Multilingual Indian Language Neural Machine Translation System Using mT5 Transformer

Abstract

This paper demonstrates the development and evaluation of a multilingual neural machine translation system for Indian languages based on the mT5 transformer, successfully utilized to develop multiple state-of-the-art NLP models. We used the modified Asian Language Treebank multilingual dataset to train the system for developing a Machine Translation model capable of translating text in English, Hindi and Bengali amongst each other. Our system was able to achieve acceptable BLEU scores of over 20 in five of the six language pairs, with the English to Bengali system achieving a maximum BLEU score of 49.87 and the Bengali to English system achieving an average BLEU score of 42.43.