Answers are grounded in this paper and cite their sources.
This study quantitatively compares the translation performance of two neural machine translation models, Google Translate and DeepL, alongside the AI chatbot ChatGPT, in translating food safety regulations from the Codex Alimentarius Commission. These regulations contain domain-specific terminology related to nutrition, food processing, non-sentence-list items, and polysemous words. The evaluation uses BLEU, chrF++, TER, and COMET metrics. While Google Translate achieved the highest BLEU score (0.197), DeepL, which added Arabic support in 2024, outperformed Google Translate and ChatGPT across all other metrics, leading to its selection for manual analysis using the Multidimensional Quality Metrics error typology. DeepL’s output revealed four types of errors: fluency, accuracy, style, and terminology, but it also demonstrated a level of eloquence and readability comparable to human translations. The findings suggest that DeepL should be considered an alternative to Google Translate and can serve as a new baseline for Arabic machine translation research. Moreover, despite its popularity, ChatGPT lags in specialized translation tasks. Finally, professional translators should consider integrating machine translation systems into their workflows to enhance efficiency in general domains and specialized fields.
No takes yet. Share an insight, caveat, or question.
Altakarli et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: