Toward a Discourse Theory for Annotating Causal Relations in Japanese

PACLIC 28 !460 Toward a Discourse Theory for Annotating Causal Relations in Japanese Kimi Kaneko Ochanomizu University kaneko.kimi@is.ocha.ac.jp Daisuke Bekki Ochanomizu University National Institute of Informatics CREST, JST bekki@is.ocha.ac.jp Abstract We present a revised discourse theory based on segmented discourse represen- tation theory and provide a method for building a Japanese corpus suitable for causal relation extraction. This extends and refines the framework proposed in Kaneko and Bekki (2014), and we evalu- ate our corpus and compare it with that work. 1 Introduction In recent years, considerable attention has been paid to deep semantic processing. Many stud- ies, including Bethard et al. (2008), Inui et al. (2007), Inui et al. (2003), and Riaz and Girju (2013), have recently been conducted on deep semantic processing, and causal relation extrac- tion (hereinafter, CRE) is one of the specific tasks of deep semantic processing. Research on CRE is still progressing, and there are many ob- stacles that must be overcome. In Inui et al. (2003), cause and effect pairs were acquired from Japanese texts by using keywords such as “node” and “kara”. In (1), for example, the antecedent ame-ga hut-ta (“it rained”) denotes an event taken as a cause, and the consequent mizutamari-ga dekita (“puddles emerged”) denotes an event taken as an effect. (1) Ame-ga rain-NOM hut-ta-node fall-past-because mizutamari-ga puddles-NOM deki-ta. emerge-past ‘Because it rained, puddles emerged.’ However, antecedents do not always denote causes or reasons for consequents, as illustrated by the following example. (2) Kesa this.morning kubi-ga neck-NOM itakat-ta-node have.a.pain-past-because netigae-ta-no strain.my.neck-past-attr daroo. may. ‘Because I had a pain in my neck this morning, I might have strained my neck while sleeping.’ In example (2), the antecedent kesa kubi-ga itakat-ta (“I had a pain in my neck this morn- ing”) is not taken as the cause of the consequent netigae-ta (“I strained my neck while sleeping”) but as the basis for the judgment expressed by the consequent. In this example, the consequent denotes the cause, and the antecedent denotes its effect. For a computer to automatically rec- ognize causal relations in text, it is important to distinguish cases like (2) from cases like (1). However, existing studies have not dealt with these kinds of problems. To solve such problems, Kaneko and Bekki (2014) (henceforth K&B) analyzed the infor- mation necessary for acquiring more accurate cause–effect knowledge and proposed a method for creating a Japanese corpus suitable for CRE. However, as is explained below, some problems remain: first, the coverage of discourse relations is not sufficient; second, annotating at two lev- els (i.e. fact and epistemic levels) is sometimes redundant. Copyright 2014 by Kimi Kaneko and Daisuke Bekki 28th Pacific Asia Conference on Language, Information and Computation pages 460–469 PACLIC 28 !461 We try to solve these remaining problems in K&B; toward that end, we propose a new method for building a Japanese corpus for causal relation extraction. In addition, we evaluate the validity of our method in terms of agreement and frequency, and analyze the results. 2 Previous Studies In this section, …

Content truncated to 500 words