Word-Level Reordering Model for Phrase-Based SMT

The complicated alignment and small translation unit make the word based approaches extremely complex and thereby hard to achieve promising performance. The employment of phrase largely addresses the alignment problem. On the other hand, the phrase-based SMT (PBSMT) models suffer more from data sparse problem and behave less flexible than word-based model because of the larger translation unit-- phrase. Therefore we conduct our research on enhancing phrase based SMT with word-level reordering model (based on source dependency tree). Experimental results on the NIST Chinese-English machine translation data show that our reordering models significantly improve the baseline, a state-of-the-art reordering model, which is widely used in phrase-based SMT system.