New Advances in Reordering for Statistical Machine Translation

Abstract:

Phrase-based statistical machine translation delivers good performance for machine translation. Nevertheless, the difference in word order between different languages poses a major challenge to this approach, especially for language pairs with significant differences in word order. This thesis tackles the reordering problem by exploiting dependency parse trees in the phrase-based statistical machine translation approach.

We propose a novel approach to detect translation ordering of two words and apply sparse dependency swap features in translation decoding to encourage good translation output word order, which gives a significant improvement in Chinese-to-English translation. We then design a neural dependency-based reordering model applied within phrase-based translation decoding, resulting in a further improvement on Chinese-to-English translation. Experiments on other language pairs further demonstrate the strength of our proposed approach. We also explore system combination with the recently proposed end-to-end neural machine translation, which shows the competitiveness of our proposed approach