Interpretation of Chinese Discourse Connectives for Explicit Discourse Relation Recognition

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 632–643, Dublin, Ireland, August 23-29 2014. Interpretation of Chinese Discourse Connectives for Explicit Discourse Relation Recognition Hen-Hsen Huang, Tai-Wei Chang, Huan-Yuan Chen, and Hsin-Hsi Chen Department of Computer Science and Information Engineering National Taiwan University No. 1, Sec. 4, Roosevelt Road, Taipei, 10617 Taiwan {hhhuang, twchang}@nlg.csie.ntu.edu.tw; {b00902057, hhchen}@ntu.edu.tw Abstract This paper addresses the specific features of Chinese discourse connectives, including types (word-pair and single-word), linking directions (forward and backward linking), positions and ambiguous degrees, and discusses how they affect the discourse relation recognition. A semi- supervised learning method is proposed to learn the probability distributions of discourse func- tions of connectives from a small labeled dataset and a big unlabeled dataset. The statistics learned from the dataset demonstrates some interesting linguistic phenomena such as connec- tive synonyms sharing similar distributions, multiple discourse functions of connectives, and couple-linking elements providing strong clues for discourse relation resolution. 1 Introduction Discourse relation labeling determines how two discourse units cohere to each other. A discourse unit may be a clause, a sentence, or a group of sentences. The labeled relation has many potential applica- tions. Coherence is considered as a metric to evaluate the essay writing by essay scorer (Lin et al., 2011). Discourse relations are used to order sentences in an event in a summarization system (Der- czynski and Gaizauskas, 2013). Sentiment transition of two clausal arguments is identified based on their discourse relation in sentiment analysis (Hutchinson, 2004; Zhou et al., 2011; Wang et al., 2012; Huang et al., 2013). The pioneer research of discourse has been established by Hobbs (1985), Polanyi (1988), Hovy and Maier (1992), and Asher and Lascarides (1995). Various discourse relation types have been defined in the frameworks such as Sanders et al. (1992), Hovy and Maier (1992), RST-DT (Carlson et al., 2002), Wolf and Gibson (2005), and PDTB (Prasad et al., 2008). Temporal, Contingency, Comparison, and Expansion, the four classes on the top level of PDTB sense hierarchy, are common used in the dis- course relation labeling tasks. When two arguments are temporally related, they form a Temporal rela- tion. The Contingency relation talks about the situation that the event in one argument casually affects the event in the other argument. Comparison is used to show the difference between two arguments. The last one relation, Expansion, is the most common. An Expansion relation either expands the in- formation for one argument in the other one or continues the narrative flow. In the recent years, discourse relation recognition has been studied for different languages (Afan- tenos et al., 2012, Cartoni et al., 2013). In explicit English discourse relation labeling tasks, the accu- racy of the approach using just the connectives is already quite high, 93.67%, and incorporating the syntactic features raises performance to 94.15% (Pitler and Nenkova, 2009). In our previous work, we investigate Chinese intra-sentential relation detection and show an accuracy of 81.63% and an F-score of 71.11% in the two-way classification (Contingency …

Content truncated to 500 words