LLM-Generated Synthetic Data for Low-Resource Urdu–English Neural Machine Translation

Urdu is spoken by more than 230 million people, yet it remains critically under-resourced in machine translation research due to the scarcity of high-quality parallel corpora relative to high-resource language pairs. We investigate whether large language model (LLM)-generated synthetic data can improve a small fine-tuned translation model in this low-resource setting, without requiring new human-annotated parallel text. Starting from an mT5-small baseline fine-tuned on 19,793 Urdu–English sentence pairs from OPUS-100, we construct an augmented training set using two complementary synthetic data generation techniques: (i) back-translation of monolingual English sentences into Urdu using NLLB-200-distilled-600M, and (ii) paraphrase-based diversification of existing English source sentences using Qwen2.5-1.5B-Instruct, keeping the original Urdu target fixed. Both models are trained under identical architectures and hyperparameters to isolate the effect of the added data. The augmented model improves over the baseline on both BLEU (7.37 → 8.21) and chrF (25.02 → 26.10). A length-stratified error analysis shows that the two models diverge far more often on longer source sentences, suggesting that the synthetic data primarily improves generalization to harder, longer inputs rather than producing uniform gains across the test set.