Improving English-to-Myanmar Statistical Machine Translation by using Recurrent Neural Network Language Model and Hierarchical Reordering

Abstract

In this paper, a comparison between neural based network that adopts Recurrent Neural Network (RNN) language model and statistical based one with N-gram language model is conducted for English-to- Myanmar phrase-based statistical machine translation (PBSMT). In this comparison, lexicalized reordering models such as word-based, phrase-based and hierarchical orientation models are used as an additional reordering model to investigate the overall performance of PBSMT. The perplexity value evaluation of both language models showed that the use of RNN obtains a more excellent result. According to the obtained BLEU and RIBES scores and additional human visual inspection, the English-to- Myanmar PBSMT with RNN language model and hierarchical reordering model is the best one in terms of improving adequacy and fluency.