Chinese Discourse Relation Recognition

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1442–1446, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Chinese Discourse Relation Recognition Hen-Hsen Huang Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan hhhuang@nlg.csie.ntu.edu.tw Hsin-Hsi Chen Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan hhchen@csie.ntu.edu.tw Abstract The challenging issues of discourse relation recognition in Chinese are addressed. Due to the lack of Chinese discourse corpora, we construct a moderate corpus with human- annotated discourse relations. Based on the corpus, a statistical classifier is proposed, and various features are explored in the experi- ments. The experimental results show that our method achieves an accuracy of 88.28% and an F-Score of 63.69% in four-class clas- sification and achieves an F-Score of 93.57% in the best case. 1 Introduction A discourse relation is the way that two succes- sive arguments logically connected. Recognizing discourse relations attracts much attetions in re- cent years due to many potential applications. In the annotation scheme of Penn Discourse Tree- bank 2.0 (PDTB-2.0), the first level of discourse relations includes four classes such as Temporal, Contingency, Comparison, and Expansion (Pra- sad et al., 2008). The first challenge of discourse relation recognition is the lack of corpus. To construct a discourse corpus has several difficulties. First, the definition of discourse is unclear and varied over different areas. Thus, finding the clear boundary of a discourse argument is a vexing problem by itself. Second, the relationship be- tween arguments is often difficult to decide and inherently subjective. Thus, the annotation and the evaluation are problematic and labor- intensive. In recent years, the study of discourse relation recognition is growing in the English domain rapidly. One of the reasons is the availability of English corpora with discourse annotations. The two most popular discourse corpora are the Rhe- torical Structure Theory Discourse Treebank (RSTDT) (Carlson et al., 2001) and PDTB-2.0. Both of them are based on the Wall Street Jour- nal corpus with human-annotated discourse in- formation. The PDTB-2.0 consists of 36,592 pairs of successive arguments and is tagged with three classes, including Implicit, Explicit, and AltLex, and with the relation types at three levels. Based on these corpora, a number of aspects on discourse relation are explored in these years. Compared to the English corpora, there is still no Chinese discourse corpus worldwide available. For this reason, the dataset is the first challenge encountered in the study of Chinese discourse relation recognition. To address this issue in this work, we construct a moderate Chinese discourse corpus as a starting point. A supervised statistical classifier is trained and tested on this data set to deal with the problem. Various features are ex- tracted from the corpus and evaluated in the ex- periments. The rest of this paper is organized as follows. First, we review the related work in Section 2. In Section 3, the Chinese discourse relation prob- lem is illustrated. Our corpus and the details on the annotation work …

Content truncated to 500 words