We propose three novel reordering models for statistical machine translation. These reordering models use dependency tree to improve the translation quality. All reordering models are utilized as features in a log linear framework and therefore guide the decoder to make better decisions about reordering. These reordering models are tested on two English-Persian parallel corpora with different statistics and domains. The BLEU score is improved by 2.5 on the first corpus and by 1.2 on the other.

Keywords: statistical machine translation, reordering model, dependency tree, discriminative reordering model, discriminative decoder, long range reordering, maximum entropy, distortion model

Full-Text [PDF 2202 kb]   (1422 Downloads)