基于《生死疲劳》语料库的大语言模型文学翻译质量评估

Translation Quality Assessment of Literary Translation by Large Language Models Based on the Corpus of Life and Death Are Wearing Me Out

摘要: 中华优秀传统文化“走出去”战略的深入实施，对翻译产出的质量标准提出了更高要求。与此同时，数智技术的快速发展为大语言模型(LLMs)介入文学翻译提供了技术支撑。基于此，本研究以莫言《生死疲劳》为例，构建“1对3”汉英平行语料库(源文本、葛浩文译本以及DeepSeek-V3.2与文心一言4.5译本)。研究运用BERTScore定量评估译文质量，并进一步引入豪斯(House)翻译质量评估模型从语场、语旨、语式三维度对大语言模型的翻译表现展开定性分析。研究结果表明，大语言模型译文的整体语义保真度接近人工译本，但是在低分例句中仍普遍存在文化信息缺失、情感表达弱化、文本风格趋同等共性问题。研究为数字化语境下中华文化的国际传播提供了实证支撑，并指出未来可以通过构建结构化提示词(Structured Prompting)等策略进一步优化大语言模型的翻译质量。

Abstract: The strategy of promoting the excellent traditional Chinese culture to go global has imposed higher requirements on translation quality, while the rapid advancement in digital intelligence technologies has provided technical support for applying large language models (LLM) into literary translation. On this basis, this study takes Mo Yan’s Life and Death Are Wearing Me Out as a case, constructing a “one-to-three” Chinese-English parallel corpus consisting of the source text, Howard Goldblatt’s translation, and translations generated by DeepSeek-V3.2 and ERNIE 4.5. The study employs BERTScore for quantitative assessment of translation quality, and further introduces House’s Translation Quality Assessment model to conduct a qualitative analysis of the LLMs’ translation performance from the three dimensions of field, tenor, and mode. The findings indicate that the overall semantic fidelity of large language model translations is close to that of human translations, yet low-scoring sentences reveal common problems such as loss of cultural information, weakening of emotional coloring, and stylistic convergence. This study provides empirical evidence for the digitally-driven international dissemination of Chinese culture, and suggests that the translation efficacy of LLMs can be further enhanced through structured prompt optimization in the future.