English-Manipuri Machine Translation: An empirical study of different Supervised and Unsupervised Methods

English-Manipuri MT is a challenging task because of the limited availability of parallel data and language divergences between English and Manipuri. This work investigates the English-Manipuri MT with different supervised and unsupervised methods of SMT and NMT. In the supervised approach, we have experimented on Phrase-based SMT and Transformer-based NMT. Furthermore, unsupervised methods are also examined, namely Unsupervised NMT(Artetxe et al.), Unsupervised SMT(Lample et al.), XLM(Cross-lingual Language Model Pretraining), and MASS(Masked Sequence to Sequence Pretraining). Common practices for improving low resource Machine Translation like Back-translation and Self-Training are also tried. Detailed error analysis and the challenges faced in the English-Manipuri MT system are also presented in this paper. The paper also explains how a new corpus of 74K parallel English-Manipuri sentences is constructed. This is the largest English-Manipuri parallel corpus reported till date.