中文模板

Size: px

Start display at page:

Download "中文模板"

漩毛阎
6 years ago
Views:

1 /2005/ Journal of oftware Vol6, No9 +, 2,,, ,00084 emantc Analyss and tructured Language Models LI Mng-Qn +, LI Juan-Z 2 ANG Zuo-Yng, LU Da-Jn Department of Electronc Engneerng, snghua Unversty, Beng 00084, Chna 2 Department of Computer cence and echnology, snghua Unversty, Beng 00084, Chna + Correspondng author Phn , E-mal lmq@thspeetsnghuaeducn, http//wwwtsnghuaeducn Receved ; Accepted L MQ, L JZang ZY, Lu DJ emantc analyss and structured language models Journal of oftware, 2005, DOI 0360/os6523 Abstract An ntegrated semantc analyss system s presented, and the structured language models are proposed based on t he semantc analyss system can automatcally tag semantc class for each word and analyze the semantc dependency structure between words wth the precson of 9085% and 7584% respectvely In order to descrbe sentence structure and long-dstance dependency, two nds of structured language models are examned and analyzed Fnally, these two language models are evaluated on the tas of Chnese speech recognton Experments show that the best semantc structured language model headword trgram model acheves 08% absolute error reducton and 8% relatve error reducton over the trgram model Key words semantc analyss; dependency analyss; language model; speech recognton,, 9085% 7584%,,,, 08%, 8% ; ; ; P8 A upported by the Natonal Hgh-ech Research and Development Plan of Chna under Grant No200AA ,,,,,, ;964,,,,CCF,,, ; 935,,,, ; 928,,,,

2 524 Journal of oftware 2005,69 N [],,,,,, [2] rgger [3] [4] ppng [5],, [6,7] [8] [9] [0],,, 9085% 7584%,,,,,, 08%, 8%,, 2 3 4, [], 343,, [2],,,,, 59, 70,,,,,,,,,, /Experencer,,,,,,,, /content,, N W w, w,, w }, { 2 N RL { R, R2,, R N }, R H, R Rsemantc relaton, R, R H, R, ernel word 2, H

3 525 w /s /Dd5 /Ae3 /Ca /Ka0 /Gb2 /Aa04 /Hc05 /Da4 /Kd0 /Ie0 /H28 Englsh hese years, Doctor Yang pays a lot of attenton to the popularzaton and applcaton of hs nventon a he sentence tagged wth semantc classes a experencer tme content degree restrctve restrctve de restrctve dependency coordnaton [Yang] [Doctor] [these years] [a lot] [hs] [producton] [popularzaton] [pay attenton to] [nventon] [of] [applcaton] target b he semantc dependency tree b Fg emantc dependency structure of a Chnese sentence w /s /Dd5 /Ae3 /Ca /Ka0 /Gb2 /Aa04 /Hc05 /Da4 /Kd0 /Ie0 /H28 Englsh hese years, Doctor Yang pays a lot of attenton to the popularzaton and applcaton of hs nventon - 2 a he sentence tagged wth semantc classes a R Modfer HeadWord H Index Word Index Word emantc relaton R /Yang 2 /Doctor /Restrctve 2 /Doctor 5 /pay attenton to /Experencer 3 /these years 5 /pay attenton to /me 4 /a lot 5 /pay attenton to /Degree 5 /pay attenton to - - /Kernel word 6 /hs 8 /producton /Restrctve 7 /nventon 8 /producton /Restrctve 8 /producton 0 /popularzaton /Content 9 /of 8 /product // De dependency 0 /popularzaton 5 /pay attenton to /arget /applcaton 0 /popularzaton /Coordnaton b emantc dependency relaton lst b Fg2 emantc dependency relaton lst of a Chnese sentence 2 2 W w, w,, w } s, s,, s } RL { R, R2,, { 2 N { 2 N R N }, P, W, N P, W { P s,, W P w, P, }

