Discriminative Reordering Model for Machine Translation

We have built a discriminative reordering model for the phrase-based machine translation system Moses, which is developed at the University of Edinburgh. The model is a maximum entropy classifier which incorporates a variety of feature functions to predict phrase orientation for machine translation. Two kinds of features reported in literature, namely lexical features and dependency path feature have been tested in the discriminative model. We have also proposed and tested a novel feature named dependency orientation and modified the dependency path feature with lexicalization. Two baseline models are used in evaluation, namely, distance-based model without lexicalized reordering, and the lexicalized reordering model. We are able to achieve significant BLEU gains over the distance based model by up to 0.95 absolute points, and BLEU gains over the lexicalized reordering model by up to 0.74 absolute points. Discriminative reordering model is a very generic framework and is open to many more features for further improvement.