Extremely Low-resource Multilingual Neural Machine Translation for Indic Mizo Language

Machine translation requires a vast amount of parallel data in order to generate high-quality translations. Since many Indian languages lack sufficient resources, enhancing translation performance for these language pairs can have a significant impact. This study aims to address the issue of low-resource neural machine translation between Mizo and English by utilizing other Indian languages in a multilingual framework. We explore the use of multilingual techniques for 13 pairs of Indic languages in both many-to-one and one-to-many setups, proposing a method for Multilingual Neural Machine Translation to enhance the translation quality of the low-resource Mizo language. We assess the effectiveness of ensemble decoding and transliteration into the Roman script through qualitative and quantitative approaches. The empirical findings demonstrate that incorporating transliteration and utilizing ensemble decoding with checkpoint ensembles leads to improved translation quality.