自然言語処理19_3

Size: px

Start display at page:

Download "自然言語処理19_3"

そよゆのもと
7 years ago
Views:

1 Wikipedia, Stijn De Saeger 1Q84 Wikipedia 2 1,925, % 2,719, % 6,347,472 Wikipedia Generating Information-Rich Taxonomy Using Wikipedia Ichiro Yamada, Chikara Hashimoto, Jong-Hoon Oh, Kentaro Torisawa, Kow Kuroda,, Stijn De Saeger, Masaaki Tsuchida and Jun ichi Kazama Hyponymy relation acquisition has been extensively studied. However, the informativeness of acquired hypernyms has not been sufficiently discussed. We found that the hypernyms in automatically acquired hyponymy relations are often too vague for their hyponyms. For instance, work is a vague hypernym for work Seven Samurai and work 1Q84. These vague hypernyms sometimes cause the lower accuracy for NLP applications such as information retrieval or question answering. In this paper, we propose a method of making (vague) hypernyms more specific ex- NHK, Science Technology Research Laboratory, Japan Broadcasting Corporation, National Institute of Information and Communications Technology, Kyoto University, Conprehensive Research Organization, Waseda University, NEC Corporation

2 Vol. 19 No. 1 March 2012 ploting Wikipedia. For instance, our method generates two intermediate nodes work by Akira Kurosawa and work by film director for a original hyponymy relation work Seven Samurai. We show that our method acquires 2,719,441 hyponymy relations with the first intermediate concepts (such as work by Akira Kurosawa ) with 85.3% weighted precision and 6,347,472 hyponymy relations with the second intermediate concepts (such as work by film director ) with 78.6% weighted precision. Furthermore, we confirm that hyponymy relaitons acquired by our method can be interpreted as object attribute value. Key Words: Hyponymy relation acquisition, Object-attribute-value acquisition, Wikipedia 1 (Hearst 1992; Hovy, Kozareva, and Riloff 2009; Oh, Uchimoto, and Torisawa 2009; Ponzetto and Strube 2007; 2009; Suchanek, Kasneci, and Weikum 2007; Nastase and Strube 2008; Snow, Jurafsky, and Ng 2005) A B A B 1 (1) (2) (3) 3 (1) (3) 2 C A B A B C B A 1 A B A B 4

3 Saeger, Wikipedia Wikipedia A B Wikipedia Wikipedia 2,719, % 2 3 Wikipedia ( 2009) 4 Wikipedia ( 2009) 5

4 Vol. 19 No. 1 March A B (Hearst 1992) , , , ,542 73,995 TV 54,145 TV 53,591 51,971 47,399 32,305 26,883 26,764 23,325 22,000 20,742 18,831 18,481 17,443 17,072 14,990 2 TREC QA task (Dang, Lin, and Kelly 2006, 2007) 6

5 Saeger, Wikipedia 3 Wikipedia Wikipedia ( 2009) 4 Wikipedia 1 Wikipedia Apple Apple ipod iphone 1 Wikipedia 7

6 Vol. 19 No. 1 March 2012 Mac mini MacBook MacBook Air term 1 Wikipedia Wikipedia MediaWiki 2 MediaWiki == == *** Mac mini 2 term 1 1 Mac mini Mac mini SVM (Vapnik 1995) X = * 1 2 Media Wiki 8

7 Saeger, Wikipedia 3 Wikipedia 29, Wikipedia 1,925,676 90% 3(a) 4 Wikipedia 2 G Wikipedia 522,709 1,472,035 90% ) ) ) T- ( ) G- ( T- ( ) ) ( ) ( ) ( ) (a) (b) T- (c) G index.html 9

8 Vol. 19 No. 1 March 2012 Wikipedia 4.1 Wikipedia T- T- T- 3(b) T- T- G- G- T- 4.2 T- G- G- 3(c) 4.1 T- Wikipedia Wikipedia T- Wikipedia T- T- Wikipedia T- T- T- 3(b) 4.2 G- T- Wikipedia T- Wikipedia G- T- Wikipedia G- G- Wikipedia Wikipedia 3 Wikipedia 1 10

