Phrase-based statistical machine translation for a low-density language pair

Abstract

We present a phrase-based statistical machine translation (SMT) system for Bangla to English that incorporates a novel transliteration module, and a specialized component for handling prepositions and Bangla compound words. We evaluate our components through their impact on the BLEU score for the phrase-based SMT system. According to the experimental results, the transliteration component has the most significant impact on the BLEU score. We also provide a new test set with multiple references between Bangla and English for MT evaluation purposes. Finally we propose a new manual evaluation approach for the MT community and evaluate our components using the new manual evaluation approach. © 2010 Springer-Verlag Berlin Heidelberg.