUOQFG ECUG =?
E\&5&3UHVV//&
PKEJQNUGPVGPEG MCE\[PUMKDWTTGNN CNIGTKCOCUUCETG TKVVGTKTCS PGVCP[CJWRCNGUVKPKCP NGDCPQPJG\DQNNCJ RTQVGUVCPVKTGNCPF OEXGKIJPCX[ ICPFJKPGJTW LCICPIW[CPC MC\KPUM[VGF QJJGCFNKPG KEGUVQTO LCEMGVDKMMGODGTI MWTFMWTFKUJ LQURKPLQDNGUU FGPVWTGVTCP GVCDCUSWG ENQPGEGNN PWENGCTOKUUKNG EWDCRQRG KVCNKEUJTCFGT FKUCTOG\KNM DCOKCPVCNKDCP DQCV[CEJV
IJCPKOCVUJCDCJ OWOEWJKVOGP TGUVKVWVKQPLGVNKPGT OQVJGTJQQFPGGF[ HNGGVYQQFKPFWEV DWORGTNGZW CTPGEVENCTM \JWHCEVQT[ ITKUYQNFDKUJQR HKUJMGCVG EQYNQTFPCPEG FKUEQPVGPVGCPIGT XKGVEQPIVGV ECUVTQEWDC EJGTQMGGLGGR VJCKNCPFHWPF FQYKPFWUVTKCN NWVJGTNGUQVJQ JWVWVWVUK IKNNGVVGTQDDGT[ FJMRKPEKTNKM JCTT[JCWUGPCTIQPCWV KPUWTCPEGOQKUVWTG RQNTQWIG DGGVNGXQNMUYCIGP
FIGURE 12.2 Screen snapshot of the Lighthouse system that was created to portray TDT topic clusters and their relationships.
12.6 Using TDT Interactively 6&6 KU XKGYGF CU CP GPCDNKPI VGEJPQNQI[ HQT C TCPIG QH VCUMU VJCV YCPV VQ KORQUG GXGPVDCUGF QTICPK\CVKQP QP PGYU UVQTKGU 7NVKOCVGN[ VJG VGEJPQNQI[ OWUV DG KPEQT RQTCVGF KPVQ U[UVGOU UQ VJCV RGQRNG ECP WUG KV CPF UGG YJGVJGT KV KU JGNRHWN +V JCU VWTPGF QWV VJCV VJKU V[RG QH QTICPK\CVKQP KU WPHCOKNKCT VQ RGQRNG UQ KU FKHſEWNV VQ RTGUGPV +P VJKU UGEVKQP YG FKUEWUU VYQ YC[U VJCV 6&6 VGEJPQNQI[ ECP DG GZRQUGF
12.6.1 Demonstrations .KIJVJQWUG KU C RTQVQV[RG U[UVGO VJCV XKUWCNN[ RQTVTC[U KPVGTFQEWOGPV UKOKNCTKVKGU VQ JGNR VJG WUGT ſPF TGNGXCPV OCVGTKCN OQTG SWKEMN[ =? +V TGRTGUGPVU FQEWOGPVU CU URJGTGU KP URCEG CPF RNCEGU VJG URJGTGU UWEJ VJCV JKIJN[ UKOKNCT FQEWOGPVU CTG PGCTD[ KP URCEG CPF VJG NGUU CNKMG VJG[ CTG VJG HCTVJGT CRCTV VJG[ UJQWNF DG 6JG RQTVTC[CN QH FQEWOGPV ENWUVGTKPI VJCV KU RQUUKDNG KP .KIJVJQWUG JCU DGGP UJQYP VQ CNNQY C UVCVKUVKECNN[ UKIPKſECPV KORTQXGOGPV KP GHHGEVKXGPGUU QXGT +4ŏU ENCUUKECN TCPMGF NKUV 9G ETGCVGF C XGTUKQP QH .KIJVJQWUG YJGTG VJG RQTVTC[GF QDLGEVU YGTG VQRKEU CPF
E\&5&3UHVV//&
