ANVITA : Machine Translation System for Low-resource Indian Languages - Khasi, Nagamese and Tagin

Abstract

In this paper, we present ANVITA machine translation system submitted to WMT 2026 shared task on Low-Resource Indic Language Translation, where the team participated for six translation directions: { Khasi, Nagamese, and Tagin } ↔ English. Two core ideas shaped the design of ANVITA system. Firstly, transfer learning by symmetric fine-tuning of T5-base model and enhanced cross-lingual transfer from related languages (Assamese to Nagamese) by mitigating script mismatch through transliteration. Secondly, to alleviate data scarcity, primary submissions are trained on datasets augmented with related-language corpora and paraphrased pairs (English sentences), utilizing only the organizers provided datasets. For the contrastive submissions, training corpora are further expanded by compiling synthetic parallel data and re-lated language data followed by corpora distillation with suitable selection strategies. AN-VITA submissions are evaluated on the official test sets using multiple metrics which include BLEU, METEOR, TER, CHRF++, BERT score, and COMET. In the primary submissions, ANVITA achieved first rank for English → Tagin translation. For the contrastive submissions, ANVITA secured first rank for English → Khasi and Khasi → English with BLEU scores of 32.89 and 26.48 respectively. Furthermore, the contrastive sub-missions for Nagamese ↔ English and Tagin → English ranked second, underscoring efficacy of the data augmentation strategies applied.