@inproceedings{j-etal-2025-anvita,
    title = "{ANVITA} : A Multi-pronged Approach for Enhancing Machine Translation of Extremely Low-Resource {I}ndian Languages",
    author = "J, Sivabhavani  and
      Kankanwadi, Daneshwari  and
      Mishra, Abhinav  and
      Paul, Biswajit",
    editor = "Haddow, Barry  and
      Kocmi, Tom  and
      Koehn, Philipp  and
      Monz, Christof",
    booktitle = "Proceedings of the Tenth Conference on Machine Translation",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-emnlp/2025.wmt-1.101/",
    pages = "1240--1247",
    ISBN = "979-8-89176-341-8",
    abstract = "India has a rich diverse linguistic landscape including 22 official languages and 122 major languages. Most of these 122 languages fall into low, extremely low resource categories and pose significant challenges in building robust machine translation system. This paper presents ANVITA Indic LR machine translation system submitted to WMT 2025 shared task on Low-Resource Indic Language Translation covering three extremely low-resource Indian languages Nyshi, Khasi, and Kokborok. A transfer learning based strategy is adopted and selected suitable public pretrained models (NLLB, ByT5), considering aspects such as language, script, tokenization and fine-tuned with the organizer provided dataset. Further, to tackle low-resource language menace better, the pretrained models are enriched with new vocabulary for improved representation of these three languages and selectively augmented data with related-language corpora, supplied by the organizer. The contrastive submissions however made use of supplementary corpora sourced from the web, generated synthetically, and drawn from proprietary data. On the WMT 2025 official test set, ANVITA achieved BLEU score of 2.41-11.59 with 2.2K to 60K corpora and 6.99-19.43 BLUE scores with augmented corpora. Overall ANVITA ranked first for {\{}Nyishi, Kokborok{\}}{\ensuremath{\leftrightarrow}}English and second for Khasi{\ensuremath{\leftrightarrow}}English across evaluation metrics including BLUE, METEOR, ROUGE-L, chrF and TER."
}