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ſGF CV C NCVGT XKGYKPI # RKG UNKEG KU FGRKEVGF QP VJG VQR QH GCEJ URJGTG VJCV KPFKECVGU VJG RTQRQTVKQP QH VJG VQRKE OCFG WR QH UVQTKGU VJCV CRRGCTGF UKPEG VJG NCUV VKOG VJG XKGY YCU WRFCVGF 6JKU HGCVWTG CNNQYU JKIJN[ XQNCVKNG QT GPVKTGN[ PGY VQRKEU VQ DG TGEQIPK\GF CV C INCPEG # VQRKE ECP DG UGNGEVGF UQ VJCV VJG UVQTKGU YKVJKP KV ECP DG TGCF 6JG .KIJVJQWUGDCUGF 6&6 U[UVGO YCU HWP CPF ƀCUJ[ CPF RTQXKFGF OQUV QH VJG HWPEVKQPCNKV[ C WUGT YQWNF RTQDCDN[ NKMG *QYGXGT KV YCU CYMYCTF VQ WUG RGQRNG FQ PQV WPFGTUVCPF URJGTGU ƀQCVKPI KP URCEG CPF VJGKT KPVGTTGNCVKQPUJKRU OCFG PQ UGPUG 9G CTG EWTTGPVN[ YQTMKPI QP C UWDUVCPVKCNN[ EJCPIGF RTGUGPVCVKQP QH VJG UCOG KFGCU DWV DCUGF WRQP VJG őſNG HQNFGTŒ OGVCRJQT EQOOQP QP EQORWVGT FGUMVQRU
12.6.2 Timelines #PQVJGT YC[ QH RTGUGPVKPI 6&6 KPHQTOCVKQP KU WUKPI C VKOGNKPG VQ UJQY PQV QPN[ YJCV VJG VQRKEU CTG DWV JQY VJG[ QEEWT KP VKOG 6JG KPVGTGUVKPI RCTV QH VJG RTQDNGO KU ſPFKPI YC[U VQ EQPUVTWEV VJG VKOGNKPG CWVQOCVKECNN[ = ? 6JG YQTM KP VJKU CTGC KU PQV RCTV QH 6&6 FKTGEVN[ DWV NGXGTCIGU UKOKNCT KFGCU 9G UVCTV D[ GZVTCEVKPI CNN PCOGU CPF PQWP RJTCUGU HTQO C EQNNGEVKQP QH PGYU CUUWO KPI VJCV PCOGU CPF VJKPIU CTG VJG EGPVTCN EQORQPGPVU QH OQUV QH VJG PGYU 9G VJGP UECP VJTQWIJ VJQUG HGCVWTGU EQPUKFGTKPI C FC[ QH PGYU CV C VKOG (QT GCEJ FC[ YG WUG VJG ¾ OGCUWTG VQ FGVGTOKPG YJGVJGT QT PQV VJCV HGCVWTG KU QEEWTTKPI QP VJCV FC[ KP CP WPWUWCN YC[ōV[RKECNN[ VJCV OGCPU VJCV KV QEEWTU OWEJ OQTG QHVGP QP VJCV FC[ VJCP QP QVJGT FC[U (QT GZCORNG VJG YQTF Oklahoma QEEWTTGF KP VJG PGYU OWEJ OQTG QHVGP UJQTVN[ CHVGT VJG 1MNCJQOC %KV[ DQODKPI KP VJCP KV FKF FWTKPI VJG OCP[ OQPVJU QH PGYU DGHQTG VJCV 6JG ¾ OGCUWTG RKEMU VJCV WR TGCFKN[ )KXGP C UGV QH HGCVWTGU VJCV CTG KPVGTGUVKPI YKVJKP C RCTVKEWNCT VKOG RGTKQF YG ITQWR VJGO VQIGVJGT DCUGF QP YJGVJGT VJG[ EQQEEWT HTGSWGPVN[ 5Q Oklahoma OKIJV ITQWR YKVJ McVeigh DWV PQV YKVJ QVJGT HGCVWTGU GI Simpson HTQO VJG 1, 5KORUQP OWT FGT VTKCN VJCV QEEWTTGF CV CDQWV VJG UCOG VKOG 0QY VJCV C ITQWR QH HGCVWTGU JCU DGGP HQWPF VJG EQNNGEVKQP QH UVQTKGU VJCV VCNM CDQWV VJCV VQRKE KU GCU[ VQ KUQNCVG 6JG VKOGNKPG ECP VJGP DG EQPUVTWEVGF CU KP (KIWTG 'CEJ VQRKE KU FGRKEVGF CU C TGEVCPING 6JG NGHVTKIJV URCP QH VJG VQRKE TGƀGEVU VJG FWTCVKQP QH TGRQTVKPI QP VJG VQRKE +VU CTGC KU FGVGTOKPGF D[ VJG VQVCN PWODGT QH UVQTKGU QP C VQRKE UQ C ƀCV TGEVCPING JCU C OQFGTCVG PWODGT QH UVQTKGU QXGT C NQPI RGTKQF YJGTG C XGT[ VCNN TGEVCPING KPFKECVGU OWEJ TGRQTVKPI KP C XGT[ UJQTV RGTKQF 6JG XGTVKECN RQUKVKQPKPI QH VJG TGEVCPING KU FGVGTOKPGF D[ KVU őKORQTVCPEGŒ QT UWTRTKUG
