基于语料库的《呼兰河传》人机英译词汇风格比较

A Corpus-Based Comparison of Lexical Style in Human and GPT English Translations of Tales of Hulan River

摘要: 大语言模型进入文学翻译实践后，人机译文比较已从准确性评价拓展至风格研究，但现有研究多依赖节选或个别译例，难以判断局部差异是否构成稳定倾向。本文以《呼兰河传》全部正文为源文本，建立中文原文、葛浩文英译本与GPT英译文三方语料库，采用词汇多样性、近似词汇密度和词长指标进行全书及分章比较，并结合文化负载词作语境分析。结果显示，葛译词汇多样性较高，GPT译文近似词汇密度较高，两者词长构成接近。文化负载词案例表明，葛译会根据具体语境改变译法，GPT较多保留源语词汇构成和文化形式。研究认为，人机译文的词汇风格各自表现为相对稳定但并非固定的用词倾向，其具体表现会随文本内容和文化词项变化。全书指标、分章复核与译例细读相结合，可为人机文学翻译风格研究及译者主体性讨论提供语言证据，也有助于思考文学翻译中的人机协作模式。

Abstract: With large language models increasingly used in literary translation, human–machine translation comparison has expanded beyond accuracy assessment to stylistic analysis. However, existing studies often rely on excerpts or isolated examples, making it difficult to determine whether local differences constitute stable tendencies. Using the complete text of Tales of Hulan River as its source, this study constructs a tripartite corpus comprising the Chinese original, Howard Goldblatt’s English translation, and a GPT-generated translation. Lexical diversity, approximate lexical density, and word length are compared at both whole-book and chapter levels, supplemented by contextual analysis of culture-specific items. The results show that Goldblatt’s translation has greater lexical diversity, whereas the GPT translation has higher approximate lexical density; the two translations display similar word-length profiles. The culture-specific items examined show that Goldblatt adjusts his lexical choices across contexts, whereas GPT more often retains source-language lexical components and cultural forms. The findings indicate that human and machine lexical styles represent relatively stable but not fixed patterns of lexical choice whose specific manifestations vary across textual contexts and culture-specific items. By combining whole-book measures, chapter-level validation, and close reading of translation examples, this study provides linguistic evidence for research on human-machine stylistic differences in literary translation and discussions of translator agency, while also informing approaches to human-AI collaboration in literary translation.