ESP [10] 2 (1) (2),,, Extracting Domain-Specific Expressions from ESP Corpora in View of Syntacti

Size: px

Start display at page:

Download "ESP [10] 2 (1) (2),,, Extracting Domain-Specific Expressions from ESP Corpora in View of Syntacti"

こうしょしろみず
4 years ago
Views:

1 ESP [10] 2 (1) (2),,, Extracting Domain-Specific Expressions from ESP Corpora in View of Syntactic Structures TANAKA, Shosaku KOYAMA, Yukie Ritsumeikan University, College of Letters Toji-inkitamachi 56-1, Kita-ku, Kyoto, JAPAN Nagoya Institute of Techonology, Center for Research & Development in Higher Education Gokiso-cho, Showa-ku, Nagoya, Aichi, JAPAN Abstract This paper proposes a sophisticated method for extracting domain-specific expressions on the basis of a classification model The advanced points of the proposed method are as follows: (1) it extracts domain-specific expressions per grammatical role by applying syntactic analysis to sentences and (2) it restricts expressions to extract by grammatical patterns The results show that the proposed method can extract compact and easily comprehensible expressions from ESP corpora Keyword English for specific purpose, Domain-specific expression, Syntactic analysis, Classification model 1 English for Specific Purposes: ESP English for General Purposes

2 ESP ESP [1, 5, 7, 10] ESP (1) (2) / (3) [10] (1) (2) [10] ESP / [10] 2 ESP ESP [10] Multi-Word Expression 21 ESP 2 1 ESP 2 ESP 1 m [5] 1 MI-score B-score ESP bigram [1] 2 ESP BNC applied science British National Corpus (BNC) 9 [10] 2 9 ESP / [10] [1, 5] (1)

3 (2) (3) [10] [10] 22 [10] [10] 1 1 S t, S s (a) S = S t S s / M (b) M S α 2 (c) S t, S s (b) S t, S s (a) 2 (a) M F (b) f F f S s 3 S f F 1 1-(a) M [8] Boosting Algorithm for Tree Classification: BACT BACT Support Vector Machine 1 x, t y {+1, 1} h h t,y (x) = { y y t x otherwise t x t x t, y

4 Rank F Rank F 1 ) * 11 CDCD 2 Fig 12 algorithm 3 data 13 ] 4 image 14 performance 5 CD 15 using * 6 system 16 model 7 The * 17 results 8 images 18 CD 9 Figure 19 This * 10 = 20 are * 1: NLP-PROC 20 x ( ) sgn α t,y h t,y (x) t,y α t,y h t,y (x) 1-(b) α 1-(c) S 2 1 S s C-value[2, 3] a C-value ( C-value(a) = ( a 1) n(a) t(a) ) c(a) a a n(a) a t(a), c(a) a a n(a) a a 3 1 [10] Nature Scientific American NLP-PROC NLP-PROC NLP-PROC 135 C-value 13, C-value 100 A1 [10] CD ### * 0

5 NLP-PROC 1 ### 2 ### The can be 3 of the is a/the 4 Figure Figure CD shows Figure CD shows the 3 [10] 1 / Charniak parser 1 [11] 1 NP This paper

6 S1 S NP VP DT NN VBZ NP This paper proposes NP PP DT JJ NN IN NP a new approach for NN word NN similarity NN measurement 1: This paper proposes a new approach for word similarity measurement

7 32 22 [10] 2 4 [10] 2-(b) C-value C-value [10] of the that to s 2 Figure CD shows DT DT C-value for example, for example [10] b a C-value(a) C-value(b) a b b\a 1 b = for example, a = for example C-value(b) > C-value(a) 2 b\a =, a, b C-value b = for example, a = for example C-value C-value 2

8 33 [10] 2 S t, S s 0 S t, S s i = 1, 2,, n G t,i, G s,i i i i 1 (a) G i = G t,i G s,i G t,i /G s,i M i (b) M i G i α 2 (c) G t,i, G s,i (b) G t,i, G s,i (a) 2 (a) M i F i (b) f F i 32 f G s,i 3 G i f F i Web 41 [10] 14 ELPA 3 New Crown, New Horizon, Sunshine 4 Dream-Maker, English 21, Phoenix, Spectrum 21 ESP Nature The

