Improving an English-Arabic Transformer-Based Machine Translation Model

Abstract

Arabic is widely recognized as one of the most challenging languages for translation due to its rich morphology, complex syntax, and context-sensitive structures. Despite its global importance, Arabic has received significantly less attention in machine translation research compared to European languages, highlighting the pressing need for further investigation into high-quality Arabic machine translation systems. This paper proposes an enhanced model for English–Arabic machine translation tailored to the news domain, based on the Transformer architecture, which currently underpins most state-of-the-art machine translation. The model is augmented by incorporating external lexical alignment data at each decoding step. This integration is designed to improve the system’s handling of polysemy, contextual ambiguity, and missing word detection, thereby enhancing translation accuracy, cohesion, and semantic fidelity. The proposed model was evaluated based on BLEU scores and F measure. Results showed that the proposed model outperformed, achieving an accuracy of 87% based on BLEU score, and an 84% score on the F-measure test, surpassing the Google translate by 6%, on the same tested data. These findings confirm that incorporating lexical knowledge into the Transformer framework significantly improves English–Arabic translation quality. The model not only provides more translations that are accurate but also demonstrates improved cohesion and contextual understanding, particularly in morphologically rich and domain-specific texts. This research underscores the value of domain-aware and linguistically informed machine translation approaches, paving the way for more effective Arabic-language translation systems in practical applications.