A Lexicon-based Investigation of Research Issues in Japanese Factuality Analysis

International Joint Conference on Natural Language Processing, pages 587–595, Nagoya, Japan, 14-18 October 2013. A Lexicon-based Investigation of Research Issues in Japanese Factuality Analysis Kazuya Narita∗, Junta Mizuno† and Kentaro Inui∗ ∗Graduate School of Information Sciences, Tohoku University, Japan †National Institute of Information and Communications Technology (NICT), Japan {narita, junta-m, inui}@ecei.tohoku.ac.jp Abstract Event factuality is information about whether events mentioned in natural lan- guage correspond to either actual events that have occurred in the real world or events that are of uncertain interpreta- tion. Factuality analysis is useful for infor- mation extraction and textual entailment recognition, among others, but sufficient performance has not yet been achieved by the machine learning-based approach. It is now important to take a closer look at the linguistics phenomena involved in factual- ity analysis and identify the technical re- search issues more precisely. In this paper, we discuss issues regarding lexical knowl- edge through error analysis of a Japanese factuality analyzer based on lexical knowl- edge and compositionality. 1 Introduction Event factuality is information about whether events mentioned in natural language correspond to either actual events that have occurred in the real world or events that are of uncertain interpre- tation. (1) a. 彼はさきほど部屋を出た。 kare-wa sakihodo heya-wo de-ta. (He left the room a little while ago.) b. もう遅いから、彼は先に帰っ::::::::: たのだろう。 mou osoi-kara, kare-wa saki-ni kaet:::::::::: -ta-no-daro-u. (It’s late now, so he::: may:::: have gone home.) c. 問題が発生するのを ::::: 防いだ。 mondai-ga hassei-suru-no-wo :::::: fusei-da. (We ::::::: prevented the occurrence of the problem.) For example, we can interpret that the event “de” (leave) in (1a) is factual in the real world, the event “kaet” (go home) in (1b) is possibly fac- tual because of the modal auxiliary “-ta-no-daro- u” (may have -ed), and the event “hassei-suru” (occurrence) in (1c) is counterfactual because of the implicative predicate “fusei-da” (prevented). Factuality analysis is useful for a broad range of NLP applications such as information extrac- tion, question answering, and textual entailment recognition. Prior work on factuality analysis has made considerable efforts for designing and cre- ating corpora manually annotated with factuality- related information (Saur´ı and Pustejovsky, 2009; Matsuyoshi et al., 2010; Tanaka et al., 2013, etc.) and several empirical studies on those re- sources are reported revealing the difficulties of the task (Inui et al., 2008; Matsuyoshi et al., 2010; Morante and Blanco, 2012; Saur´ı and Pustejovsky, 2012). For Japanese, Matsuyoshi et al. (2010) re- port that their factuality classes are highly skewed and the minority classes are very difficult for their machine learning-based models to precisely iden- tify. The minority classes include uncertain state- ments as in example (1b) and counterfactual state- ments as in (1c). Such “marked” statements are far less frequent than unmarked statements (i.e. cer- tain factual statements) and thus are not as easy to collect as unmarked statements. While the label distribution is reported to be less skewed in English (Szarvas et al., 2008), still uncertain and counterfactual statements constitute minority classes. In addition, uncertain and counterfactual statements exhibit a very broad variety of linguis- …

Content truncated to 500 words