E\&5&3UHVV//&
OQPKECNGYKPUM[
CNNGICVKQP
LQPGUDQTQ YGUVUKFGOKFFNGUEJQQN YKNNG[
VGTT[PKEJQNU
FGCVJRGPCNV[ DTQPEQU UWRGTDQYN
UGZWCNCFXCPEG PQTVJGTPKTGNCPF
FCKONGTDGP\ GCUVGT
KPFWUVTKCNOGTIGT QTGIQP UEJQQNUJQQVKPI NQWKUGYQQFYCTF
FKUOKUUVJKUOGPW JKFGPCOGU JKFGRNCEGU JKFGQTIU QTGIQP URTKPIHKGNF MKRMKPMGN MKPMGN CR VJWTUVQPJKIJUEJQQN QTG ,CPWCT[
(GDTWCT[
/CTEJ
#RTKN
/C[
CWRCKT
MKRMKPMGN FKUOKUUVJKUOGPW /CMGVKVNG )GV*KUVQITCO UJQYEQPVGZV ;QWTCFJGTG ,WPG
FIGURE 12.3 Overview of January-June 1998. The topic labeled monica lewinsky allegation is the highest ranked topic by the ¾ measure. The pop-up on oregon school shooting shows significant named entities for that event. The other pop-up displays a sub-menu for obtaining more information on the name kip kinkel. XCNWG DCUGF QP VJG ¾ XCNWG QH KVU HGCVWTGU 6QRKEU VJCV CTG XGT[ WPWUWCN CTG RTGUGPVGF CV VJG VQR QH VJG ITCRJ 6JG ITCRJ KP (KIWTG KPENWFGU VJG VGP OQUV őKORQTVCPVŒ VQRKEU KP VJG VKOG RGTKQF # PKEG HGCVWTG QH VJCV KORQTVCPEG XCNWG KU VJCV KV ECP DG WUGF VQ MGGR VQRKEU QP VJG FKURNC[ CV CP[ NGXGN QH ITCPWNCTKV[ VJG OQUV KORQTVCPV QT CP[ QVJGT PWODGT VQRKEU KP C RCTVKEWNCT RGTKQF QH VKOG ECP CNYC[U DG UGNGEVGF
12.7 Modeling Events #NN QH VJG CDQXG FKUEWUUKQP KU CDQWV OQFGNKPI VQRKEU YKVJKP 6&6 YKVJ C UOCVVGTKPI QH FKUEWUUKQP QP RTGUGPVKPI VJG TGUWNVU QH VJG YQTM +PVGTGUVKPIN[ CNOQUV PQVJKPI KP VJG TGUGCTEJ NKVGTCVWTG HQT 6&6 CVVGORVU VQ OQFGN VQRKEU CU OQTG VJCP C őDCI QH YQTFUŒ VJCV CTG YGKIJVGF CRRTQRTKCVGN[ 6QRKEU KP VJG PGYU CTG TGNCVGF VQ GXGPVU 'XGPVU CTG CDQWV RGQRNG 6JG[ VCMG RNCEG CV C RCTVKEWNCT RNCEG CPF KP C IKXGP VKOG 9KVJ VJG GZEGRVKQP QH VJG WUG QH PCOGF GPVKVKGU =? CPF VJG GZRNKEKV KPENWUKQP QH VKOG =? VJGTG KU CNOQUV PQ TGEQIPKVKQP VJCV VJQUG EQORQPGPVU QH VJG PGYU VQRKEU OKIJV DG WUGHWN 5Q HCT VJGTG KU XKTVWCNN[ PQ FKUVKPEVKQP DGVYGGP VJG VGEJPQNQI[ WUGF HQT 6&6 CPF VJCV WUGF HQT FQEWOGPV TGVTKGXCN 6JCV NCEM QH FKUVKPEVKQP OKIJV PQV DG CP KUUWG GZEGRV VJCV 6&6 RGTHQTOCPEG KU PQV
E\&5&3UHVV//&
