Automatic Detection of Metonymies using Associative Relations between Words

Automatic Detection of Metonymies using Associative Relations between Words Takehiro Teraoka (teraoka@sfc.keio.ac.jp) Graduate School of Media and Governance, Keio University, 5322 Endoh, Fujisawa, Kanagawa, Japan Ryuichiro Higashinaka (higashinaka.ryuichiro@lab.ntt.co.jp) NTT Cyber Space Laboratories, NTT Corporation, 1-1 Hikarinooka Yokosuka, Kanagawa 239-0847 Japan Jun Okamoto (juno@kaetsu.ac.jp) Department of Business Innovation, Kaetsu University, 2-8-4 Hanakoganeiminamicho, Kodaira, Tokyo 187-0003, Japan Shun Ishizaki (ishizaki@sfc.keio.ac.jp) Graduate School of Media and Governance, Keio University, 5322 Endoh, Fujisawa, Kanagawa, Japan Abstract Table 1: Metonymic expressions with spatial contiguity and temporal contiguity. It is crucial for computers to detect metonymic expressions be- cause sentences including them may have different meanings from literal ones. In previous studies, detecting metonymies has been done mainly by rule-based and statistical approaches. The problem of current metonymy detection is that using syn- tactic and semantic information may be not enough to de- tect metonymic expressions. In this study, we propose an approach for detecting them with associative information be- tween words. We evaluated our method by comparing it with a baseline that uses syntactic and semantic information. As a result, our method showed significantly better accuracy (0.84) of judging words as metonymic or literal expressions than that of the baseline. Keywords: Metonymy; Association Experiment; Associative Concept Dictionaries; Verbs; Nouns Metonymic patterns -spatial contiguity- Container for Content Producer for Product Controller for Controlled Object Used for User Material for Product Others Introduction Metonymy is a figure of speech, where one item’s name represents another which usually has a close relation with the first one. The metonymic relation, as shown in Table 1 (Lakoff & Johnson, 1980; Taniguchi, 2003; Yamanashi, 1988), has different patterns which are classified predomi- nately into two types: spatial contiguity and temporal con- tiguity (Taniguchi, 2003). Below is a Japanese example for ‘Container for Content’: kare-ga isshoubin-wo nomihoshita (He drank up a 1.8-liter bottle.) The Japanese sentence above means literally that he drank up the bottle. Of course, it does not mean that he drank or ate the bottle itself, but its content, usually Japanese sake. Japanese sake is generally in a large bottle made from glass, and called bin in Japanese. It has a capacity of 1.8 liters, isshou. Therefore, the above example sentence where is- shoubin is a metonymic expression means that he drank up Japanese sake in a 1.8-liter bottle. Since a sentence includ- ing metonymy is grammatically correct on a literal level, it is difficult for computers to grasp its true meaning as humans do. Metonymic patterns -temporal contiguity- Result for Cause Cause for Result Examples of sentences (metonymic reading) kare-ha glass-wo nonda ‘He drank the glass (liquid).’ kare-ha Mahler-wo kiita ‘He listened to Mahler (symphony).’ Nixon-ga Hanoi-wo bakugekishita ‘Nixon (government) bombed Hanoi.’ gakuseifuku-ga aruiteiru ‘The school uniform (student) is walking.’ kare-ha caffeine-wo nonda ‘He drank caffeine (soft drink).’ riron-ga sore-wo jisshoushita ‘The theory (proposer) claimed that.’ Examples of sentences (metonymic reading) kanojo-ga sekimensuru ‘She is blushing.’(She is ashamed) kare-ga sakazuki-wo katamukeru ‘He is tipping the sake cup.’ (He is drinking the Japanese sake) In English metonymy detection, most previous studies have taken mainly rule-based and statistical approaches. The rule-based approach uses semantic networks and hand- crafted rules to detect metonymies (Bouaud, Bachimont, & Zweigenbaum, 1996; Fass, 1991; Iverson & Helmre- ich, 1992). The representative work of statistical approach used corpus-based metonymy resolution on location names (Markert & Nissim, 2003). Moreover, by using syntactic, semantic, encyclopedic, or collocation information as ma- chine learning features, some conventional studies for detect- ing metonymic expressions were suggested (Markert & Nis- sim, 2007; Nastase & Stube, 2009). Their methods are effec- tive, but they only dealt with metonymies on country names and companies. When considering the variety of metonymic patterns in Table 1, it is desirable to be able to detect various