This paper describes a submission to the WMT 2026 Low-Resource Arabic-Asian Language Translation shared task, covering three directions: Urdu→Arabic, Arabic→Urdu, and Arabic→English. For each direction we submit a primary system (official WMT data only) and a contrastive system trained on an augmented corpus. For the low resource Urdu-Arabic pair, we build a silver standard corpus by mining CC-100 Urdu monolingual text, applying domain adaptation to filter for relevance against WMT DevTest anchors using LEALLA sentence embeddings, and translating the retained sentences into Arabic — adding 27,172 domain matched pairs to the 20,299 official pairs. For Arabic-English, we augment the WMT data with News Commentary to expand the training pool roughly fivefold. All systems are built through transfer learning, fine-tuning the pretrained multilingual NLLB-200 model across all three directions. Our contrastive systems outperform our own primary systems in every direction (BLEU gains of 1.6–3.2 points), and on the official leaderboard our contrastive systems rank 1st in Arabic→Urdu, 3rd in Urdu→Arabic, and 5th in Arabic→English. Our results show that domain adaptation, rather than raw data volume, is the more effective lever for improving low resource NMT.

We have explored the effect of in domain knowledge during parallel sentence filtering from in domain corpora. Models built with sentences mined from in domain corpora without domain knowledge performed poorly, whereas model performance improved by more than 2.3 BLEU points on average with further domain centric filtering. We have used Large Language Models for selecting similar and domain aligned sentences. Our experiments show the importance of inclusion of domain knowledge in sentence selection methodologies even if the initial comparable corpora are in domain.