Inducing Discourse Connectives from Parallel Texts

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 610–619, Dublin, Ireland, August 23-29 2014. Inducing Discourse Connectives from Parallel Texts Majid Laali and Leila Kosseim Department of Computer Science and Software Engineering, Concordia University, Montreal, Quebec, Canada {m laali, kosseim}@cse.concordia.ca Abstract Discourse connectives (e.g. however, because) are terms that explicitly express discourse rela- tions in a coherent text. While a list of discourse connectives is useful for both theoretical and empirical research on discourse relations, few languages currently possess such a resource. In this article, we propose a new method that exploits parallel corpora and collocation extraction techniques to automatically induce discourse connectives. Our approach is based on identifying candidates and ranking them using Log-Likelihood Ratio. Then, it relies on several filters to fil- ter the list of candidates, namely: Word-Alignment, POS patterns, and Syntax. Our experiment to induce French discourse connectives from an English-French parallel text shows that Syntac- tic filter achieves a much higher MAP value (0.39) than the other filters, when compared with LEXCONN resource. 1 Introduction Discourse relations are often categorized as being implicit or explicit depending on how they are marked linguistically (Prasad et al., 2008). Implicit relations between two text spans are inferred by the reader even if they are not explicitly connected through lexical cues. On the other hand, explicit relations are explicitly identified with syntactically well-defined terms, so called discourse markers or discourse connectives (DCs). A list of DCs is a valuable resource to help the automatic detection of discourse relations in a text. Discourse parsers (e.g. (Lin et al., 2010)) often use DCs as a powerful distinguishing feature to tag discourse relations (Pitler and Nenkova, 2009). A list of DCs is also instrumental in generating annotated training data which, in turn, is critical for training data-driven parsers (Prasad et al., 2010). In this article, we propose an automatic method to induce a list of DCs for one language from a parallel corpus. We present an experiment in inducing a French DC list from an English-French parallel text. Our approach is based on the hypothesis that discourse relations are retained during the translation process. Therefore, if a reliable discourse tagger exists in a language, we can produce a corpus with discourse annotation labels in any language that has a parallel text with that language. Fortunately, according to (Versley, 2011), in English, the discourse usage of DCs can be automatically identified and labeled with their relation with 84% precision; a result that is close to the reported inter-annotator agreement. Moreover, with the advancement of statistical machine translation, today English parallel corpora for several languages are publicly available. Although we can expect little variability in the usage of discourse relations in parallel texts, this is not the case for DCs. In other words, translated texts may not always reproduce DCs of the source texts. Since discourse relations can be conveyed either explicitly with a DC or implicitly, a translator may choose to remove explicit DCs in the source …

Content truncated to 500 words