CFGSWCVG HQT CP[VJKPI QVJGT VJCP NKOKVGF WUGU KVU GTTQT TCVGU CTG UVKNN VQQ JKIJ (WT VJGT VJG GHHGEVKXGPGUU QH KPHQTOCVKQP TGVTKGXCN U[UVGOU CRRGCTU VQ JCXG RNCVGCWGF KP VJG NCUV UGXGTCN [GCTU UWIIGUVKPI VJCV ICKPU KP GHHGEVKXGPGUU CTG PQV NKMGN[ VQ DG JCF HTQO VJCV FKTGEVKQP ß +V OC[ DG RQUUKDNG VQ KPEQTRQTCVG VJG OKUUKPI CURGEVU QH GXGPVDCUGF VQRKEU D[ OQFGN KPI VJGO GZRNKEKVN[ 'KEJOCPP CPF 5TKPKXCUCP =? DWKNV FKHHGTGPV XGEVQTU HQT RGQRNG QTICPK\CVKQPU CPF UQ QP EQORCTGF VJGO UGRCTCVGN[ CPF VJGP OGTIGF VJGO WUKPI C NKPGCT EQODKPCVKQP QH VJG RKGEGYKUG UKOKNCTKVKGU 6JKU CRRTQCEJ YCU PQV JKIJN[ UWE EGUUHWN DWV KV ENGCTN[ ECRVWTGF UQOG QH VJG PQVKQPU QH GXGPVU 9G URGEWNCVG VJCV VJG RTQDNGO KU UVCTVKPI VJKU CRRTQCEJ HTQO VJG XGEVQT URCEG RGTURGEVKXG YJKEJ IKXGU PQ VJGQTGVKECN LWUVKſECVKQP QT OQVKXCVKQP HQT CP[ QH VJG UVGRU CNQPI VJG YC[ 9J[ UJQWNF PCOGU CPF QTICPK\CVKQPU DG UGRCTCVG! 9JCV CTG VJG[ VT[KPI VQ ECRVWTG! 9J[ C NKPGCT EQODKPCVKQP! 2GTJCRU KV YQWNF DG DGVVGT VQ NQQM HQT VJG YJQ YJGTG YJCV CPF YJGP QH PGYU UVQTKGU VQ GZRNKEKVN[ OQFGN VJG UWDLGEV QH C VQRKE QT GXGPV QT VT[ VQ KFGPVKH[ VJG NQECVKQP QH VJG JCRRGPKPI -PQYKPI VJCV VJG U[UVGO KU VT[KPI VQ ECRVWTG VJQUG URGEKſE RKGEGU QH VJG GXGPV OCMGU KV RQUUKDNG VQ GXCNWCVG VJQUG KVGOU FKTGEVN[ TCVJGT VJCP GXCNWCVKPI QPN[ CV VJG NGXGN QH VQRKE OCVEJ 9G JCXG DGIWP UQOG YQTM KP VJKU CTGC JQRKPI VQ DWKNF C TKEJ VQRKE OQFGN VJCV ECR VWTGU VJG XCTKQWU CURGEVU QH GXGPVU 9G CTG OQVKXCVGF D[ VJG UVCVKUVKECN NCPIWCIG OQFGNKPI CRRTQCEJGU CPF JQRG VJCV YG ECP GZVGPF VJCV YGNN GPQWIJ VQ KORTQXG GH HGEVKXGPGUU 4GUWNVU CU QH VJKU YTKVKPI ECP QPN[ DG FGUETKDGF CU RTQOKUKPI 4GICTFNGUU QH YJGVJGT QWT YQTM KU UWEEGUUHWN QT YJGVJGT 'KEJOCPP CPF 5TKPKXCUCPŏU CRRTQCEJ KU VJG QPG VJCV YQTMU VJG KORQTVCPV RQKPV KU VJCV UOCNN UVGRU CTG DGKPI OCFG VQYCTF OQFGNKPI GXGPVU GZRNKEKVN[ 6&6 VCUMU CTG PQV NKMGN[ VQ KORTQXG UWDUVCPVKCNN[ KP CEEWTCE[ CU NQPI CU DTQCFGT OQTG IGPGTCN +4 VGEJPQNQI[ KU VJG QPN[ CRRTQCEJ WUGF