4 526 Journal of oftware 2005,69,,,, U,,, P s, P s s, s 2 2 P w, P w s, w 3 P, 3 22 P, W,,, N Pc, W { Pc w,, W Pc s,, W P, } P w, P w w, w 5 c 2 P s, P s w, s 6 c, W, W 3,, 3,,,,,, r hw hs,,, {, / }, 70 /, ;,, NULL NULL r,,,,, 3,,,,, P, P τ, P q,, q,, q 4, 7, q τ,, 3, 2, Restrcton,rght Restrcton,rght P q,, q,, q,, q l,, q l, g P q r, r P q hwl, hs g l, hw, g hs g core q,, q,, q λ log P q wl, s,, log, l w g s g + λ P q rl r g l g 8, λ, λ 03 P q r, r P q hw, hs, hw, hs l g l l g g

5 527 [0],, Experencer,rght, me,rght, Degree,rght, arget,left, Restrctve,rght, De dedependency,left, Restrctve,rght, Restrctve,rght, Restrctve,rght, Coordnaton,left [Yang] [Doctor] [these years] [a lot] [hs] [producton] [popularzaton] [pay attenton to] [nventon] [ de ] ['de'] [applcaton] Fg3 One of the bnary semantc parse trees for the semantc dependency tree n Fg 3 2,,,, 2 W W P W P, W 9, [0], 9 P Pr W P,, Pr, 0 P W,, Pr, * *,, arg max P, W, 0 * * P W P,, 0

6 528 Journal of oftware 2005, , [6],, 4, 9, rgram model a radtonal word trgram model a experencer Headword trgram model restrctve 2 3 tme degree restrctve de restrctve dependency b Headword trgram model b Fg4 4,, P 2 2 w W PWP w,, W P,, W, 4 P WP h h,, 2 P WP P WP w, P w, 2 P h, h 2 Pr 2 w W PWP w,, W ρ,, W, ρ,,, N,, W Pr PN,, W 3 4 P, ρ 5, P, Chelba [6] 6 N P, { P s,, W P w, P, }, P, P q,, q,, q τ W,, W, 6,,, τ 6

7 529,,,, 5 [ ] τ, P s,, W P w, P, J P N 7 7, 4, 7, 3 P N 3, semantc dependency net, DN [2] 9394 DN, ~994 2, 2, DN, 7 2, DN 9394, DN [3] [0] ;, 9394, 9394-DN;, 9394-DN, CR CR, w, w, R,, H w, w, R, H 2 RCR RCR 3 CR CR, em class tagger 2 Dep parser [0],Dep, 00%, W W 4 able Results of semantc analyss system CR % RCR % CR % em class tagger 94 Dep parser W-Parser W-Parser ,W W W [3] W,W

8 530 Journal of oftware 2005,69 P s s 2, s 2, W W,, 33,W W,,,, 32 HP,HP, 6,5, 20, 25 [6,7],,, P W λp W + λ P 8, W P I W P W P w P w2 w P w w 2, w N 3, P w W λ P w W + λ P w w, w I PPL,,, K PPL exp log PM w / w,, w K 0,, P w w,,, w w, 3 W- 20 λ, 2, λ 04, rgram 28%, 07% able 2 Perplexty of headword trgram model under dfferent nterpolaton weghts 2 λ PPL , 2 [6] 2, h w, ρ, 5 δ x, y D 2 N Pr P d δ D, d ρ, y x, δ x, y ;, δ x, y 0 HP, 5 4% 2, , 323, 45 MFCC, 4 MFCC, DDBHMMduraton dstrbuton based hdden Marov model [4],, trgram, 00 ; 9 2

9 53, 3 3 trgram -best n n-best Probablty Depth Fg5 Depth dstrbuton of headword trgram model 5 able 3 Chnese character error rates of -best and n-best paths of baselne % 3 -best n-best % -best 5-best 20-best 00-best FDM BDM HM, able 4 Results of all semantc structure language models % 4 % CER relatve error reducton FDM BDM HM W-Parser W-Parser FDMBDM,, W-Parser W-Parser 780% 299% BDM HM, 3,, 67%,,, 2,,,,,,, 08%, 84%,,