9 53, ,840 Nature 81,330 1,974,934 ACL 159,080 3,861,550 2: Association for Computational Linguistics(ACL) 3 ESP TreeTagger 4 31 Charniak parser CD a/an/the DT 1-(b) α 099 BACT 10,000 2-(b) 7 7-gram : This paper S NP 2 NP 46,071 ACL 154, ,464 ACL 149, ACL 4 TreeTagger - a language independent part-of-speech tagger, 5 Eugene Charniak, 6 ACL all of that, each boy in DT group, DT picture story, DT United States goverment ACL that one, they both, that of knot and dale, japanese and arabic

10 S1 S NP VP This paper propose a new approach for word similarity measurement = : this paper 2: Rank F Rank F 1 ) 11 model 2 table 12 text 3 system 13 method 4 section 14 algorithm 5 sentence 15 user 6 we 16 noun 7 knowledge 17 DT * they 8 data 18 result 9 approach 19 rule 10 information 20 it 3: ACL 20 ACL C-value 1, A2 [10] ,325 [10] 1,633 3 DT * they table, section, figure ACL sentence, knowledge, semantic, corpus, algorithm DT DT f DT number of DT set of

11 S1 S ADVP, However, NP VP these issues will not be addressed here = : however, SUBJ PRED 3: However, these issues will not be addressed here S1 S NP VP This paper proposes a new approach for word similarity measurement = : SUBJ PRED 4: This paper proposes a new approach for word similarity measurement 43 2: 431 However, these issues will not be addressed here However,, 1 S NP,VP S S NP VP SUBJ,PRED However, 3 however, SUBJ PRED 4 SUBJ PRED 46,071 ACL 154,569

12 Rank F Rank F 1 section 11 thus 2 if * then 12 model 3 system 13 approach 4, * SUBJ * PRED 14 word 5 but *, * SUBJ * PRED 15 given 6 ) 16 each * PRED 7 DT * * * PRED 17, SUBJ * PRED 8 *, 18 and *, SUBJ * PRED 9 sentence 19, * SUBJ * always 10 then *, SUBJ * PRED 20 point 4: ACL ,276 ACL 152,756 7 ACL C-value 1, A3 [10] 4 According to DT bayes decision rule, SUBJ PRED In constructing DT power plant, SUBJ PRED 8 however, therefore, finally, thus for example, in addition, on DT other hand,, in particular word, SUBJ PRED corpus, SUBJ PRED 5 [10] 2 7 ACL no, SUBJ PRED, well, according to this science magazine, SUBJ PRED, or perhaps SUBJ PRED, after that, SUBJ PRED ACL but SUBJ PRED, so SUBJ PRED!, then DT few year ago, SUBJ PRED 8 [10]

13 [1],, :,, 12, pp (2005) [2] Franzti, K, Ananiadou: Extracting Nested Collocations, Proceedings of 16th International Conference on Computational Linguistics, pp (1996) [3] Franzti, K, Ananiadou, S and Mima, H: Automatic Recognition of Multi-Word Terms: the C-value/NC-value Method, Digital Libraries, Vol 3, No 2, pp (2000) [4] Kita, K, Kato Y, Omoto, T and Yano, Y: A Comparative Study of Automatic Extraction of Collocations from Corpora: Mutal Information vs Cost Criteria,, Vol 1, No 1, pp (1994) [5] : Multi-Word Expression, 216 ESP, pp (2008) [6] : -Nature -, (in press) [7] : 1110, (2009) [8], :,, Vol 45, No 9, pp (2004) [9],, :,, Vol 10, No 1, pp (2003) [10], : ESP, 233, pp (2009) [11] Tjong Kim Sang, E, F and Buchholz, S: Introduction to the CoNLL-2000 Shared Task: Chunking, Proceedings of CoNLL-2000 and LLL-2000, pp (2000) A A1 [10] Rank C-value Expression ### The of the [ CD ] [ CD ] ### in the

