Relation Acquisition using Word Classes and Partial Patterns

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 825–835, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Relation Acquisition using Word Classes and Partial Patterns Stijn De Saeger†∗ Kentaro Torisawa† Masaaki Tsuchida§ Jun’ichi Kazama† Chikara Hashimoto† Ichiro Yamada‡ Jong Hoon Oh† Istv´an Varga† Yulan Yan† † Information Analysis Laboratory, National Institute of Information and Communications Technology, 619-0289 Kyoto, Japan {stijn,torisawa,kazama,ch,rovellia,istvan,yulan}@nict.go.jp § Information and Media Processing Laboratories, NEC Corporation, 630-0101 Nara, Japan m-tsuchida@cq.jp.nec.com ‡ Human & Information Science Research Division, NHK Science & Technology Research Laboratories, 157-8510 Tokyo, Japan yamada.i-hy@nhk.or.jp Abstract This paper proposes a semi-supervised rela- tion acquisition method that does not rely on extraction patterns (e.g. “X causes Y” for causal relations) but instead learns a combi- nation of indirect evidence for the target re- lation — semantic word classes and partial patterns. This method can extract long tail instances of semantic relations like causality from rare and complex expressions in a large Japanese Web corpus — in extreme cases, pat- terns that occur only once in the entire cor- pus. Such patterns are beyond the reach of cur- rent pattern based methods. We show that our method performs on par with state-of-the-art pattern based methods, and maintains a rea- sonable level of accuracy even for instances acquired from infrequent patterns. This abil- ity to acquire long tail instances is crucial for risk management and innovation, where an ex- haustive database of high-level semantic rela- tions like causation is of vital importance. 1 Introduction Pattern based relation acquisition methods rely on lexico-syntactic patterns (Hearst, 1992) for extract- ing relation instances. These are templates of natu- ral language expressions such as “X causes Y ” that signal an instance of some semantic relation (i.e., causality). Pattern based methods (Agichtein and Gravano, 2000; Pantel and Pennacchiotti, 2006b; Pas¸ca et al., 2006; De Saeger et al., 2009) learn many ∗ This work was done when all authors were at the National Institute of Information and Communications Technology. such patterns to extract new instances (word pairs) from the corpus. However, since extraction patterns are learned us- ing statistical methods that require a certain fre- quency of observations, pattern based methods fail to capture relations from complex expressions in which the pattern connecting the two words is rarely observed. Consider the following sentence: “Curing hypertension alleviates the deteriora- tion speed of the renal function, thereby lower- ing the risk of causing intracranial bleeding” Humans can infer that this sentence expresses a causal relation between the underlined noun phrases. But the actual pattern connecting them, i.e., “Cur- ing X alleviates the deterioration speed of the re- nal function, thereby lowering the risk of causing Y ”, is rarely observed more than once even in a 108 page Web corpus. In the sense that the term pat- tern implies a recurring event, this expression con- tains no pattern for detecting the causal relation be- tween hypertension and intracranial bleeding. This is what we mean by “long tail instances” …

Content truncated to 500 words