MTG-LIRA: Scaling Neural Machine Translation for Low-Resource Indian Languages through Multilingual Learning and Synthetic Data

Abstract

Low-resource Indian languages continue to face significant challenges in neural machine translation due to the scarcity of high-quality parallel corpora and the diversity of scripts and linguistic characteristics. This paper presents MTG-LIRA’s submission to the WMT 2026 Low-Resource Indic Language Translation Shared Task, focusing on scalable multilingual translation for several low-resource Indian language pairs. Our approach is built upon a unified multilingual Transformer architecture that enables parameter sharing and knowledge transfer across related languages. To further improve translation quality, we investigate multiple adaptation strategies, including multilingual transfer learning, synthetic data augmentation, task-specific fine-tuning, checkpoint averaging, and weighted model averaging, selecting the most effective approach for each language pair. We submitted systems for low resource language pairs like English↔Assamese, English→Mizo, English→Manipuri (Bengali), English↔Manipuri (Meitei Mayek), and English→Bodo with both primary and contrastive submissions where applicable.. In addition, we conducted exploratory experiments on several other low-resource Indian languages to assess the generalizability of our multilingual framework. The submitted systems demonstrate that combining a shared multilingual foundation with language-specific adaptation strategies provides an effective and scalable solution for neural machine translation in low-resource Indian languages.