14 , the CD ] ### CD ] ### In [ CD , and to the ### Figure CD ### Figure can be ] ### ### Figure CD ) ### on the ### In this and the Figure CD CD, CD of a ### Figure CD : for the CD, ### A from the in [ CD ] ### For number of Figure CD is a CD, CD ] that the [ CD, CD ( CD ) ### Figure CD shows ### In this paper, we to be ### In this paper is the CD ) by the ### In this paper, the number of Figure CD : ### In the CD, CD ] ### [ CD, CD ] in a , CD in [ CD ] ### CD, CD, CD with the in [ CD In this paper, we Fig ### In this [ CD ], CD, CD, the same , CD ] based on [ CD, CD ] In this paper, , CD ] ### CD ) ### Figure CD shows , which set of , a ### CD in Fig ### used to as the [ CD, , CD, CD In this paper as a this paper, we a set of ### For example, to a ( CD shown in Fig ### CD : ### These ### Figure CD shows the CD ], this paper CD, CD, CD ] [ CD, CD, CD , CD, , it is with a ### ( shown in this paper, A2 Rank C-value Expression DT system this paper figure CD et al, CD DT number of table CD

15 DT word al, CD ) section CD DT user DT set of DT use of such as DT algorithm DT sentence DT result of DT text DT problem this approach DT following DT method DT rule DT corpus natural language DT grammar DT goal of DT input DT feature DT parser DT first CD as well as DT training this work this type of DT data DT verb DT task DT performance of DT document DT output DT purpose of , CD )(( DT information DT lexicon DT question this kind of DT best DT language DT advantage of DT value of DT example of DT dialogue ( eg, DT experiment DT difference DT process DT translation DT reason DT domain DT fact natural language processing DT term DT context DT analysis DT definition of DT evaluation DT size of for each described in this paper DT first step DT noun more than CD difference between DT accuracy DT effect of DT meaning of p ( DT idea DT notion of DT probability of DT relation DT list of only CD DT query DT discourse DT structure DT choice of DT decision DT basic idea as well DT remaining DT aim of this research DT representation DT generation et al, CD ; DT node each word DT program DT precision DT answer DT rest of DT combination of DT annotation part of speech DT source DT next section shown in figure CD presented in this paper DT application et al, CD )(( training data DT knowledge DT average number of DT total number of DT pattern DT learning than CD

16 ( b ) DT search DT constraint DT target DT development of DT network DT speaker DT remainder of this technique associated with DT case DT content of DT score DT dictionary language processing this study DT procedure DT speech we have DT majority of between DT CD DT tag relation between DT interpretation of DT tool information retrieval CD of them DT argument DT variety of DT time machine learning DT database scoring function fundamental problem sense distinction statistical analysis better performance sense extension annotation graph relevant document ultimate goal main task processing time input data error type surface form interpretation process each state not have each step main function linguistic structure processing technique anaphoric expression question classification annotation process further improvement main focus probability estimate score function important difference text meaning user utterance chart parser language learning parsing system query language empirical study database query candidate translation first argument similarity score word similarity key idea further analysis main clause clause with han character native speaker one with web page little attention A3 Rank C-value Expression however, SUBJ PRED for example, SUBJ PRED this paper, SUBJ PRED case, SUBJ PRED therefore, SUBJ PRED finally, SUBJ PRED thus, SUBJ PRED in addition, SUBJ PRED first, SUBJ PRED on DT other hand, word, SUBJ PRED , then SUBJ PRED in particular, SUBJ PRED for instance, SUBJ PRED furthermore, SUBJ PRED system, SUBJ PRED

