英语专八试题人机翻译句法复杂度对比分析研究

A Comparative Study on Syntactic Complexity between Human and ChatGPT-5 Translations of TEM-8 Translation Task

摘要: 随着生成式人工智能技术在自然语言处理领域的深度渗透，ChatGPT等大语言模型在翻译任务中的应用备受关注，但现有评估多聚焦准确性与流畅性，对反映翻译专业性的句法复杂度关注不足。本研究以1996~2025年(不含2020年)英语专业八级(专八)汉译英试题为对象，构建ChatGPT-5译文与专八官方参考答案的平行语料库，采用BFSU系列句法分析工具，从基础句法指标、复杂句法结构指标、句法多样性与规范性三个维度，结合独立样本t检验开展量化对比。结果显示：两类语料在基础句法规模与句法多样性上无显著差异；人类翻译语料在从属分句、补语从句等复杂结构的使用密度，以及句法规范性、场景适配稳定性上显著优于ChatGPT-5；ChatGPT-5存在复杂句法结构使用保守、异常值占比偏高的问题。本研究首次系统揭示了二者在句法复杂度上的深层差异，丰富了AI翻译质量评估的理论体系，为AI翻译模型优化、英语专业教学及专八翻译备考提供了实证依据与实践指导。

Abstract: With the deep integration of generative artificial intelligence technologies into the field of natural language processing, large language models such as ChatGPT have attracted considerable attention in translation tasks. However, existing evaluations mainly focus on accuracy and fluency, while relatively neglecting syntactic complexity, which reflects translation professionalism. This study takes the Chinese-English translation tasks of the Test for English Majors Band 8 (TEM-8) from 1996 to 2025 (excluding 2020) as research data. A parallel corpus consisting of ChatGPT-5 translations and the official reference translations was constructed. The BFSU series of syntactic analysis tools were employed to conduct quantitative comparisons from three dimensions: basic syntactic indices, complex syntactic structure indices, and syntactic diversity and well-formedness, combined with independent-samples t-tests. The results show that there is no significant difference between the two corpora in terms of basic syntactic scale and syntactic diversity. However, the human translation corpus demonstrates significantly higher density in the use of complex structures such as subordinate clauses and complement clauses, as well as stronger syntactic well-formedness and greater stability in contextual adaptability. ChatGPT-5 shows a conservative tendency in the use of complex syntactic structures and exhibits a relatively higher proportion of outliers. This study systematically reveals the underlying differences between the two in syntactic complexity, enriches the theoretical framework of AI translation quality evaluation, and provides empirical evidence and practical guidance for AI translation model optimization, English-major teaching, and TEM-8 translation preparation.