Traduction humaine et postédition : contrôle qualité en contexte académique

Our study is a follow-up to a 2018 comparative experimental research on human translation (HT) and machine translation (MT) post-editing (PE) (English-French) that was conducted with third-year bachelor students in translation (Schumacher 2020b). In this article we provide the results of a 2nd controlled experiment conducted in 2021 with master’s students. The objectives are to compare both HT and MTPE products (DeepL and Google Translate) and to contribute to the debate on adapting translator training programs. Quantitative analysis of this 2nd experiment shows that post-edited texts were judged of better overall quality than HT. The human quality assessment also shows very high success rates in PE (especially with DeepL) and confirmed the existence of a levelling effect in PE on the target text quality: PE–particularly with Google Translate–gives rise to an inverse relationship between a student’s performance in HT and the quality of his/her post-edited product. In other words, students with poor-quality human-translated productions are the ones who benefit most from PE and vice versa. Finally, it appears the quality of a post-edited text depends on neural MT system since students performed better with DeepL.