Abstract This study examines the workload, translation quality, and user perception associated with Chinese–English (C–E) post-editing (PE) performed by student translators using three tools: DeepL, ChatGPT-4o, and DeepSeek-V3. Thirty Master of Translation and Interpreting (MTI) students from a comprehensive university in China participated in the study. Source texts of varying types, but similar complexity level, served as the materials for raw machine translation outputs. By analyzing variables such as task duration, keyboard events and pause length, translation quality assessment data of 120 post-edited files, and questionnaire responses, the study yielded the following findings: (1) The use of neural machine translation (NMT) or large language models (LLMs) did not significantly impact overall PE workload, though students assisted by LLMs required fewer keyboard events when editing popular science texts; (2) LLM-assisted PE demonstrated higher quality than those edited with DeepL, as evidenced by fewer error counts and lower error scores; (3) Participants perceived ChatGPT-4o to be the most effective tool for error correcting ability and task efficiency. DeepSeek-V3 was rated highest in terms of user experience, while DeepL was regarded as the most reliable in terms of information accuracy. By comparing students’ PE workload, translation quality and user perception across different scenarios, this study offers new insights into human–AI collaboration in PE workflows and its implications for translation pedagogy in the AI era.
Shao et al. (Fri,) studied this question.