Evaluating Large Language Models Against Human in Stage Play Translation

Stage play translation is uniquely challenging because it demands more than literal languagetransfer; translators must preserve dramatic tone, cultural nuance, and metaphorical reso-nance to maintain the text’s performability. Capturing metaphors embedded in dialogueis especially critical for audience impact. This study evaluates the ability of large languagemodels (LLMs) to translate metaphor-rich stage play snippets. We constructed a dataset of50 passages reflecting the figurative density and stylistic expressiveness of theatrical scriptsand compared LLM-generated translations against human renderings. Translations were as-sessed for metaphor fidelity, readability, and dramatic suitability. Results show that LLMsconsistently produce natural and performable text comparable to human translations inreadability and dramatic tone, but they are less reliable in preserving metaphorical fidelity.These findings highlight both the promise of LLMs as tools for creative translation andtheir current limitations, suggesting future adaptations such as fine-tuning may be neededto enhance their handling of figurative language in the performing arts.