Dependency reordering features for Japanese-English phrase-based translation

Abstract

Translating Japanese into English is very challenging because of the vast difference in word order between the two languages. For example, the main verb is always at the very end of a Japanese sentence, whereas it comes near the beginning of an English sentence. In this thesis, we develop a Japanese-to-English translation system capable of performing the long-distance reordering necessary to fluently translate Japanese into English. Our system uses novel feature functions, based on a dependency parse of the input Japanese sentence, which identify candidate translations that put dependency relationships into correct English order. For example, one feature identifies translations that put verbs before their objects. The weights for these feature functions are discriminatively trained, and so can be used for any language pair. In our Japanese-to-English system, they improve the BLEU score from 27.96 to 28.54, and we show clear improvements in subjective quality. We also experiment with a well-known technique of training the translation system on a Japanese training corpus that has been reordered into an English-like word order. Impressive results can be achieved by naively reordering each Japanese sentence into reverse order. Translating these reversed sentences with the dependency-parse-based feature functions gives further improvement. Finally, we evaluate our translation systems with human judgment, BLEU score, and METEOR score. We compare these metrics on corpus and sentence level and examine how well they capture improvements in translation word order.