12.8 Conclusion 9G JCXG VCNMGF CDQWV VJG VQRKE FGVGEVKQP CPF VTCEMKPI 6&6 TGUGCTEJ RTQITCO CPF UMGVEJGF UGXGTCN QH VJG CRRTQCEJGU VJCV JCXG DGGP WUGF VQ CFFTGUU VJG VCUMU 9G VCNMGF CDQWV XGEVQT URCEG OQFGNU CPF UVCVKUVKECN NCPIWCIG OQFGNU VJG VYQ FQOKPCPV RCTCFKIOU KP VJG 6&6 TGUGCTEJ NKVGTCVWTG 9G FKUEWUUGF UGXGTCN QH VJG VGEJPKSWGU VJCV U[UVGOU JCXG WUGF VQ DWKNF QT GPJCPEG VJQUG OQFGNU CPF NKUVGF OGTKVU QH OCP[ QH VJGO 9G EQPENWFGF D[ VCNMKPI CDQWV VJG HCKNWTG QH VQRKE OQFGNU VQ KPEQTRQTCVG VJG PQVKQP QH őGXGPVŒ GZRNKEKVN[ TGN[KPI QP VGEJPQNQI[ VJCV KU LWUV CU WUGHWN HQT VJG UWDLGEV ß +PHQTOCVKQP TGVTKGXCN research JCU PQV RNCVGCWGF 6JG YQTM KU DGKPI GZVGPFGF KPVQ C YKFG TCPIG QH PGY CTGCU CPF PGY VGEJPQNQIKGU CPF KFGCU CRRGCT EQPUVCPVN[ #HVGT KORTQXKPI C [GCT HQT UGXGTCN [GCTU TCPMGF FQEWOGPV TGVTKGXCN JCU PQV KORTQXGF UWDUVCPVKCNN[ UKPEG VJG NCVG U =?
E\&5&3UHVV//&
DCUGF VQRKEU HCOKNKCT VQ +4 FQEWOGPV TGVTKGXCN 9G DGNKGXG VJCV 6&6 TGUGCTEJGTU JCXG ENGCTN[ FGOQPUVTCVGF VJG GZVGPV VQ YJKEJ +4 VGEJPQNQI[ ECP DG WUGF VQ UQNXG 6&6 RTQDNGOU *QYGXGT YG CNUQ DGNKGXG VJCV 6&6 VGEJPQNQI[ OWUV CPF ECP DG UWDUVCPVKCNN[ KORTQXGF CPF VJCV VJG QPN[ CXGPWG VQ VJCV IQCN KU VJTQWIJ KPEQTRQTCVKPI KPHQTOCVKQP CDQWV GXGPVU KPVQ VJG OQFGNU FKTGEVN[
References =? , #NNCP +PETGOGPVCN TGNGXCPEG HGGFDCEM +P Proceedings of Conference on Information Retrieval Research (SIGIR) RCIGU Ō <WTKEJ =? , #NNCP , %CTDQPGNN ) &QFFKPIVQP , ;COTQP CPF ; ;CPI 6QRKE FGVGEVKQP CPF VTCEMKPI RKNQV UVWF[ (KPCN TGRQTV +P Proceedings of the DARPA Broadcast News Transcription and Understanding Workshop RCIGU Ō =? , #NNCP * ,KP / 4CLOCP % 9C[PG & )KNFGC 8 .C XTGPMQ 4 *QDGTOCP CPF & %CRWVQ 6QRKEDCUGF PQXGNV[ FGVGEVKQP UWOOGT YQTMUJQR CV %.52 ſPCN TGRQTV #XCKNCDNG CV JVVRYYYENURLJWGFWYUVFV =? , #NNCP 4 2CRMC CPF 8 .CXTGPMQ 1PNKPG PGY GXGPV FGVGEVKQP CPF VTCEM KPI +P Proceedings of Conference on Information Retrieval Research (SIGIR) RCIGU Ō =? ,COGU #NNCP &GVGEVKQP CU OWNVKVQRKE VTCEMKPI Information Retrieval 8QN PWODGT RCIGU -NWYGT #ECFGOKE 2TGUU =? ,COGU #NNCP 8KEVQT .CXTGPMQ &CXKF (TG[ CPF 8KMCU -JCPFGNYCN 7/CUU CV 6&6 0QVGDQQM RWDNKECVKQP HQT RCTVKEKRCPVU QPN[ 0QXGODGT =? ,COGU #NNCP 8KEVQT .CXTGPMQ CPF *WDGTV ,KP %QORCTKPI GHHGEVKXGPGUU KP 6&6 CPF +4 6GEJPKECN 4GRQTV +4 7PKXGTUKV[ QH /CUUCEJWUGVVU %GPVGT HQT +PVGNNKIGPV +PHQTOCVKQP 4GVTKGXCN =? ,COGU #NNCP 8KEVQT .CXTGPMQ CPF 4WUUGNN 5YCP 'ZRNQTCVKQPU YKVJKP VQRKE