Automatically identifying implicit discourse relations using annotated data and raw corpora (Identification automatique des relations discursives « implicites » à partir de données annotées et de corpus bruts) [in French]

Automatically identifying implicit discourse relations using annotated data and raw corpora This paper presents a system for identifying « implicit » discourse relations (that is, relations that are not marked by a discourse connective). Given the little amount of available annotated data for this task, our system also resorts to additional automatically labeled data wherein unambiguous connectives have been suppressed and used as relation labels, a method introduced by (Marcu et Echihabi, 2002). As shown by (Sporleder et Lascarides, 2008) for English, this approach doesn’t generalize well to implicit relations as annotated by humans. We show that the same conclusion applies to French due to important distribution differences between the two types of data. In consequence, we propose various simple methods, all inspired from work on domain adaptation, with the aim of better combining annotated data and artificial data. We evaluate these methods through various experiments carried out on the ANNODIS corpus : our best system reaches a labeling accuracy of 45.6%, corresponding to a 5.9% significant gain over a system solely trained on manually labeled data. MOTS-CLÉS : analyse du discours, relations implicites, apprentissage automatique.