Discriminative source side dependency tree reordering model

In this research a novel discriminative reordering model for statistical machine translation is proposed. Source dependency tree is used to define the orientation classes of the reordering model. We use maximum entropy principle to train the model. In addition to the common features used in the discriminative reordering models, two new and effective features are introduced. They are phrase number and orientation memory features. The proposed model is integrated to the decoding phase of the translation. The performance of this method and effect of each individual feature are evaluated on two Persian-English corpora. We observe a relative 5% improvement in terms of BLEU score.