Apprentissage discriminant des modèles continus de traduction
Do, Quoc Khanh, Alexandre Allauzen, François Yvon · HAL (Le Centre pour la Communication Scientifique Directe) · 2015
This paper proposes a new discriminative framework to train translation models based on neural network. This framework relies on the definition of a new objective function that allows us to introduce the evaluation metric in the learning process as well as to consider how the model interacts with the translation system. Moreover, this approach is compared with the state of the art estimation methods, such as the maximum likelihood criterion and the noise contrastive estimation. Experiments are carried out on the English to French translation task of TED Talks . The results show the efficiency of the proposed approach, whereas the initialization has a strong impact. We show that with a tailored initialization scheme significant improvements can be obtained in terms of BLEU scores.