1、1. . , , , . ( , , ), . , . (character) , 10. , , . 1-best n-best . , 1-best . . 1-best . . . , , . . O * * , *ywyoonO, jhm, gbleepostech.ac.krError detection and correction in speech recognition by using lexico-semantic patternsYongWook YoonO HanMin Jung*, Gary Geunbae Lee*Graduate School of Inform
2、ation *Dept. of Computer Science and Engineering, Technology, POSTECH POSTECH . , . , . . . , . , script . .2 / , 3 . 4 , 5 .2. 2.1 Noisy channel modelRingger et al.1 Noisy Channel Model . noisy channel . word sequence , W word sequence , W |P . , language model , |W error channel model . |P indepen
3、dence ,| |iiwP 2 language model . . , 1:1 1:n n:1 (3.1 ). 1 fertility , k fertility k .2.2 Error Pattern MatchingSatoshi Kaki3 word string error pattern matching . random . error pattern database . database error pattern string error pattern error database correct string . . error string corpus(stri
4、ng-database) . error string key string-database string string error string . . string 3.1 n:m . . 4 error pattern 2.3 Error Pattern 5 string error pattern / 3 pattern matching , error database error pattern bigram . . n-best lattice , pair - . error pattern . pair . 3 . , n-best . . . 3 .3. 3.1 (ins
5、ertion), (deletion), (replacement) 3 .1. (): : 2. (): : 3. ( 1): : ( 2): : ( 3): : 80% (4 ). , . , 2 3 n:1 1:m , 1:1 .3.2 L S P ( ) L S P ( ) 1. 1 . . ( , , ) . . 3.3 (LSP: Lexico-Semantic Pattern) open-domain query type matching 67. Kim8 , , sequence . 9 . . .: LSP : %action|xsp|ef|%|pa|ef|: LSP: %
6、shop|%here|j|subway|%from|%most-near|ef%action %shop % , subway . xsp, ef, j , , . . matching / Finite State Automata(FSA) / .3.4 1 . off-line . 1,011 6 . 42,385 , 2 . 2. / 2 . 6 , , , , , 1 “*” (dont care) . “*” sequence , sequence . “*” . “*” , . . () %action|%action|*|xsp|ef|%name ncp %action|%ac
7、tion|ncp|*|ef|%name xsp %action|%action|ncp|xsp|*|%name ef %action|%action|ncp|xsp|ef|%name * %action|%action|ncp|xsp|ef|* %name %action|%to|%action|*|ef|% xsp %action|%to|%action|xsp|ef|* % %action|%to|*|xsp|ef|% %action %action|*|%action|xsp|ef|% %to|j %action|*|%together|pv|ef|% j uknc|j|*|ep|ep|
8、 pv uknc|j|subway|%near|%from|* pv uknc|j|subway|%near|%from|pv * “ .” (: ) %wash|%and|%shop|jcs|%together|pa * ( ) *|%and|%shop|jcs|%together|pa %wash %wash|*|%shop|jcs|%together|pa %and %wash|%and|*|jcs|%together|pa %shop %wash|%and|%shop|*|%together|pa %jcs %wash|%and|%shop|jcs|*|pa %together %wa
9、sh|%and|%shop|jcs|%together|* pa , , 6 . , . 3.5 n:m . n:m (:) error pattern database , . Satoshi3 error pattern . .- “ “ - “ “ “ “ - “ “ “ “ - “ “ “ “ - “ “ “ “ - “ “ “ - “ “ “ - “ “ “ - “ “ “ “ - “ “ LSP . / . , , , . . , . , , , . / . 3 . 4. ByVoice1 . Windows platform 1-best Windows API . ByVoic
10、e 30 , . Linux application client Windows web application . navigation 1,011. 3 . database . baseline ( ). 1. ( )Level # of input# of incorrectcorrectnessSentence 540 224 58.5 %Word 4243 341 92.0 %1 http:/www.byvoice.co.kr/ 540 107 2 . 2. (word ) Word 15 9 95 119 - 8 4 55 67 56.3 15 9 71 95 79.8 2 l
11、exical level , . . , . 35, 10 . . . 5. “ “ # 0 0-0 ncn (1 0) %wash # 1 1-1 j (0 0) %and # 2 2-2 uknc (1 0) uknc # 3 3-3 j (0 0) jcs # 4 4-4 ma (1 0) %together # 5 5-5 pa (1 0) pa # 6 6-6 ef (0 0) jxc # 7 7-7 unoun (1 0) %name # 8 8-8 unoun (1 0) nbn # 9 9-9 j (0 0) %from # 10 10-10 ncn (1 0) %price
12、# 11 11-11 j (0 0) jcs # 12 12-14 ncp (1 0) %most-cheap # 13 0-0 (1 0) # 14 0-0 (0 0) # 15 15-15 ef (0 0) ef # 16 16-16 nbn (1 0) %location # Speech Recognition Error: %wash|%and|uknc|jcs|%together|pa # Speech Recognition Hypothesis: %wash|%and|*|jcs|%together|pa %shop|ncn # Speech Recognition Error
13、: %and|uknc|jcs|%together|pa|jxc # Speech Recognition Hypothesis: %and|*|jcs|%together|pa|jxc %serv|%shop|%wash|ncn # Speech Recognition Error: uknc|jcs|%together|pa|jxc|%name # Speech Recognition Hypothesis: *|jcs|%together|pa|jxc|%name %shop|%wash|ncn : (): %shop (3), ncn (3), %wash (2), %serv (1)
14、 LSP : %shop ncn # Speech Recognition Correction: uknc %shop at 2th # shop: , , ; %shop entry# - ; by minimal edit distance from 3. . . corpus . . 35 . window 2 , . n-to-m model , ( % ) POS feature feature corpus . .6. 1 E.K.Ringger et al., “A Fertility Model forPost Correction of Continuous SpeechR
15、ecognition”. ICSLP96, pp.897-900, 1996.2 F. Jelinek ,“Self-Organized LanguageModeling for Speech Recognition”, Readings in Speech Recognition, Morgan Kaufmann Publishers, Inc., San Mateo, CA, pp.450-506, 1990.3 Satoshi Kaki et al., 98, “A Method forCorrecting Speech Recognition Using the Statistical
16、 features of Character Co-occurrence”. COLING-ACL98, p.653-657, 1998.4 James F. Allen et al., “A Robust System forNatural Spoken Dialogue”, Proceedings ofthe 34th Annual Meeting of the ACL 96,pp.62-70, 1996.5 , , “ ”, 2000 ,pp.441-443, 2000.6 Sanda Harabagiu, Dan Moldovan, Marius Pasca,Rada Mihalcea
17、, Mihai Surdeanu, Razvan Bunescu, Roxana Girju, Vasile Rus and Paul Morarescu “The Role of Lexico-Semantic Feedback in Open-Domain Textual Question-Answering”, in Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics (ACL-2001), Toulouse France, pp.274-281, 2001.7 Gary Geunbae Lee, Jungyun Seo, Seungwoo Lee,Hanmin Jung, Bong-Hyun Cho, Changki Lee, ByungKwan Kwak, Jeongwon Cha, Dongseok Kim, JooHui An, Harksoo Kim, Kyun
