Translation model adaptation for an Arabic/French news translation system by lightly-supervised training

Most of the existing, easily available paral-lel texts to train a statistical machine trans-lation system are from international organi-zations that use a particular jargon. In this paper, we consider the automatic adaptation of such a translation model to the news do-main. The initial system was trained on more than 200M words of UN bitexts. We then ex-plore large amounts of in-domain monolingual texts to modify the probability distribution of the phrase-table and to learn new task-specific phrase-pairs. This procedure achieved an im-provement of 3.5 points BLEU on the test set in an Arabic/French statistical machine trans-lation system. This result compares favorably with other large state-of-the-art systems for this language pair.