VTCEMKPI CPF FGVGEVKQP +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Event-based Information Organization RCIGU Ō -NWYGT #EC FGOKE 2WDNKUJGTU $QUVQP =? #OKV $CIIC CPF $TGEM $CNFYKP 'PVKV[DCUGF ETQUUFQEWOGPV EQTGHGTGPEKPI WUKPI VJG XGEVQT URCEG OQFGN +P In Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics and the 17th International Conference on Computational Linguistics (COLING-ACL’98) RCIGU Ō =? & $GGHGTOCP # $GTIGT CPF , .CHHGTV[ 6GZV UGIOGPVCVKQP WUKPI GZRQPGPVKCN OQFGNU +P Proceedings for Empirical Methods in NLP
E\&5&3UHVV//&
=? % $WEMNG[ ) 5CNVQP , #NNCP CPF # 5KPIJCN #WVQOCVKE SWGT[ GZRCP UKQP WUKPI 5/#46 64'% +P Proceedings of the Text Retrieval Conference (TREC-3) RCIGU Ō 0+56 =? ,CKOG %CTDQPGNN ;KOKPI ;CPI ,QJP .CHHGTV[ 4CNH & $TQYP 6QO 2KGTEG CPF :KP .KW %/7 TGRQTV QP 6&6 5GIOGPVCVKQP FGVGEVKQP CPF VTCEMKPI +P Proceedings of the DARPA Broadcast News Workshop RCIGU Ō /QT ICP -CWHHOCP 2WDNKUJGTU =? *UKP*UK %JGP CPF .WP9GK -W #P 0.2 +4 CRRTQCEJ VQ VQRKE FGVGEVKQP +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Event-based Information Organization RCIGU Ō -NWYGT #ECFGOKE 2WDNKUJGTU $QUVQP =? %JTKUVQRJGT %KGTK 5VGRJCPKG 5VTCUUGN &CXKF )TCHH 0KK /CTVG[ -CTC 4GPPGTV CPF /CTM .KDGTOCP %QTRQTC HQT VQRKE FGVGEVKQP CPF VTCEMKPI +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Event-based Information Organization RCIGU Ō -NWYGT #ECFGOKE 2WDNKUJGTU $QUVQP =? 9 $TWEG %TQHV 5VGRJGP %TQPGP6QYPUGPF CPF 8KEVQT .CXTGPMQ 4GNGXCPEG HGGFDCEM CPF RGTUQPCNK\CVKQP # NCPIWCIG OQFGNKPI RGTURGEVKXG +P Proceedings of the DELOS-NSF Workshop on Personalization and Recommender Systems in Digital Libraries RCIGU Ō =? 5 &JCTCPKRTCICFC / (TCP\ ,5 /E%CTNG[ 6 9CTF CPF 9, <JW 5GI OGPVCVKQP CPF FGVGEVKQP CV +$/ +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Event-based Information Organization RCIGU Ō -NWYGT #ECFGOKE 2WDNKUJGTU $QUVQP =? &CXKF 'KEJOCPP CPF 2CFOKPK 5TKPKXCUCP # ENWUVGTDCUGF CRRTQCEJ VQ DTQCF ECUV PGYU +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Eventbased Information Organization RCIGU Ō -NWYGT #ECFGOKE 2WDNKUJ GTU $QUVQP =? & 'XCPU CPF 4 .GHHGTVU &GUKIP CPF GXCNWCVKQP QH VJG %.