Effective Use of Discontinuous Phrases for Hierarchical Phrase-based Translation

Hierarchical phrase-based (HPB) models have shown strong capability in generalization and reordering. However, they are heavily dependent on continuous phrases and are difficult for modeling natural linguistic discontinuities directly. In this paper, we propose a novel approach for integrating discontinuous phrases into the Chinese-to-English HPB system. We focus on the extraction method of discontinuous phrases which retrieves various linguistic information missed in the HPB model, such as set phrases and long-distance reordering of adverbials, etc. After being transformed into the similar form to HPB rules, the translation rules with discontinuities can be seamlessly integrated into the CKY decoder. Experimental results show that the proposed approach for incorporating the linguistic discontinuities achieves statistically significant improvements over the traditional HPB system.