From translational data to contrastive knowledge

Textual aligning consists in pairing segments (e.g. sentences or phrases) that are translational equivalents across corpora of translations. An interesting application of textual aligning is the automatic extraction of bilingual lexicons. As it has been pointed out during previous evaluation campaigns, such as Arcade, lexical aligning remains problematic. In order to solve problems of consistency linked with the concept of translational compositionality, a redefinition of lexical aligning task is proposed, introducing the concept of lexical correspondence. Simple techniques dedicated to lexical correspondences extraction are then evaluated. Thus, it appears that adapted statistical filters allow to extract very accurately significant regularities that are relevant at the contrastive level. More generally, these methods prove to be adapted not only for bilingual lexicons extraction: they could be used to study a wide range of contrastive phenomena on empirical basis.