#4+6 64'% U[UVGO +P Proceedings of the Text Retrieval Conference (TREC-2) RCIGU Ō 0+56 =? ,QPCVJCP ) (KUEWU CPF )GQTIG 4 &QFFKPIVQP 6QRKE FGVGEVKQP CPF VTCEMKPI GXCNWCVKQP QXGTXKGY +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Event-based Information Organization RCIGU Ō -NWYGT #ECFGOKE 2WD NKUJGTU $QUVQP =? &CXKF (TG[ 4CJWN )WRVC 8KMCU -JCPFGNYCN 8KEVQT .CXTGPMQ #PVQP .GWUMK CPF ,COGU #NNCP /QPKVQTKPI VJG PGYU C 6&6 FGOQPUVTCVKQP U[UVGO +P Proceedings of the Human Language Technology Conference (HLT) RCIGU Ō =? ,5 )CTQHQNQ %)2 #W\CPPG CPF '/ 8QQTJGGU 6JG 64'% URQMGP FQE WOGPV TGVTKGXCN VTCEM # UWEEGUU UVQT[ +P Proceedings of the Text Retrieval Conference (TREC-8) 0+56 URGEKCN RWDNKECVKQP
E\&5&3UHVV//&
=? 8KEVQT .CXTGPMQ ,COGU #NNCP 'F &G)W\OCP &CPKGN .C(NCOOG 8GGTC 2QN NCTF CPF 5VGRJGP 6JQOCU 4GNGXCPEG OQFGNU HQT VQRKE FGVGEVKQP CPF VTCEMKPI 6GEJPKECN 4GRQTV +4 7PKXGTUKV[ QH /CUUCEJWUGVVU %GPVGT HQT +PVGNNKIGPV +PHQTOCVKQP 4GVTKGXCN =? 8KEVQT .CXTGPMQ CPF 9 $TWEG %TQHV 4GNGXCPEGDCUGF NCPIWCIG OQFGNU +P Proceedings of ACM SIGIR Conference on Research in Information Retrieval RCIGU Ō =? 6KO .GGM 4KEJCTF 5EJYCTV\ CPF 5TKPKXCUC 5KUVC 2TQDCDKNKUVKE CRRTQCEJGU VQ VQRKE FGVGEVKQP CPF VTCEMKPI +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Event-based Information Organization RCIGU Ō -NWYGT #EC FGOKE 2WDNKUJGTU $QUVQP =? #PVQP .GWUMK CPF ,COGU #NNCP .KIJVJQWUG 5JQYKPI VJG YC[ VQ TGNGXCPV KPHQT OCVKQP +P Proceedings of the IEEE Symposium on Information Visualization 2000 (InfoVis 2000) RCIGU Ō =? #PVQP .GWUMK CPF ,COGU #NNCP +ORTQXKPI TGCNKUO QH VQRKE VTCEMKPI GXCNWCVKQP 6GEJPKECN 4GRQTV +4 7PKXGTUKV[ QH /CUUCEJWUGVVU %GPVGT HQT +PVGNNKIGPV +PHQTOCVKQP 4GVTKGXCN =? %& /CPPKPI CPF * 5EJiWV\G Foundations of Statistical Natural Language Processing /+6 2TGUU %CODTKFIG /CUUCEJWUGVVU =? # /CTVKP ) &QFFKPIVQP 6 -COO CPF / 1TFQYUMK 6JG &'6 EWTXG KP CUUGUUOGPV QH FGVGEVKQP VCUM RGTHQTOCPEG +P EuroSpeech RCIGU Ō =? &CXKF /KNNGT 4KEJCTF 5EJYCTV\ 4CNRJ 9GKUEJGFGN CPF 4GDGEEC 5VQPG 0COGF GPVKV[ TGEQIPKVKQP HTQO DTQCFECUV PGYU +P Proceedings of the DARPA Broadcast News Workshop =? .QPI 0IW[GP 5R[TQU /CVUQWMCU ,CUQP &CXGPRQTV ,C[ $KNNC 4KEJ 5EJYCTV\ CPF ,QJP /CMJQWN 6JG $$0 $;$.15 Z46 DTQCFECUV PGYU VTCP UETKRVKQP U[UVGO +P Proceedings of the 2000 Speech Transcription Workshop =? &CXKF & 2CNOGT ,QJP & $WTIGT CPF /CTK 1UVGPFQTH +PHQTOCVKQP GZVTCEVKQP HTQO DTQCFECUV PGYU URGGEJ FCVC +P Proceedings of the DARPA Broadcast News Workshop =? 