@InProceedings{karim-EtAl:2026:wmt,
  author    = {Karim, Misbahul  and  Ahmed, Faruk  and  Islam, Mishbahul Al  and  Laskar, Sahinur Rahman  and  Laskar, Rabul Hussain},
  title     = {NERDS-NITS at WMT2026: Pivot-Based and Direct Neural Machine Translation for Arabic–Asian Low-Resource Language Pairs},
  booktitle      = {Proceedings of the Eleventh Conference on Machine Translation},
  month          = {October},
  year           = {2026},
  address        = {Budapest, Hungary},
  publisher      = {Association for Computational Linguistics},
  pages     = {2325--2330},
  abstract  = {This paper presents NERDS-NITS's sub mission to the WMT 2026 Low-Resource Arabic–Asian Language Translation Shared Task, which covers the following language pairs: Arabic↔English, Arabic↔Hindi, Arabic↔Bengali, Arabic↔Urdu, and Arabic↔Indonesian. For the three Indic language pairs, we propose a pivot-based framework that combines a fine-tuned Arabic↔English NLLB-200 model with pretrained IndicTrans2 checkpoints, motivated by the scarcity of direct Arabic–Indic parallel data. For Arabic↔Indonesian, where Indic Trans2 offers no coverage, we instead fine-tune NLLB-200-1.3B directly using QLoRA under memory-constrained conditions. Our Arabic↔English backbone, built on fine-tuned NLLB-200-600M model, forms the pivot leg for all Indic-language systems. Across all evaluated pairs, fine-tuning yields consistent improvements over zero-shot baselines, and the proposed pivot architecture further outperforms direct multilingual fine-tuning for Arabic–Indic pairs, with gains of up to 6.68 BLEU and 0.02 COMET},
  url       = {https://aclanthology.org/2026.wmt-1.173}
}