Hypernym Discovery Based on Distributional Similarity and Hierarchical Structures

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 929–937, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Hypernym Discovery Based on Distributional Similarity and Hierarchical Structures Ichiro Yamada†, Kentaro Torisawa†, Jun’ichi Kazama†, Kow Kuroda†, Masaki Murata†, Stijn De Saeger†, Francis Bond† and Asuka Sumida‡ †National Institute of Information and Communications Technology 3-5 Hikaridai, Keihannna Science City 619-0289, JAPAN {iyamada,torisawa,kazama,kuroda,murata,stijn,bond}@nict.go.jp ‡Japan Advanced Institute of Science and Technology 1-1 Asahidai, Nomi-shi, Ishikawa-ken 923-1211, JAPAN a-sumida@jaist.ac.jp Abstract This paper presents a new method of devel- oping a large-scale hyponymy relation data- base by combining Wikipedia and other Web documents. We attach new words to the hy- ponymy database extracted from Wikipedia by using distributional similarity calculated from documents on the Web. For a given tar- get word, our algorithm first finds k similar words from the Wikipedia database. Then, the hypernyms of these k similar words are assigned scores by considering the distribu- tional similarities and hierarchical distances in the Wikipedia database. Finally, new hy- ponymy relations are output according to the scores. In this paper, we tested two distribu- tional similarities. One is based on raw verb- noun dependencies (which we call “RVD”), and the other is based on a large-scale clus- tering of verb-noun dependencies (called “CVD”). Our method achieved an attachment accuracy of 91.0% for the top 10,000 rela- tions, and an attachment accuracy of 74.5% for the top 100,000 relations when using CVD. This was a far better outcome com- pared to the other baseline approaches. Ex- cluding the region that had very high scores, CVD was found to be more effective than RVD. We also confirmed that most relations extracted by our method cannot be extracted merely by applying the well-known lexico- syntactic patterns to Web documents. 1 Introduction Large-scale taxonomies such as WordNet (Fell- baum 1998) play an important role in informa- tion extraction and question answering. However, extremely high costs are borne to manually en- large and maintain such taxonomies. Thus, appli- cations using these taxonomies tend to face the drawback of data sparseness. This paper presents a new method for discovering a large set of hy- ponymy relations. Here, a word1 X is regarded as a hypernym of a word Y if Y is a kind of X or Y is an instance of X. We are able to generate large-scale hyponymy relations by attaching new words to the hyponymy database extracted from Wikipedia (referred to as “Wikipedia relation database”) by using distributional similarity cal- culated from Web documents. Relations ex- tracted from Wikipedia are relatively clean. On the other hand, reliable distributional similarity can be calculated using a large number of docu- ments on the Web. In this paper, we combine the advantages of these two resources. Using distributional similarity, our algorithm first computes k similar words for a target word. Then, each k similar word assigns a score to its ancestors in the hierarchical structures of the Wikipedia relation database. The hypernym that has the highest score for …

Content truncated to 500 words