@InProceedings{sharma-EtAl:2026:wmt,
  author    = {Sharma, Pushkar  and  Ahtasam, Mo  and  Singh, Kshetrimayum Boynao  and  Kumar, Deepak  and  Ekbal, Asif},
  title     = {NLP-IIT Patna at WMT 2026: Corpus-Driven Adaptation of Multilingual MT Models for Low-Resource Arabic–Asian Machine Translation},
  booktitle      = {Proceedings of the Eleventh Conference on Machine Translation},
  month          = {October},
  year           = {2026},
  address        = {Budapest, Hungary},
  publisher      = {Association for Computational Linguistics},
  pages     = {2336--2344},
  abstract  = {We investigate low-resource Arabic-Asian machine translation across six Arabic-centric translation directions: Arabic↔English, Arabic↔Hindi, and Arabic↔Urdu. Fine-tuned MADLAD-400 ranks first in four of the six directions on the official primary leaderboard and places within the top three in the remaining two. We fine-tune three pretrained multilingual models, MADLAD-400, NLLB-200, and GemmaX2, and compare them with their zero-shot counterparts using BLEU, chrF2++, COMET-22, and TER. Across all language pairs, supervised fine-tuning consistently outperforms zero-shot inference. Among the evaluated models, fine-tuned MADLAD-400 achieves the strongest overall performance, NLLB-200 delivers competitive results, and GemmaX2-28-9B exhibits the largest relative improvement over its zero-shot baseline. These findings demonstrate the effectiveness of supervised adaptation for low-resource Arabic-Asian machine translation and highlight MADLAD-400 as the most robust model across the evaluated translation directions.},
  url       = {https://aclanthology.org/2026.wmt-1.175}
}

