Rule Based Katakana to Myanmar Transliteration for Post-editing Machine Translation

Phrase based statistical machine translation (PBSMT) is a current state-of-the-art approach to machine translation, however its outputs often contain various types of errors such as lexical errors, and syntax errors (Koehn et al., 2003)(Bojar et al., 2013)(Bojar, 2011b). Incorporating deep linguistic knowledge directly into PBSMT is not easy and rarely leads to improvements in translation performance (Bojar, 2011a). One of the possible solution is to make automatic corrections on translated output in a post-editing process. This paper presents a rule based post-editing scheme for fixing translation errors based on out of vocabulary (OOV) Katakana words produced by Japanese to Myanmar PBSMT. Our experiments indicate that applying rule based Katakana to Myanmar transliteration leads to substantial improvements of translation quality both in terms of BLEU scores and OOV coverage.