Graph-based lexicalized reordering models for statistical machine translation

Lexicalized reordering models are very important components of phrase-based translation systems. By examining the reordering relationships between adjacent phrases, conventional methods learn these models from the word aligned bilingual corpus, while ignoring the effect of the number of adjacent bilingual phrases. In this paper, we propose a method to take the number of adjacent phrases into account for better estimation of reordering models. Instead of just checking whether there is one phrase adjacent to a given phrase, our method firstly uses a compact structure named reordering graph to represent all phrase segmentations of a parallel sentence, then the effect of the adjacent phrase number can be quantified in a forward-backward fashion, and finally incorporated into the estimation of reordering models. Experimental results on the NIST Chinese-English and WMT French-Spanish data sets show that our approach significantly outperforms the baseline method.