An efficient approach to rule redundancy reduction in hierarchical phrase-based translation

Hierarchical phrase-based machine translation model is a popular syntax model that makes use of the expressive power of synchronous context-free grammars (SCFG) to address the reordering problem in statistical machine translation. The model, however, generally suffers from a great amount of redundancy in the extracted translation rules. In this paper, we re-introduce the concept of rift into the rule extraction procedure to force the rules with reordering power to concentrate on where reordering has actually happened. Our approach brings a dramatic reduction in the training time and the number of the rules, with only minor sacrifice in translation quality.