Impact of Automatic Post-Editing and Prompting Strategies on the Linguistic Features of English-to-French Editorial Translations

This article consists in a comparison of the raw machine translation (MT), from English to French, and automatic post-editing (APE) of editorials, using three systems for MT (DeepL, Google Translate and GPT-4o) and three prompts with different levels of instruction specificity for APE. According to linguistic metrics, lexical and syntactic diversity increase across the board in APE as compared to MT, with the least constrained prompt generally leading to higher gains than the most constrained one. Meanwhile, quality estimation scores follow an opposite trend: less constrained prompts achieve lower COMETKiwi scores. A qualitative comparison of some machine-translated and automatically post-edited excerpts shows that APE can correct MT errors, reduce fluency errors or calques, and lead to more natural translations overall, but may also introduce new errors.