Research on prompt engineering for translation with large language models (LLMs) completely overlooks interpreting or fails to distinguish it from translation. Unlike translators, interpreters rely heavily on paraphrasing, summarizing, and concise expression—skills at which LLMs excel. This master's thesis explores whether LLMs can generate translations that match the quality of human interpretation when guided by an augmented prompt designed to imitate interpreting strategies. The method employs chain-of-thought prompting, guiding the model through a three-step process: 1) paraphrasing the text in clear and concise language, 2) transferring it into the target language, and 3) refining it using natural collocations while ensuring coherence. This approach mirrors the core interpreting technique of deverbalization and reformulation. The prompt was tested with Claude 3.5 Sonnet and GPT-4o on speeches from EU institutions. The translations were evaluated in terms of content accuracy, terminology and vocabulary usage, syntax, style and grammar. The results suggest that while the customized prompt had only a minor effect on semantic accuracy, it significantly improved cohesion, syntax, and idiomatic quality of the translations—to the extent that the final outputs were comparable to human interpretation. This study is the first to explore how imitating interpreters’ cognitive process influences translation quality.
|
