Combining Natural and Artificial Examples to Improve Implicit Discourse Relation Identification

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1694–1705, Dublin, Ireland, August 23-29 2014. Combining Natural and Artificial Examples to Improve Implicit Discourse Relation Identification Chloé Braud ALPAGE, Univ Paris Diderot & INRIA Paris-Rocquencourt 75013 Paris - France chloe.braud@inria.fr Pascal Denis MAGNET, INRIA Lille Nord-Europe 59650 Villeneuve d’Ascq - France pascal.denis@inria.fr Abstract This paper presents the first experiments on identifying implicit discourse relations (i.e., relations lacking an overt discourse connective) in French. Given the little amount of annotated data for this task, our system resorts to additional data automatically labeled using unambiguous connec- tives, a method introduced by (Marcu and Echihabi, 2002). We first show that a system trained solely on these artificial data does not generalize well to natural implicit examples, thus echoing the conclusion made by (Sporleder and Lascarides, 2008) for English. We then explain these ini- tial results by analyzing the different types of distribution difference between natural and artificial implicit data. This finally leads us to propose a number of very simple methods, all inspired from work on domain adaptation, for combining the two types of data. Through various experiments on the French ANNODIS corpus, we show that our best system achieves an accuracy of 41.7%, corresponding to a 4.4% significant gain over a system solely trained on manually labeled data. 1 Introduction An important bottleneck for automatic discourse understanding is the proper identification of implicit relations between discourse units. What makes these relations difficult is that they lack strong surface cues like a discourse marker. This point is illustrated in the French examples (1) and (2).1 In (1), the connective mais (but) triggers a relation of contrast, whereas in (2), there is no explicit connective to signal the explanation relation, and the relation has to be inferred through other ways (in this case, a causal relation between having injured players and loosing). (1) La hulotte est un rapace nocturne, mais elle peut vivre le jour. The tawny owl is a nocturnal bird of prey, but it can live in the daytime. (2) L’équipe a perdu lamentablement hier. Elle avait trop de blessés. The team lost miserably yesterday. It had too many injured players. Implicit relations are very widespread in naturally-occurring data. Thus, they make up between 39.5% and 54% of the annotated examples in the Penn Discourse TreeBank (PDTB) (Prasad et al., 2008), depending on the relation types used.2 A quick look at other discourse corpora suggests that the problem is as pervasive (if not more) in other languages. The French ANNODIS corpus does not annotate the distinction between explicit and implicit relations, but a projection of a French connective lexicon on the data gives a proportion of 47.4 to 71% of implicit relations, depending on the set of relations.3 For the German discourse corpus of (Gastel et al., 2011), (Versley, 2013) report 65% of implicit relations. In this paper, we tackle the problem of automatically identifying implicit discourse relations in French. To date, the large majority of studies …

Content truncated to 500 words