9 Saeger, Wikipedia SVM G- T- T- G- G Wikipedia G- 2 G- G- 4 G- G- T- T- Wikipedia 1,925,676 6,347,472 G- 2,719,441 T- T- 2,719,441 T- 2,113,040 Wikipedia 2 2 G- G- T- SF WALL-E M.O IRIS Crimson J D A Boy in France 11

10 Vol. 19 No. 1 March G- T- T- 342,884 Wikipedia G- G T- G- G- 22 T- G- 200 T- 200 G- G- 178 Good: Less good: Bad: 4 Kappa

11 Saeger, Wikipedia 3 Good Less good Bad Good (100/200) (92/200) (8/200) G (125/178) (30/178) (23/178) T (170/200) (1/200) (29/200) = #Good 1 + #Less good #Bad 0 #Good + #Less good + #Bad #Good #Less good #Bad Good 1 Bad Less good Pasca (Pasca 2007, 2009) Good 3 G- T- Good SVM SVM Wikipedia ( 2009) 5 SVM Good Less Good SVM G- SVM 3 G- 4 SVM SVM (1) 5 13

12 Vol. 19 No. 1 March SVM Good Less good Bad Good G (97/178) (30/178) (51/178) 3 Good Less good Bad G- T- Less good Good Bad T % 2 Wikipedia T- 3.4% 3 Wikipedia 69.0% 1 3 wrapper 1 2 Wikipedia 14

13 Saeger, Wikipedia Wikipedia 3 Wikipedia T % 3 3 term T- G- Wikipedia term 30 term 3 5 T G G- 5 Good Less good Bad Good (75/150) (70/150) (5/150) G (99/129) (22/129) (8/129) T (140/150) (1/150) (9/150) 15

14 Vol. 19 No. 1 March % T- 8.4% T- 2,719,441 1,958,117 G- 6,347,472 4,960,751 6 G- T- 6.1 G- G- G- 6 G- 20 G- 6 G- 20 G- G- 59,890 G- 54, Kappa G- G- G- TV 6 16

15 Saeger, Wikipedia 78.0% 20 54, % 1,199, % 6.2 T- T- Wikipedia Wikipedia Wikipedia 3 T- T- Wikipedia IRIS Crimson T- IRIS Crimson 5 T- term Wikipedia term Wikipedia term T- T- 5 G- 200 T- Wikipedia Wikipedia T- SVM 17

16 Vol. 19 No. 1 March Wikipedia Wikipedia T Vital: Okay: Wrong: Kappa (1) Vital 1.0 Okey 0.5 Wrong 0 (Pasca 2007, 2009) 7 T- 94.0% T- 3 T T- Vital Okay Wrong T (188/200) (0/200) (12/200) (105/200) (4/200) (91/200)

17 Saeger, Wikipedia 53.5% Wikipedia 2 term 2 term Wikipedia term term 7 (Hearst 1992; Ando, Sekine, and Ishizaki 2004) (Pantel and Ravichandran 2004; Etzioni, Cafarella, Downey, Popescu, Shaked, Soderland, Weld, and Yates 2005) HTML (Shinzato and Torisawa 2004) Wikipedia ( 2009; Oh et al. 2009; Yamada, Torisawa, Kazama, Kuroda, Murata, De Saeger, Bond, and Sumida 2009) Hovy (Hovy et al. 2009) Hovy Doubly-Anchored Pattern bootstrap people / Shakespeare writers animals people 2 Wikipedia Wikipedia (Kazama and Torisawa 2007; Ponzetto and Strube 2007; Suchanek et al. 2007; Nastase and Strube 2008; 2009; Oh et al. 2009; Yamada et al. 2009) Wikipedia Wikipedia 19

18 Vol. 19 No. 1 March Wikipedia 2,719,441 T- 85.3% 6,347,472 G- 78.6% 1,958,117 T- 93.7% 4,960,751 G- 85.3% ( 2009) 1,925, % G- 20 G- 59, % (Dang et al. 2006, 2007) T- Wikipedia Wikipedia T- Torisawa (Torisawa 2001) 2 Torisawa Ando, M., Sekine, S., and Ishizaki, S. (2004). Automatic Extraction of Hyponyms from Japanese Newspaper Using Lexico-syntactic Patterns. In Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC), pp Dang, H., Lin, J., and Kelly, D. (2006). Overview of the TREC 2006 Question Answering 20

