Adaptive Morphology-Aware Contrastive Translation for Urdu and Arabic Translations

Abstract

 

Neural machine translation (NMT) has made significant progress through the introduction of sequence-to-sequence learning and attention mechanisms. However, translating morphologically complex languages ​​such as Arabic and Urdu remains challenging due to complex word formation, inflectional variations, and data scarcity. This article presents a multilingual, morphology-aware NMT system to improve translation quality for resource-poor and morphologically complex language pairs. The proposed system, Adaptive Morphology-Aware Contrastive Translation (AMACT) extends a conventional Seq2Seq approach with attention mechanisms by explicitly modelling morphological features so it allows for better precise alignment along with preserving semantics during the translation process. To work on this model, a continuous experimental framework has been developed that allows to integrates data preprocessing, vocabulary building, model training, evaluation, and real-time translation into a unified graphical user interface. Experiments have been conducted using the popular parallel Arabic-English and also Urdu-English corpora along with standard BLEU evaluation metric that corresponds and helps in performance evaluation. The evaluation shows better results basically states that proposed morphology-aware model significantly outperforms the basic approach at all n-gram levels. For Arabic-English translation, the proposed model achieves a BLEU-1 score of 0.9721 and a BLEU-4 score of 0.9655, compared to 0.9455 and 0.8681 specifically for the basic model. In the same way for Urdu to English translation, the proposed method shows significant improvements, thereby achieving a BLEU-4 score of 0.0824, as compared to earlier 0.0028 for the basic model. These results also confirm that integrating morphological knowledge into NMT architectures increases translation accuracy, particularly for languages ​​with complex morphology. This work underlines the effectiveness of morphology-based modelling and creates a scalable basis for future resource-efficient NMT model.