10 532 Journal of oftware 2005,69 4, [5], 677%,,,, [6,7],Chelba W, Chelba Upenn reeban,,, 2, W W,,, Chelba W 3 7, 6, Chelba , W , Chelba 008%, W 937% our Our normalzaton method Chelba's Chelba s normalzaton method Probablty Depth Fg6 Depth dstrbuton comparson between headword trgram models wth and wthout our normalzaton method 6 5 [0],,, 9085% 7584%,

11 533,, 07%, 08%, 8%,, 3 ;2,,, ;3, P WP References [] Jelne F elf-organzed language modelng for speech recognton In Wabel A, Lee KF, eds Readngs n peech Recognton an Mateo Morgan Kaufmann Publshers, [2] Brown PF, DellaPetra VJ, Deouza PV, La JC, Mercer RL Class-Based n-gram models of natural language Computatonal Lngustcs, 992, [3] Lau R, Rosenfeld R, Rouos rgger-based language models A maxmum entropy approach In ullvan BJ, ed Proc of the Int l Conf on Acoustcs, peech, and gnal Processng ICAP, Vol II [4] Bellegarda JR A mult-span language modelng framewor for large vocabulary speech recognton IEEE rans on peech Audo Processng, 998, [5] Gao JF, uzu Hen Y Explorng headword dependency and predctve clusterng for language modelng In Hac J, Matsumoto Y, eds Proc of the Emprcal Methods n Natural Language Processng EMNLP [6] Chelba C Explotng syntactc structure for natural language modelng [PhD hess] Johns Hopns Unversty, 2000 [7] Xu P, Chelba C, Jelne F A study on rch syntactc dependences for structured language modelng In Proc of the 40th Annual Meetng of the Assocaton for Computatonal Lngustcs ACL ACL, [8] Roar B Probablstc top-down parsng and language modelng Computatonal Lngustcs, 200, [9] Gao JF, uzu H Unsupervsed learnng of dependency structure for language modelng In Proc of the 4st Annual Meetng of the Assocaton for Computatonal Lngustcs ACL ACL, http//researchmcrosoftcom/~fgao/paper/ dlm-acl03pdf [0] L MQ, L JZang ZY, Lu DJ A statstcal model for parsng semantc dependency relatons n a Chnese sentence Chnese Journal of Computers, 2004, n Chnese wth Englsh abstract [] Me JJ, Zhu YM, Gao YQ, Yn HX ongyc Cln Dctonary of ynonymous Words hangha hangha Cshu Publsher, 983 n Chnese [2] L MQ, L JZ, Dong ZDang ZY, Lu DJ Buldng a large Chnese corpus annotated wth semantc dependency In Ma Q, Xa F, eds Proc of the 2nd IGHAN Worshop on Chnese Language Processng [3] Zhang JP A study of language model and understandng algorthm for large vocabulary spontaneous speech recognton [PHD hess] Beng Department of Electronc Engneerng, snghua Unversty, 999 n Chnese wth Englsh abstract [4] Wang ZY, Xao X Duraton dstrbuton based HMM speech recognton models Chnese Journal of Electroncs, 2004, n Chnese wth Englsh abstract [5] Zhou M A bloc based dependency parser for unrestrcted Chnese text In Proc of the 2nd Chnese Language Processng Worshop http//researchmcrosoftcom/chna/papers/robust_dependency_parser_chnese_extpdf [0],,,2004, [],,,,983 [3] [ ],999 [4], HMM,2004,

基于词语关联度的查询缩略*

基于词语关联度的查询缩略* * 基于词语关联度的查询缩略陈炜鹏 1, 付瑞吉 1, 胡熠 2, 秦兵 1, 刘挺 (1. 哈尔滨工业大学计算机科学与技术学院社会计算与信息检索研究中心, 黑龙江省哈尔滨市,150001; 2. 腾讯公司搜索平台部, 广东省深圳市,518057) 摘要 : 冗