Discriminant Models for Word Alignment Modèles discriminants pour l'alignement mot à mot

Word alignment aims to link each word of a translated sentence to its related words in the source sentence. Nowadays, Giza++ is the most used word alignment system. This toolkit implements the generative IBM models. Despite its popularity, several limitations remain. We thus propose to address this task using discriminative models (Maximum Entropy and Conditional Random Fields) which can easily make use of additional features. These models are evaluated in terms of Alignment Error Rate (AER) using two language pairs (French/english and Arabic English). Our results show that discriminant models are well suited for this task and that they can outperform IBM models.