Learning Better Classification-Based Reordering Model for Phrase-Based Translation

Reordering is of a challenging issue in phrase-based statistical machine translation systems. This paper proposed three techniques to optimize classification-based reordering models for phrase-based translation under the bracket transduction grammar framework. First, a forced decoding technique is adopted to learn reordering samples for maximum entropy model training. Secondly, additional features are learned from the context of two consecutive phrases to enhance the prediction ability of the reordering classifier. Thirdly, the reordering model score is integrated as two feature functions (STRAIGHT and INVERTED) into the log-linear model to improve its discriminative ability. Experimental result demonstrates significant improvements over the baseline in two translation tasks such as Chinese to English and Chinese to Japanese translation.