A Multilingual Neural Machine Translation Model for Low Resource North Eastern Languages

Abstract Multilingual neural machine translation (MNMT) has emerged as a powerful approach to facilitate cross-lingual communications. However, developing MNMT systems for low resource languages in linguistically diverse regions remains a significant challenge due to scarcity of parallel data. This study introduces an innovative MNMT model designed for the linguistically diverse North-East Indian region, leveraging the transformer architecture. The proposed system employs shared encoders and language- specific decoders for English to Indic language translation while using language-specific decoders for and a shared decoder for the reverse direction. To address data scarcity, we incorporate back-translation techniques to augment the parallel corpus. Experiments on six North-East Indian languages (Assamese, Bodo, Khasi, Manipuri, Mizo and Nepali) demonstrated good results across different evaluation metrics like BLEU, ROUGE,METEOR and TER.