Non-Autoregressive and Iterative Refinement Non-Autoregressive Transformer with Additional Graph Layer for Urdu-to-English Machine Translation

Neural Machine Translation (NMT) has achieved state-of-the-art performance using autoregressive Transformer. However, the trade-off between translation quality and decoding speed remains a central challenge. This problem is further intensified for low-resource languages such as Urdu, where limited parallel data and complex linguistic structure result in poor translation quality and reduced model effectiveness. To overcome this problem Non-Autoregressive Transformer (NAT) and Non-Autoregressive Transformer with Iterative Refinement (iNAT) are proposed. This paper investigates NAT and iNAT for low resource language Urdu-English machine translation. We analyze translation accuracy and their effectiveness for low-resource language pairs. Experimental results show that both NAT and iNAT outperform the baseline Autoregressive Transformer (ART) in terms of translation accuracy. However, the overall translation quality remains limited due to the complex linguistic structure and resource scarcity of the Urdu-English pair. To address this limitation, we further refine these architectures by introducing an additional graph-based learning layer that captures latent relationships among tokens through attention-driven graph representations. The proposed graph-enhanced model learns contextual dependencies by dynamically constructing graphs over attention representations. Experimental evaluation demonstrates that the proposed architecture improves the translation performance of both NAT and iNAT models. On the Urdu-English translation task, the proposed model achieves a 3.0% higher BLEU score than the baseline models. The findings highlight the significance of incorporating graph-based layer to enhance translation quality, making NMT systems more effective for low-resource and linguistically complex languages such as Urdu.