Simple but Effective Approaches to Improving Tree-to-tree Model

Tree-to-tree translation model is widely studied in statistical machine translation (SMT) and is believed to be much potential to achieve promising translation quality. However, the existing models still suffer from the unsatisfactory performance due to the limitations both in rule extraction and decoding procedure. According to our analysis and experiments, we have found that tree-to-tree model is severely hampered by several rigid syntactic constraints: the both-side subtree constraint in rule extraction, the node constraint and the exact matching constraint in decoding. In this paper we propose two simple but effective approaches to overcome the constraints: utilizing fuzzy matching and category translating to integrate bilingual phrases and using head-out binarization to binarize the bilingual parsing trees. Our experiments show that the proposed approaches can significantly improve the performance of tree-to-tree system and outperform the state-of-the-art phrase-based system Moses.