ENLUS: Bridging Language Divide with English-Mizo Machine Translation Corpus

In this work, we present ENLUS, a parallel and monolingual corpus of the English-Mizo language pair developed for machine translation. ENLUS monolingual corpora have been collated from publicly available news and media websites in Mizo. The parallel data has been collected from various online and offline sources. Machine translation has recently made significant progress in high-resource language scenarios. The quantity of linguistic data mostly determines how well the translation models work. In a multilingual culture like India, most languages are under-resourced with scant to no data. Therefore, there is an evident need for reliable, freely accessible resources to carry out the translation. Mizo is an Indian language that is under-explored due to limited resources. We outline the procedures used to create and validate both corpora. We train popular machine translation techniques, including unsupervised and supervised models, on this corpus. We also compare the translation results to those of Google Translate, which is a commercial translation service. For the purpose of comparing all the models, we use the publicly available evaluation benchmark, FLORES-200, which recently included the Mizo language. Our empirical findings reveal that the multilingual NMT system outperforms existing bilingual techniques and commercial translation tools.