17 in section CD, SUBJ PRED , SUBJ also PRED hand, SUBJ PRED second, SUBJ PRED moreover, SUBJ PRED experiment, SUBJ PRED that be, SUBJ PRED result, SUBJ PRED model, SUBJ PRED here, SUBJ PRED approach, SUBJ PRED in fact, SUBJ PRED in other word, SUBJ PRED figure CD, SUBJ PRED in general, SUBJ PRED text, SUBJ PRED as DT result, SUBJ PRED corpus, SUBJ PRED sentence, SUBJ PRED data, SUBJ PRED above, SUBJ PRED in contrast, SUBJ PRED work, SUBJ PRED language, SUBJ PRED so, SUBJ PRED method, SUBJ PRED problem, SUBJ PRED similarly, SUBJ PRED instead, SUBJ PRED task, SUBJ PRED time, SUBJ PRED et al, CD ) table CD, SUBJ PRED set, SUBJ PRED unfortunately, SUBJ PRED hence, SUBJ PRED next, SUBJ PRED way, SUBJ PRED process, SUBJ PRED context, SUBJ PRED information, SUBJ PRED domain, SUBJ PRED nevertheless, SUBJ PRED , as of course, SUBJ PRED for example, if specifically, SUBJ PRED application, SUBJ PRED study, SUBJ PRED DT number of consequently, SUBJ PRED al, CD ) document, SUBJ PRED algorithm, SUBJ PRED level, SUBJ PRED rule, SUBJ PRED english, SUBJ PRED again, SUBJ PRED , SUBJ often PRED recently, SUBJ PRED analysis, SUBJ PRED currently, SUBJ PRED relation, SUBJ PRED indeed, SUBJ PRED additionally, SUBJ PRED grammar, SUBJ PRED reason, SUBJ PRED step, SUBJ PRED structure, SUBJ PRED secondly, SUBJ PRED term, SUBJ PRED view, SUBJ PRED tree, SUBJ PRED feature, SUBJ PRED in practice, SUBJ PRED stage, SUBJ PRED sense, SUBJ PRED phase, SUBJ PRED purpose, SUBJ PRED such as at DT same time, SUBJ , SUBJ only PRED question, SUBJ PRED point of view, SUBJ PRED third, SUBJ PRED evaluation, SUBJ PRED point, SUBJ PRED end, SUBJ PRED generation, SUBJ PRED further, SUBJ PRED knowledge, SUBJ PRED rather, SUBJ PRED firstly, SUBJ PRED all, SUBJ PRED in DT next section, SUBJ below, SUBJ PRED phrase, SUBJ PRED typically, SUBJ PRED , SUBJ still PRED be found, SUBJ PRED verb, SUBJ PRED overall, SUBJ PRED representation, SUBJ PRED found, SUBJ PRED DT consequence, SUBJ PRED as DT consequence, SUBJ PRED constraint, SUBJ PRED otherwise, SUBJ PRED error, SUBJ PRED earlier, SUBJ PRED processing, SUBJ PRED translation, SUBJ PRED

18 clearly, SUBJ PRED training data, SUBJ PRED present, SUBJ PRED comparison, SUBJ PRED in DT following section, SUBJ pair, SUBJ PRED name, SUBJ PRED noun, SUBJ PRED lexicon, SUBJ PRED , SUBJ usually PRED first of all, SUBJ PRED more than CD intuitively, SUBJ PRED dialogue, SUBJ PRED framework, SUBJ PRED so far, SUBJ PRED as shown in obviously, SUBJ PRED used, SUBJ PRED type, SUBJ PRED test, SUBJ PRED query, SUBJ PRED system with annotated with polysemous word not provide not directly such system general case text structure same as different domain many application lexical entry different type preliminary experiment test collection world knowledge problem with processing system user utterance early stage z ) defined as combination with document with target word word meaning not found since word english sentence aligned sentence text planner language modeling be using since both parallel corpus other approach linguistic feature mapping between semantic constraint correlation between feature value data structure each type verb sense context-free grammar previous experiment generation process learning approach learning method local context

1 Fig. 1 Extraction of motion,.,,, 4,,, 3., 1, 2. 2.,. CHLAC,. 2.1,. (256 ).,., CHLAC. CHLAC, HLAC. 2.3 (HLAC ) r,.,. HLAC. N. 2 HLAC Fig. 2

1 Fig. 1 Extraction of motion,.,,, 4,,, 3., 1, 2. 2.,. CHLAC,. 2.1,. (256 ).,., CHLAC. CHLAC, HLAC. 2.3 (HLAC ) r,.,. HLAC. N. 2 HLAC Fig. 2 CHLAC 1 2 3 3,. (CHLAC), 1).,.,, CHLAC,.,. Suspicious Behavior Detection based on CHLAC Method Hideaki Imanishi, 1 Toyohiro Hayashi, 2 Shuichi Enokida 3 and Toshiaki Ejima 3 We have proposed a method for