Microsoft Word - Rocling 2012 Yu.doc

Size: px

Start display at page:

Download "Microsoft Word - Rocling 2012 Yu.doc"

玉聊蓬
5 years ago
Views:

1 Skp N-gram Modelng for Near-Synonym Choce Shh-Tng Chen Department of Informaton Management Yuan Ze Unversty We-Cheng He Department of Informaton Management Yuan Ze Unversty Phlps Kokoh Prasetyo Lvng Analytcs Research Centre Sngapore Management Unversty Lang-Chh Yu Department of Informaton Management Yuan Ze Unversty (Near-Synonym) (Pontwse Mutual Informaton, PMI) N (N-gram) (Skp N-gram) SemEval

2 arm (weapon)(bodypart) :arm weapon weapon arsenal arm weapon arsenal [1,2] (Computer-Asssted Language Learnng, CALL) [3,4] (1) coffee [5] Near-Synonym{strong, powerful} (2) ghastly [6] Near-Synonym{error, mstake} (3) under the bay [7] Near-Synonym{brdge, overpass, tunnel} (1)(2)(1) strong, powerful strong strong coffee powerful(2) error, mstake ghastly mstake mstake(3){brdge, overpass, tunnel} under the bay tunneltunnel [7] 164

3 (Skp N-gram) N (N-gram)(Skp)N N Web 1T 5-gram corpus N N N Web 1T 5 gram corpus N N N () Web 1T 5-gram Google Web 1T 5-gram corpus Google Web 1T 5-gram Google [8] Google [9] 1 Web 1T 5-gram 1 Web 1T 5-gram 24GB Tokens 1,024,908,267,229 Sentences 95,119,665,584 Ungrams 13,588,391 Bgrams 314,843,401 Trgrams 977,069,902 Fourgrams 1,313,818,354 Fvegrams 1,176,470,663 () FITB(fll-n-the-blank) FITB(fll-n-the-blank) (gap) 165

4 [10,11] 1 FITB 1 FITB () Inkpen PMI(Pontwse Mutual Informaton,)[6]PMI Gardner Dras PMI [11] N Inkpen N N N [12] PMI N WSDWord sense dsambguaton WSD [13]Dagan WSD [14] () 1. : XML XML 5 (5-gram) 2 166

5 2 (EIC) : clean : <nstance d="388"> <context>grace has the money to <head>clean</head> up.</context> </nstance> : has the money to clean the money to clean up money to clean up (EIC) : clean.v 388 :: wn 1;proft greatly 1;clear 1;prosper 1;accumulate 1;make a fortune 1; : :{wn, proft greatly, clear, prosper, accumulate, make a fortune} (EIC) :(5 ) has the money to clean : has the money to wn has the money to proft greatly has the money to clear has the money to prosper has the money to accumulate has the money to make a fortune 167

6 () N N N N 0 5 Islam Inkpen 5 [12] N () N(N-gram) N N N Google Web 1T 5-grams N (Ungram)2 (Bgram)3 (Trgram)4 (Fourgram)5 (Fvegram) 1. N : s w4w3w2w 1ww 1w2w3w w 1 5 Pww ( ), 1 Pw ( 1 w3) Pw ( 2 w 2 2) Pw ( 3 w 1) : Pw 3 ( 4 w ) 5 1 Pww n Cwn 1 n Mwn 1 Pww n Cw ( n1) nmw ( n1) Ps () ( ) ( ) (1 ) ( ) ( ) 1 M( w n1) n1 n1 n1 w 4 5-gram Mw ( ) Cw ( ) Cw ( ) (2) (1) C() N Web 1T 5-gram N N N N N N () (Skp N-gram) 5 N Web 1T 5-gram N N Skp4Skp3Skp2 6-8 Skp4Skp3Skp2 : 168

7 6 Skp4 5 : has the money to clean Skp4 * the money to clean 243 has * money to clean 0 has the * to clean 1099 has the money * clean 0 has the money to * 0 7 Skp3 5 : has the money to clean Skp3 * * money to clean 1025 * the * to clean * the money * clean 652 * the money to * 243 has * * to clean 5999 has * money * clean 0 8 Skp2 5 : has the money to clean Skp2 has the * * * 1435 has * money ** 0 has * * to * 6071 has *** clean * the money * * 652 * the * to * * the ** clean * has the money to * has the money to cleanhas the money to afford has the money to back has the money to cover... has the money to has the money to * N N N N=

8 9 N 0 5-gram gram gram gram :1703 N=5N=4 N=3 N=5 N=5 N=4 N=3 Skp4 Skp3 SemEval-2007 SemEval-2007 SemEval [15] McCarthy Navgl [15] () - Sharoff Englsh Internet Corpus(EIC) Sharoff (:SemEval-2007[15]) PoS # Noun 497 Verb 440 Adjectve 468 Adverb 298 All 1703 () - SemEval-2007 [15]

9 11 : If ths Government had been dong ts job they would have total confdence. 11 (:SemEval-2007[15]) duty functon bt responsblty duty task role job.n 433 :: duty 2;functon 1;bt 1;responsblty 1;task 1;role 1; () [15] (Recall)(Mode Recall) (Recall) 5 (Mode Recall) 12 (Recall) R a: T resa T a freq H res (3) T H freq res 12 Recall Mode R bgt m 1 f bg T m m (4) T m bg m 13 Mode Recall 171

