Prompt-induced cultural mediation and its limits: a micro-level analysis of LLM translation of Chinese tourism texts

Large Language Models (LLMs) such as ChatGPT demonstrate strong fluency in multilingual translation, yet their ability to convey culturally embedded meanings remains uneven. This study examines whether explicit prompt design can enhance LLM handling of culture-loaded elements in Chinese-to-English tourism translation. A corpus of 140 excerpts across seven cultural categories was translated under two prompting conditions, minimal and culturally explicit. Translation quality was assessed through expert cultural-adequacy ratings and BLEU scores, complemented by comparisons with expert thick-translation references. Results show that culturally oriented prompts improve performance in fact-based domains (e.g. customs, cuisine, architecture), whereas gains are limited for figurative, symbolic, or intertextually rich content (e.g. rhetoric, poetry). Qualitative analysis identifies three recurrent behaviors: cultural elaboration is activated mainly by explicit prompts; elaboration, once triggered, follows a uniform but weakly selective pattern; and the model excels at factual recall but struggles with inferential, rhetorical, and symbolic meaning. These findings indicate that prompting can reveal latent cultural knowledge but cannot replace the context-sensitive interpretive reasoning characteristic of human translators. LLMs may assist culturally enriched translation when properly guided, but human interpretive agency remains essential. Future work should examine target-reader reception, multi-turn prompting, and cross-genre extensions.