Evaluating Arabic-English Neural Machine Translation: Challenges Across Different Text Types

Résumé

Artificial intelligence and natural language processing have gained widespread recognition recently, particularly as neural machine translation (NMT) has become indispensable for translation service providers. However, despite its unprecedented technological advancements, machine translation (MT) engines continue to struggle with Arabic-English translation due to linguistic complexities, structural differences, and the limited availability of high-quality training data. These challenges are particularly evident when translating nuanced content that requires a deep understanding of context and cultural sensitivity. This paper evaluates the performance of three MT systems—Reverso, Systran, and Microsoft Azure—by analyzing their translations of three distinct text types: general, technical, and journalistic. It uses the TAUS Dynamic Quality Framework error typology to assess key aspects, including accuracy, fluency, terminology, and style. The analysis is both qualitative and quantitative, offering a comprehensive view of each system's strengths and limitations. The findings indicate that while the three MT engines generate comprehensible translations, they consistently struggle with recurring issues, such as domain-specific terminology, idiomatic expressions, and stylistic coherence. The study underscores the necessity of post-editing and highlights how MT can assist human translators in improving productivity while also emphasizing the irreplaceable value of human expertise. This research brings attention to the need to refine MT engines through domain-specific and higher-quality linguistic training resources for Arabic-English translation.