Neural machine translation for limited resources English-Nyishi pair

Neural machine translation handles sequential data over the variable length of input and output sentences and accomplishes a state-of-the-art method for the task of machine translation. Although the neural machine translation shows good performance in both low and high-resource language pairs translation, it requires adequate parallel training data. In low-resource language sets, the preparation of the corpus is strenuous and time-consuming. Automatic translation systems like Google and Bing cover under-resourced Indian languages, but lack the support of the Nyishi language. It is due to the lack of a suitable dataset. In this work, we have contributed a parallel corpus of low-resource language pairs, English-Nyishi, and reported comparative experiments on the baseline neural machine translation systems. The results are evaluated for English to Nyishi and vice-versa via well-known automatic evaluation metrics and manual evaluation.