Decoder-based Discriminative Training of Phrase Segmentation for Statistical Machine Translation

摘要

In this paper, we propose a new method of training phrase segmentation model for phrase-based statistical machine translation(SMT). We define a good segmentation as the segmentation producing a good translation. According to this definition, we propose a method that can discriminate between a good segmentation and a bad segmentation based on the translation quality. The proposed approach constructs the phrase labeled data by using the SMT decoder, so that the phrase segmentations supporting good translations can be acquired. Furthermore, our iterative training algorithm of the segmentation model can gradually improve the performance of the SMT decoder. Experimental results show that the proposed method is effective in improving the translation quality of the phrase-based SMT system.