Investigation of Data Augmentation Techniques for Assamese-English Language Pair Machine Translation

Machine translation can deliver impressive results when a substantial parallel corpus is accessible. However, in a country like India, characterized by its linguistic diversity and numerous languages with distinct origins and scripts, most languages face a scarcity of resources, making it challenging to create translation models of high quality. This work investigates a neural machine translation system with data augmentation techniques to boost the translation quality for an extremely resource-constrained language pair, i.e., English–Assamese. We experiment with back-translation, tagged back-translation, iterative back-translation, and iterative tagged back-translation. A qualitative and quantitative analysis performed on the various data augmentation techniques is performed. Furthermore, human evaluation is carried out to evaluate the adequacy and fluency of the translation. Empirical results show that data augmentation via iterative back-translation methods and tagged approach enhances translation performance in extremely low-resource settings.