The Effect of Corpus Size on Case Frame Acquisition for Discourse Analysis

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 521–529, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics The Effect of Corpus Size on Case Frame Acquisition for Discourse Analysis Ryohei Sasano Graduate School of Informatics, Kyoto University sasano@i.kyoto-u.ac.jp Daisuke Kawahara National Institute of Information and Communications Technology dk@nict.go.jp Sadao Kurohashi Graduate School of Informatics, Kyoto University kuro@i.kyoto-u.ac.jp Abstract This paper reports the effect of corpus size on case frame acquisition for discourse analysis in Japanese. For this study, we collected a Japanese corpus consisting of up to 100 bil- lion words, and constructed case frames from corpora of six different sizes. Then, we ap- plied these case frames to syntactic and case structure analysis, and zero anaphora resolu- tion. We obtained better results by using case frames constructed from larger corpora; the performance was not saturated even with a corpus size of 100 billion words. 1 Introduction Very large corpora obtained from the Web have been successfully utilized for many natural lan- guage processing (NLP) applications, such as prepo- sitional phrase (PP) attachment, other-anaphora res- olution, spelling correction, confusable word set dis- ambiguation and machine translation (Volk, 2001; Modjeska et al., 2003; Lapata and Keller, 2005; At- terer and Sch¨utze, 2006; Brants et al., 2007). Most of the previous work utilized only the sur- face information of the corpora, such as n-grams, co-occurrence counts, and simple surface syntax. This may be because these studies did not require structured knowledge, and for such studies, the size of currently available corpora is considered to have been almost enough. For instance, while Brants et al. (2007) reported that translation quality continued to improve with increasing corpus size for training language models at even size of 2 trillion tokens, the increase became small at the corpus size of larger than 30 billion tokens. However, for more complex NLP tasks, such as case structure analysis and zero anaphora resolution, it is necessary to obtain more structured knowledge, such as semantic case frames, which describe the cases each predicate has and the types of nouns that can fill a case slot. Note that case frames offer not only the knowledge of the relationships between a predicate and its particular case slot, but also the knowledge of the relationships among a predicate and its multiple case slots. To obtain such knowl- edge, very large corpora seem to be necessary; how- ever it is still unknown how much corpora would be required to obtain good coverage. For examples, Kawahara and Kurohashi pro- posed a method for constructing wide-coverage case frames from large corpora (Kawahara and Kuro- hashi, 2006b), and a model for syntactic and case structure analysis of Japanese that based upon case frames (Kawahara and Kurohashi, 2006a). How- ever, they did not demonstrate whether the coverage of case frames was wide enough for these tasks and how dependent the performance of the model was on the corpus size for case frame construction. This paper aims to address these …

Content truncated to 500 words