MTG-LIRA’s submission for the WMT 2026 Arabic–Asian Machine Translation Shared Task presents a data-centric Neural Machine Translation (NMT) pipeline to address low-resource Arabic–Hindi translation. Built using the OpenNMT-py framework with a 6-layer encoder-decoder Transformer architecture, the system relies on extensive corpus augmentation, combining official WMT data with ~4.05 million parallel sentence pairs from OPUS, alongside ~7 million synthetic Hindi–Arabic sentence pairs generated by translating the English side of the BPCC corpus (English-Hindi) into Arabic using Meta’s NLLB model. The data undergoes rigorous preprocessing, including Unicode normalization, quality filtering, deduplication, and alignment verification, before evaluating both Byte Pair Encoding (BPE) and SentencePiece tokenization (32,000 vocabulary size) across two training strategies. Results show that while BPE and SentencePiece achieve comparable performance, the impact of synthetic data depends heavily on the translation direction: augmenting with synthetic data yields substantial performance improvements for Hindi→Arabic (increasing BLEU from 7.2 to 9.5 and COMET from 0.8166 to 0.8262 using BPE), whereas it degrades translation quality for Arabic→Hindi (where the OPUS-only BPE model achieves a higher BLEU of 25.3 compared to 17.7 with synthetic data).

MTG-LIRA’s submission for the WMT 2026 Low-Resource Indic Language Translation Shared Task presents a unified, scalable multilingual Transformer architecture developed using OpenNMT-py to address severe data scarcity across several low-resource Indian languages. The framework partitions languages into family-based groupings—such as Indo-Aryan and Tibeto-Burman—to optimize cross-lingual knowledge sharing and vocabulary overlap using Byte Pair Encoding (BPE). Data preparation integrates official WMT datasets, the BPCC multilingual corpus, and ~5 million synthetic English–Mizo sentence pairs generated via Meta’s NLLB model, all cleaned using a rigorous preprocessing pipeline. Experimental results across task directions demonstrate that combining multilingual base models with targeted, language-specific adaptation strategies yields significant gains: fine-tuning on domain-specific data improved BLEU scores for English→Assamese (12.17 to 14.44) and English→Mizo (16.2 to 22.72); checkpoint averaging of the 5 best models enhanced stability for English↔Manipuri (Meitei Mayek); a weighted 3:1 checkpoint averaging strategy improved performance for English→Bodo; and zero-shot style direct transfer proved effective for Assamese→English and English→Manipuri (Bengali script) due to shared script features.