@InProceedings{koirala-bhetwal-choudhury:2026:wmt,
  author    = {Koirala, Nabin  and  Bhetwal, Rhythm  and  Choudhury, Nurul Amin},
  title     = {NITM AI Lab at WMT 2026 Shared Task: Low-Resource Arabic-Asian Language Translation},
  booktitle      = {Proceedings of the Eleventh Conference on Machine Translation},
  month          = {October},
  year           = {2026},
  address        = {Budapest, Hungary},
  publisher      = {Association for Computational Linguistics},
  pages     = {2331--2335},
  abstract  = {This paper describes the NITM AI Lab submissions to the WMT26 Low-Resource Arabic-Asian Translation shared task. We participate in two translation directions: English-to-Arabic (Sub-Task 1A) and Arabic-to-Hindi (Sub-Task 2B). To address the data scarcity inherent to these language pairs, we compare three neural architectures: a domain-specialized bilingual model (OPUS-MT-TC-Big), a distilled massively multilingual model (NLLB-200Distilled-600M), and a generalist multilingual sequence-to-sequence transformer (mBART50-Many-to-Many). Our systems contrast full-parameter fine-tuning with parameter-efficient fine-tuning (PEFT) via Low-Rank Adaptation (LoRA), and we measure the effect of rule-based orthographic and numerical normalization together with targeted data augmentation from OPUS. On the official blind test set, the domain-specialized OPUS-MT model is the strongest system for English-to-Arabic, while LoRA-tuned NLLB-200 performs best for Arabic-to-Hindi, with OPUS augmentation giving a small additional gain in the latter direction.},
  url       = {https://aclanthology.org/2026.wmt-1.174}
}