19 Saeger, Wikipedia Track. In Proceedings of the Fifteenth Text REtrieval Conference. Dang, H., Lin, J., and Kelly, D. (2007). Overview of the TREC 2007 Question Answering Track. In Proceedings of the Sixteenth Text REtrieval Conference. Etzioni, O., Cafarella, M., Downey, D., Popescu, A.-M., Shaked, T., Soderland, S., Weld, D. S., and Yates, A. (2005). Unsupervised named-entity extraction from the web: An experimental study. Artificial Intelligence, 165 (1), pp Hearst, M. A. (1992). Automatic Acquisition of Hyponyms from Large Text Corpora. In Proceedings of the 14th conference on Computational Linguistics (COLING), pp Hovy, E., Kozareva, Z., and Riloff, E. (2009). Toward Completeness in Concept Extraction and Classification. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp Kazama, J. and Torisawa, K. (2007). Exploiting Wikipedia as External Knowledge for Named Entity Recognition. In Proceedings of the Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pp Nastase, V. and Strube, M. (2008). Decoding Wikipedia Categories for Knowledge Acquisition. In Proceedings of the 23rd AAAI Conference on Artificial Intelligence (AAAI), pp Oh, J.-H., Uchimoto, K., and Torisawa, K. (2009). Bilingual Co-Training for Monolingual Hyponymy-Relation Acquisition. In Proceedings of ACL-09: IJCNLP, pp Pantel, P. and Ravichandran, D. (2004). Automatically Labeling Semantic Classes. In Proceedings of the Human Language Technology and North American Capter of the Association for Computational Linguistics Coference (HLT-NAACL), pp Pasca, M. (2007). Organizing and Searching the World Wide Web of Facts Step Two: Harnessing the Wisdom of the Crowds. In Proceedings of the 16th World Wide Web Conference (WWW), pp Pasca, M. (2009). Outclassing Wikipedia in Open-Domain Information Extraction: Weakly- Supervised Acquisition of Attributes over Conceptual Hierarchies. In Proceedings of the 12th Conference of Europian Chapter of the Association of Computational Linguistics (EACL), pp Ponzetto, S. P. and Strube, M. (2007). Deriving a Large-Scale Taxonomy from Wikipedia. In Proceeding of the 22nd Conference on the Advancement of Artificial Intelligence (AAAI), pp Shinzato, K. and Torisawa, K. (2004). Extracting Hyponyms of Prespecified Hypernyms from 21

20 Vol. 19 No. 1 March 2012 Itemizations and Headings in Web Documents. In Proceedings of the 20th Conference on Computational Linguistics (COLING), pp Snow, R., Jurafsky, D., and Ng, A. Y. (2005). Learning Syntactic Patterns for Automatic Hypernym Discovery. In Proceedings of the Neural Information Processing Systems (NIPS). Suchanek, F. M., Kasneci, G., and Weikum, G. (2007). Yago: A Core of Semantic Knowledge. In Proceedings of the 16th World Wide Web Conference (WWW), pp Torisawa, K. (2001). An Unsuperveised Method for Canonicalization of Japanese Postpositions. In Proceedings of the 6th Natural Language Processing Pacific Rim Symposium (NLPRS), pp Vapnik, V. N. (1995). The Nature of Statistical Learning Theory. Springer-Verlag New York, Inc., New York, USA. Yamada, I., Torisawa, K., Kazama, J., Kuroda, K., Murata, M., De Saeger, S., Bond, F., and Sumida, A. (2009). Hypernym Discovery Based on Distributional Similarity and Hierarchical Structures. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp Association for Computational Linguistics. (2009). Wikipedia., 16 (3), pp NHK NHK KAIST KAIST

21 Saeger, Wikipedia Stijn De Saeger: NICT MASTAR NEC NEC

114 583/4 2012

114 583/4 2012 5-5 Fundamental Language Resources HASHIMOTO Chikara, Jong-Hoon Oh, SANO Motoki, and KAWADA Takuya Fundamental language resources are classifi ed into natural language processing tools and natural language