Cross-lingual Discourse Relation Analysis: A corpus study and a semi-supervised classification system

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 577–587, Dublin, Ireland, August 23-29 2014. Cross-lingual Discourse Relation Analysis: A corpus study and a semi-supervised classification system Junyi Jessy Li1, Marine Carpuat2 and Ani Nenkova1 1 University of Pennsylvania, Philadelphia, PA 19104, USA {ljunyi, nenkova}@seas.upenn.edu 2 National Research Council Canada, Ottawa, ON K1A 0R6, Canada marine.carpuat@nrc.gc.ca Abstract We present a cross-lingual discourse relation analysis based on a parallel corpus with discourse information available only for one language. First, we conduct a corpus study to explore dif- ferences in discourse organization between Chinese and English, including differences in infor- mation packaging, implicit/explicit discourse expression divergence, and discourse connective ambiguities. Second, we introduce a novel approach to learning to recognize discourse relations, using the parallel corpus instead of discourse annotation in the language of interest. Our result- ing semi-supervised system reaches state-of-art performance on the task of discourse relation detection, and outperforms a supervised system on discourse relation classification. 1 Introduction The analysis of the way spans of text semantically connect with each other to create a coherent text has a rich theoretical and empirical tradition (Mann and Thompson, 1988; Marcu, 1997; Di Eugenio et al., 1997; Allbritton and Moore, 1999; Schilder, 2002). Because of the difficulty in annotation, however, labelled datasets were rare and rather small. The release of the Penn Discourse Treebank (PDTB) (Prasad et al., 2008) brought about a new sense of maturity in discourse analysis, finally providing a high-quality large-scale resource for training discourse parsers for English. Based on the PDTB, a number of studies have provided insightful analysis of the use of discourse connectives in English news text and have developed methods for the identification of discourse relations and their arguments (Wellner and Pustejovsky, 2007; Pitler et al., 2008; Pitler and Nenkova, 2009; Pitler et al., 2009; Lin et al., 2009; Prasad et al., 2010; Park and Cardie, 2012; Lin et al., 2014). Some have applied the insights and classifiers to standard natural language processing tasks such as assessing text coherence and text quality (Pitler and Nenkova, 2008; Lin et al., 2011), detecting causal dependencies of events (Do et al., 2011), and machine translation (Meyer and Popescu-Belis, 2012). A resource like the PDTB is extremely valuable, and it would be desirable to have a similar resource in other languages as well. Following the release of the PDTB, smaller corpora annotated with discourse relations have been developed for Hindi (Oza et al., 2009), Turkish (Zeyrek and Webber, 2008), Arabic (Al-Saif and Markert, 2010), and the effort is on-going with Chinese (Zhou and Xue, 2012). On the other hand, for the vast majority of languages, such well-annotated resource for discourse re- lations is not available. In our work we carry the valuable annotations in the PDTB over to another language—Chinese—using parallel corpora. Projecting information available in one language onto an- other has been explored in areas such as part-of-speech tagging (Yarowsky et al., 2001; Das and Petrov, 2011), grammar induction (Hwa et al., 2005; …

Content truncated to 500 words