Neural Machine Translation between Myanmar (Burmese) and Dawei (Tavoyan)

Abstract

This work explores the first evaluation of the quality of neural machine translation between Myanmar (Burmese) and Dawei (Tavoyan). We also developed Myanmar-Dawei parallel corpus (around 9K sentences) based on the Myanmar language of ASEAN MT corpus. We implemented two prominent neural machine translation systems: Recurrent Neural Network (RNN) and Transformer with syllable segmentation. We also investigated various hyper-parameters such as batch size, learning rate and cell types (GRU and LSTM). We proved that LSTM cell type with RNN architecture is the best for Dawei-Myanmar and Myanmar-Dawei neural machine translation. Myanmar to Dawei NMT achieved comparable results with PBSMT and HPBSMT. Moreover, Dawei to Myanmar RNN machine translation performance achieved higher BLEU scores than PBSMT (+1.06 BLEU) and HPBSMT (+1.37 BLEU) even with the limited parallel corpus.