La post-édition de traduction automatique en contexte d'apprentissage. Effets sur la qualité et défis pour l'enseignement de la traduction

For the past few years, the translation industry has been shaped by a technological boom,
largely driven by the advent of new translation technologies, such as neural machine
translation (NMT). Hence, professional practices are changing, and adapting translation
training programs has become a critical challenge both for translation training and for the
future of the profession. In such context, this thesis is devoted to the study of an
increasingly common practice in the language services industry: machine translation postediting (MTPE). Firstly, the main objective of our research is to investigate the effects of MTPE in an academic context on final text quality. Secondly, we seek to assess translation
students' knowledge and perceptions regarding MT and PE. Thirdly, we aim to contribute
to the debate on the introduction of MT and PE training courses in translation curricula. To
meet these goals, we conducted two controlled experiments involving translation students
to compare the outputs of two translation processes (English into French): human
translation (HT) and post-editing using both statistical and neural MT. To analyse the
empirical data gathered, we combined a qualitative approach (human quality assessment
and linguistic analysis of errors contained in students’ products) with a quantitative
approach (descriptive statistics, inferential statistics, and automatic linguistic analysis). Our findings show that PE of MT, whether statistical or neural, leads to products of an overall quality comparable to human-translated texts, or even of higher quality in the case of neural MTPE. Moreover, the fine statistical analysis mainly indicates that post-edited texts contain more calques than human-translated texts, and that NMT post-edited language contains fewer grammar and syntax errors compared to human-translated language. In addition, we note that PE quality depends on MT paradigm (statistical or neural), as well as on the neural MT engine used (Google Translate or DeepL). Our work also uncovers a leveling effect in neural MTPE on the quality of target texts, which demonstrates that the poorer the student’s translation skills, the more they will benefit from PE, and conversely, the higher their translation skills, the poorer the quality of their post-edited product. Further, part of our results suggests the existence of typical features that set post-edited language apart from human-translated language (i.e. Post-Editese). In the last part of this thesis, we substantiate our position in favour of MT Literacy training to educate informed, autonomous, and responsible users. Finally, based on our results and our pedagogical experience, we highlight the main issues and opportunities of MTPE and outline seven major challenges that need to be addressed in PE training.