Annotation précise du français en sémantique de rôles par projection cross-linguistique

This paper considers the task of the automatic induction of role-semantic annotations for new languages with high precision. To this end we test the generalisability of the language-independent, projection-based annotation framework introduced by Pado and Lapata (2005, 2006) by (a) applying it to a new, more distant, language pair (English-French), and (b), using automatic, and thus noisy, input annotation. We show that even under these conditions, high-quality role annotations for French can be obtained that rival existing results for German. We conclude that the framework has considerable potential in reducing the manual effort involved in creating role-semantic resources for a wider range of languages. Mots-cles : Multilingue, FrameNet, Annotation semantique automatique, Semantique lexicale, Projection d'annotation de roles, roles semantiques