A Transformer-Based Multilingual Neural Machine Translation Framework with Shared–Language-Specific Decoding for Low-Resource Northeast Indian Languages

Abstract

Multilingual Neural Machine Translation MNMT recently emerges as a pow erful paradigm for tackling the task of translating among multiple language pairs under one model framework Even so it can be difficult to construct highquality MNMT systems for lowresource languages particularly in linguis tically diverse regions because of the scarcity of parallel bilingual corpora In this paper we introduce a framework for MNMT based on transformer for the ecosystem of NorthEast Indian language The shared encoder with language independent decoders is used in English to Indic translation directions and vice versa utilises the language independent encoder with language specific decoder Moreover to mitigate the lack of parallel training data backtranslation is also employed to provide another source of synthetic parallel sentence pairs which enhances the training corpus of our model The system was evaluated using stan dard translation metrics including BLEU Bilingual Evaluation Understudy ROUGE METEOR and TER Translation Edit Rate Results demonstrate promising performance with BLEU scores of 029 for English–Assamese and 026 for English–Nepali with 95 confidence intervals confirming the robustness of 1 these gains over other language pairs These findings highlight the potential of our MNMT model in advancing translation quality for lowresource Northeast Indian languages and provide a foundation for further improvements using larger datasets and pretrained multilingual models.