Training MT Model Using Structural SVM

This paper presents a training method of log-linear model for statistical machine translation based on structural support vector machine. This method is designed to directly optimize parameters with respect to translation quality. By adopting maximum-margin principle of SVM, the MT model can learn from training samples with generalization capability. Experiments are carried out on a hierarchical phrase-based MT system facing Chinese to English translation. Result shows that structural SVM training has the ability of re-ranking the k-best list of MT system according to automatic evaluation criteria BLEU, and it can enhance the average quality of MT system outputs.