High-Fidelity English–Urdu And English–Arabic Neural Translation Through Metric-Guided Marianmt Optimization

Abstract

This research addresses the critical underrepresentation of Urdu in machine translation studies and extends the work to include Arabic, aiming to bridge the gap in low-resource language translation. The methodology spans from dataset preprocessing through model fine-tuning to the deployment of a user-friendly graphical interface, ensuring both technical robustness and practical usability. Leveraging a fine-tuned MarianMT-based architecture with attention mechanisms and optimized preprocessing techniques, the system ensures accurate contextual mapping between English and the target languages. A parallel corpus of English Urdu and English Arabic sentence pairs was employed for training and evaluation, incorporating tokenization, sequence padding, and maximum length constraints to enhance translation quality. Recognizing the inherent challenges in low-resource language processing, the approach integrates comprehensive evaluation using standard metrics accuracy, precision, recall, F1-score, and BLEU score alongside visualization techniques to improve transparency and interpretability. Experimental evaluation reveals outstanding performance, achieving 94.00% accuracy, 95.00% recall, 93.78% precision, and a 94.00% F1-score, surpassing several established baselines for multilingual low-resource translation tasks. Beyond its technical merit, the proposed system demonstrates significant potential in applications related to education, cross-cultural communication, and information accessibility. While this study focuses on English Urdu and English Arabic translation, the methodology is inherently adaptable to other language pairs, paving the way for scalable and inclusive multilingual translation systems.