Attention based English to Indo-Aryan and Dravidian language translation using sparsely factored NMT

Abstract

Neural machine translation (NMT) is a sophisticated technique that employs a large, singular neural network to learn and execute automatic translation tasks. Unlike statistical machine translation systems, NMT handles the entire translation process in an end-to-end manner, removing the need for additional components. This approach has shown significant promise in translation quality and has become the prevalent method. In this study, we apply sparsely factored NMT to English and several Indo-Aryan (Hindi, Bengali) and Dravidian (Tamil, Malayalam) language pairs. Specifically, we develop the machine translation system using an attention-based mechanism. A significant problem with traditional transformers is the huge memory requirement. Therefore, a sparsely factored NMT (SFNMT) is used to reduce the memory requirement but also improves the training time, thereby, reducing the computing time. In this paper, take inspiration from Vaswani transformer and modify it to get the best results. The system’s performance was evaluated using the BLEU metric. The proposed model indtrl achieves a BLUE score of 32.13 (en→hi), 29.31 (en→be), 31.21 (en→ta), 21.12 (en→ml) and 32.67 (en→hi), 29.38 (en→be), 31.75 (en→ta), 21.17 (en→ml) without backtranslation and with backtranslation. To evaluate the performance of the system, we compared the results with those of existing systems. The developed system demonstrated a marginally higher BLEU score than both AnglaMT and Google translate.