10 () 1. N-gram Skp N-gram Skp N-gram 14: 14 N-gram+Skp Recall Mode Recall N-gram N5S4N3N2N N5N4S3N2N N5N4N3S2N N5S4S3N2N N5N4S3S2N N5S4N3S2N N5S4S3S2N N-gram Skp N5S4S3N2N1 N5S4S3N2N1 5-gramSkp4Skp32-gramungram N5S4S3N2N1 N-gram Skp N-gram Skp N-gram Skp N=4 N=3 N5S4S3N2N1 2. (Accuracy) Accuracy bgt m 1 f bg orgnal word All (6) bg orgnalword All System Accuracy N-gram 30.30% N5S4N3N2N % N5N4S3N2N % N5N4N3S2N % N5S4S3N2N % N5N4S3S2N % N5S4N3S2N % N5S4S3S2N % 172

11 15 N5N4S3N2N % SemEval-2007 N5N4S3N2N1 SemEval-2007 N5S4S3N2N1 N-gram Skp N-gram N-gram Skp N-gram N-gram 3. N N : 16 N-gram Skp N-gram : I over and made a U turn whle Chrs got out, ran over and took a pcture. : pull pull stop N-gram 5-gram 5-gram pull over and made a 0 stop over and made a 0 4-gram pull over and made 0 4-gram stop over and made 0 over and made a 0 over and made a 0 Skp Skp4 Skp4 pull * and made a 0 stop * and made a 0 pull over * made a 0 stop over * made a 0 pull over and * a 1172 stop over and * a 171 pull over and made * 0 stop over and made * 0 17 N-gram Skp N-gram : Java so that all of the clone( ) methods catch the CloneNotSupportedExcepton rather than t to the caller. : pass pass hand N-gram 5-gram 5-gram than pass t to the 0 than hand t to the 50 Skp Skp 4 Skp 4 than pass * to the 0 than hand * to the 50 than pass t * the 157 than hand t * the 50 than pass t to * 0 than hand t to *

12 N stop pull N-gram pull stop 5-gram 4-gram 0 3-gram2-gram ungram 4-gram stop pull N-gram stop Skp4 pull stop pull N hand pass N-gram hand 5 50pass 0 N hand Skp4 pass hand pass N SemEval [15] N N 0 N N N SemEval-2007 N N NSC E MY3 [1] D. Moldovan and R. Mhalcea, "Usng WordNet and Lexcal Operators to Improve Internet Searches," IEEE Internet Computng, pp , [2] J. Bhogal, A. Macfarlane, and P. Smth, "A Revew of Ontology based Query Expanson," Informaton Processng & Management, pp , [3] C. Cheng, "Word-Focused Extensve Readng wth Gudance," In Proc. of the 13th Internatonal Symposum on Englsh Teachng, pp , [4] S. Ouyang, H. H. Gao, and S. N. Koh, "Developng a Computer-Facltated Tool for Acqurng Near-Synonyms n Chnese and Englsh," In Proc. of IWCS-09, pp ,

13 [5] D. Pearce, "Synonymy n Collocaton Extracton," In Proc. of the Workshop on WordNet and Other Lexcal Resources at NAACL-01, 2001 [6] D. Inkpen, "A Statstcal Model of Near Synonym Choce," ACM Trans. Speech and Language Processng, pp. 1-17, [7] L. C. Yu, C. H. Wu, R. Y. Chang, C. H. Lu, and E. H. Hovy, "Annotaton and verfcaton of sense pools n OntoNotes," Informaton Processng & Management 46(4), pp , [8] A. Islam, D. Inkpen, "Real-word spellng correcton usng Google Web 1T 3-gram, " In Proc. of EMNLP-09, pp , [9] P. Nulty, F. Costello, "Usng lexcal patters n google Web 1T corpus to deduce semantc relaton between nouns," In Proc. of the NAACL/HLT Workshop on Semantc Evaluatons: Recent Achevements and Future Drectons, pp , [10] P. Edmonds, "Choosng the Word Most Typcal n Context Usng a Lexcal Co-occurrence Net Network," In Proc. of ACL-97, pp , [11] M. Gardner and M. Dras, "Explorng Approaches to Dscrmnatng among Near-Synonyms," In Proc. of the Australasan Technology Workshop, pp , [12] A. Islam, D. Inkpen, "Near-Synonym Choce usng a 5-gram Language Model," In proc. Research n Computng Scence, pp , [13] D. McCarthy, "Lexcal Substtuton as a Task for WSD Evaluaton," In Proc. of the SIGLEX/SENSEVAL Workshop on Word Sense Dsambguaton at ACL-02, pp , [14] I. Dagan, O. Glckman, A. Glozzo, E. Marmorshten, and C. Strapparava, "Drect Word Sense Matchng for Lexcal Substtuton," In Proc. of COLING/ACL-06, pp , [15] D. McCarthy, R. Navgl, "The Englsh lexcal substtuton task," In Proceedngs of the 4th Internatonal Workshop on Semantc Evaluatons (SemEval-2007), pp ,

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