Max margin learning for statistical machine translation: Toward improvement of machine translation accuracy

Abstract

Minimum error rate training (MERT) has been a widely used learning method for statistical machine translation to estimate the feature function weights of a linear model. MERT has an advantage to incorpolate an automatic translation evaluation metrics as BLEU scores to its objective function. Weight vector can directly be optimized with Line search algorithm using error surface on a given set of candidate translations. It efficiently searches the best parameter resulting the highest BLEU scores. In this paper, we presented a new training algorithm for statisitcal machine translation, inspired by MERT and Structural Support Vector Machines. We performed MERT optimization by maximizing the margin between the oracle and incorrect translations under the L2-norm prior. Our experimental results on Japanese-English speech translation task showed that BLEU scores obtained by our proposed method were much better than those obtained by MERT. We achieved the best improvement of BLEU about +3.0 over standard MERT.