Toward Finding Semantic Relations not Written in a Single Sentence: An Inference Method using Auto-Discovered Rules

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 902–910, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Toward Finding Semantic Relations not Written in a Single Sentence: An Inference Method using Auto-Discovered Rules Masaaki Tsuchida†§∗ Kentaro Torisawa‡ Stijn De Saeger ‡ Jong-Hoon Oh‡ Jun’ichi Kazama‡ Chikara Hashimoto‡ Hayato Ohwada§ † Information and Media Processing Laboratories, NEC Corporation, Nara, Japan m-tsuchida@cq.jp.nec.com ‡ Information Analysis Laboratory, National Institute of Information and Communications Technology, Kyoto, Japan {torisawa,stijn,rovellia,kazama,ch}@nict.go.jp § Graduate school of Science and Technology, Tokyo University of Science, Chiba, Japan. ohwada@ia.noda.tus.ac.jp Abstract Recent advances in automatic knowledge acquisition methods have enabled us to construct massive knowledge bases of se- mantic relations. Most previous work has focused on semantic relations explicitly expressed in single sentences. Our goal in this work is to obtain valid non-single sentence relation instances, which are not written in any single sentence and may not be even written in a large corpus. We de- velop a method to infer new semantic rela- tion instances by applying auto-discovered inference rules, and show that our method inferred a considerable number of valid in- stances that were not written in single sen- tences even in 600 million Web pages. 1 Introduction Recent advances in automatic relation acquisi- tion methods (Agichtein and Luis, 2001; Et- zioni et al., 2004; Pantel and Pennacchiotti, 2006; Pas¸ca et al., 2006; Banko and Etzioni, 2008; De Saeger et al., 2009) have opened the way to build large knowledge bases con- taining a huge number of semantic relation in- stances such as “CAUSE(allergen, allergy)” and “PREVENTION(coffee, drowsiness)”. Such a mas- sive knowledge base is valuable for applications like innovation and risk management (Torisawa et al., 2010), like finding potential and unexpected causes of a disease, unknown side-effects of a drug, and so on. In such tasks overlooking a small piece of information can have grave consequences. In this work, our goal is to acquire Non-single Sentence relation instances (NS instances), which ∗This work was done when the author was at the National Institute of Information and Communications Technology. are not written in any single sentences, in order to reduce the possibility of overlooking valuable information. More precisely, if the two nouns in an instance (e.g., “allergen” and “allergy” in “CAUSE(allergen, allergy)”) do not appear in any single sentences in a given corpus together with the other clues indicating the relation type (e.g., the verb “cause”), the instance is an NS instance. NS instances may be indirectly written in a se- quence of several sentences using anaphora or may not be written at all even in a given corpus. In the following, we call the complement of NS in- stances Single Sentence relation instances (SS in- stances), which consist of two nouns that co-occur in some single sentences in a given corpus together with some evidence for a relation type. Most existing relation acquisition methods ac- quire relation instances using lexico-syntactic pat- terns (Pantel and Pennacchiotti, 2006; Pas¸ca et al., 2006; De Saeger et …

Content truncated to 500 words