Neural Machine Translation Using a Pivot Approach: Dogri to English

The progress of Neural Machine Translation for Indic languages is gradually achieving a noteworthy result in terms of translation accuracy. However, its advancement is impeded by the lack of rich parallel corpora, a limitation not shared by non-Indic languages. As a result, most languages of India are still categorized as low-resource or zero-resource languages. Various state-of-the-art architectures have been developed to address these challenges, with the predominant approach being the pivot-based technique. This technique, rooted in transfer learning, utilizes an auxiliary language as a pivot to assist in translation. In this study, a pivotbased Neural Machine Translation (NMT) architecture is designed for translating Dogri to English, with Hindi as the pivot. Dogri, being a low-resource language, shares linguistic similarities with Hindi. Furthermore, Hindi-to-English state-of-the-art models are accessible, ensuring high translation accuracy. Through the implementation of this pivot-based translation model, a BLEU score of 20.97, SacreBLEU of 21.54 and chrF of 50.01 is attained.