Exploiting Dependency-based Pre-ordering for English-Myanmar Statistical Machine Translation

Word reordering is one of the challenging problems for machine translation when translating between different word order languages, such as English and Myanmar. The grammatical forms of these two languages are totally dissimilar so it is very difficult to translate from English to Myanmar with proper reordering. In this paper, we employed a pre-ordering approach to learn reordering rules automatically that aims to adjust the word order of the source sentence to that of the target sentence prior to translation. The rules are learned with source-side dependency parse trees and bilingual word alignments to obtain a monotonic parallel corpus. Experiments on English-Myanmar translation show that the pre-ordering approach obtains statistical improvement by up to 1.8 BLEU points on Asian Language Tree-bank (ALT) data, compared to baseline state-of-the-art phrase-based statistical machine translation (PBSMT).