4QP 2CRMC On-line New Event Detection, Clustering, and Tracking 2J& VJGUKU 7PKXGTUKV[ QH /CUUCEJWUGVVU =? 4QP 2CRMC CPF ,COGU #NNCP 6QRKE FGVGEVKQP CPF VTCEMKPI 'XGPV ENWUVGTKPI CU C DCUKU HQT ſTUV UVQT[ FGVGEVKQP +P 9 $TWEG %TQHV GFKVQT Advances in Information Retrieval: Recent Research from the CIIR EJCRVGT RCIGU Ō -NWYGT #ECFGOKE 2WDNKUJGTU
E\&5&3UHVV//&
=? 4QP 2CRMC ,COGU #NNCP CPF 8KEVQT .CXTGPMQ 7/#55 CRRTQCEJGU VQ FG VGEVKQP CPF VTCEMKPI CV 6&6 +P Proceedings of the DARPA Broadcast News Workshop RCIGU Ō =? , 2QPVG CPF 9$ %TQHV # NCPIWCIG OQFGNKPI CRRTQCEJ VQ KPHQTOCVKQP TG VTKGXCN +P Proceedings of SIGIR RCIGU Ō =? ,C[ 2QPVG A Language Modeling Approach to Information Retrieval 2J& VJGUKU 7PKXGTUKV[ QH /CUUCEJWUGVVU =? ,C[ 2QPVG CPF 9 $TWEG %TQHV 6GZV UGIOGPVCVKQP D[ VQRKE +P Proceedings of the European Conference on Research and Advanced Technology for Digital Libraries (ECDL) RCIGU Ō =? 5VGRJGP 4QDGTVUQP CPF &CXKF # *WNN 6JG 64'% ſNVGTKPI VTCEM ſPCN TGRQTV +P Proceedings of the Text Retrieval Conference (TREC-9) RCIGU Ō =? )GTCTF 5CNVQP CPF /KEJCGN , /E)KNN Introduction to Modern Information Retrieval /E)TCY*KNN 0GY ;QTM %JCRVGT RCIGU =? , /KEJCGN 5EJWNV\ CPF /CTM ; .KDGTOCP 6QYCTFU C őWPKXGTUCN FKEVKQPCT[Œ HQT OWNVKNCPIWCIG +4 CRRNKECVKQPU +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Event-based Information Organization RCIGU Ō -NWYGT #ECFGOKE 2WDNKUJGTU $QUVQP =? #PFTGCU 5VQNEMG 'NK\CDGVJ 5JTKDGTI &KNGM *CMMCPK6WT )QMJCP 6WT
E\&5&3UHVV//&
=? ,KPZK :W CPF 9 $TWEG %TQHV +ORTQXKPI VJG GHHGEVKXGPGUU QH KPHQTOCVKQP TG VTKGXCN YKVJ NQECN EQPVGZV CPCN[UKU ACM Transactions on Information Systems (TOIS) RRŌ =? ;KOKPI ;CPI ,CKOG %CTDQPGNN 4CNH $TQYP ,QJP .CHHGTV[ 6JQOCU 2KGTEG CPF 6JQOCU #WNV /WNVKUVTCVGI[ NGCTPKPI HQT 6&6 +P ,COGU #NNCP GFKVQT Topic Detection and Tracking: Event-based Information Organization RCIGU Ō -NWYGT #ECFGOKE 2WDNKUJGTU $QUVQP
E\&5&3UHVV//&