Large-Scale Verb Entailment Acquisition from the Web

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1172–1181, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Large-Scale Verb Entailment Acquisition from the Web Chikara Hashimoto∗ Kentaro Torisawa† Kow Kuroda‡ Stijn De Saeger§ Masaki Murata¶ Jun’ichi Kazama∥ National Institute of Information and Communications Technology Sorakugun, Kyoto, 619-0289, JAPAN {∗ ch,† torisawa,‡ kuroda,§ stijn,¶murata,∥ kazama}@nict.go.jp Abstract Textual entailment recognition plays a fundamental role in tasks that require in- depth natural language understanding. In order to use entailment recognition tech- nologies for real-world applications, a large-scale entailment knowledge base is indispensable. This paper proposes a con- ditional probability based directional sim- ilarity measure to acquire verb entailment pairs on a large scale. We targeted 52,562 verb types that were derived from 108 Japanese Web documents, without regard for whether they were used in daily life or only in specific fields. In an evaluation of the top 20,000 verb entailment pairs ac- quired by previous methods and ours, we found that our similarity measure outper- formed the previous ones. Our method also worked well for the top 100,000 re- sults. 1 Introduction We all know that if you snored, you must have been sleeping, that if you are divorced, you must have been married, and that if you won a lawsuit, you must have sued somebody. These relation- ships between events where one is the logical con- sequence of the other are called entailment. Such knowledge plays a fundamental role in tasks that require in-depth natural language understanding, e.g., answering questions and using natural lan- guage interfaces. This paper proposes a novel method for verb entailment acquisition. Using a Japanese Web corpus (Kawahara and Kurohashi, 2006a) derived from 108 Japanese Web documents, we automat- ically acquired such verb pairs as snore → sleep and divorce → marry, where entailment holds be- tween the verbs in the pair.1 Our definition of “en- tailment” is the same as that in WordNet3.0; v1 entails v2 if v1 cannot be done unless v2 is, or has been, done.2 Our method follows the distributional similar- ity hypothesis, i.e., words that occur in the same context tend to have similar meanings. Just as in the methods of Lin and Pantel (2001) and Szpek- tor and Dagan (2008), we regard the arguments of verbs as the context in the hypothesis. How- ever, unlike the previous methods, ours is based on conditional probability and is augmented with a simple trick that improves the accuracy of verb entailment acquisition. In an evaluation of the top 20,000 verb entailment pairs acquired by the pre- vious methods and ours, we found that our similar- ity measure outperformed the previous ones. Our method also worked well for the top 100,000 re- sults, Since the scope of Natural Language Process- ing (NLP) has advanced from a formal writing style to a colloquial style and from restricted to open domains, it is necessary for the language re- sources for NLP, including verb entailment knowl- edge bases, to cover a broad range of …

Content truncated to 500 words