汉语隐性块状结构的人机法译显化差异研究

A Comparative Study of Explicitation in Human and Machine Translation into French of Implicit Block Structures in Chinese

摘要: 具备“强空间缩合特征”的汉语块状结构往往潜隐着丰富的深层语义，其跨语际的“识隐显译”构成了中国文化对外传播的核心难点。本研究以跨语际“时空性差异”理论与认知语言学为理论框架，聚焦大语言模型与官方人工译者在认知加工机制上的差异。通过构建汉法平行语料库并设计层级递进的提示词(Prompt 1-3)，本研究对汉语隐性块状结构的显化翻译特征展开了定量与定性相结合的系统考察。研究发现：1) 在缺乏干预的基础指令下，机器翻译受制于源语表层结构，呈现出显著的“零显译”倾向；2) 随着提示词层级的优化，大模型对源语潜隐语义的识别与显化效能显著改善，印证了提示词工程对激活机器认知推理的正向干预作用；3) 大语言模型在输出稳定性与语义连贯性上仍具较高随机性，生成文本易出现语义流失、语用偏离乃至文化误读。特别是在宏观施事主体的“完全显译”层面，大模型仍面临显著的认知瓶颈，尚难企及人工翻译的语用等效精度。本研究在“人机协同”视域下，初步构建了跨语际显化翻译的分类与多维测评框架，探明了优化大模型处理汉语隐性语义的干预路径，旨在为高校数智翻译教学范式转型及中国文化走出去提供实证支撑。

Abstract: Chinese block-like structures, characterized by “strong spatial condensation,” often conceal rich deep semantics. The cross-linguistic “recognition of the implicit and explicitation in translation” constitutes a core challenge in the international dissemination of Chinese culture. Framed within the cross-linguistic “spatio-temporal divergence” theory and cognitive linguistics, this study focuses on the differences in cognitive processing mechanisms between Large Language Models (LLMs) and official human translators. By constructing a Chinese-French parallel corpus and designing progressively layered prompts (Prompts 1-3), this research systematically investigates the translational explicitation features of Chinese implicit block-like structures through a combination of quantitative and qualitative analyses. The findings reveal that: 1) Under basic zero-shot prompts, machine translation is heavily constrained by the surface structure of the source language, exhibiting a significant tendency toward “zero explicitation”; 2) As the prompt levels are optimized, the LLM’s efficacy in recognizing and explicitating the implicit semantics of the source language improves significantly, validating the positive intervention effect of prompt engineering in activating machine cognitive reasoning; 3) LLMs still demonstrate considerable randomness in output stability and semantic coherence, with generated texts prone to semantic loss, pragmatic deviation, and even cultural misreading. Particularly in the “full explicitation” of macro-level agents, LLMs continue to face severe cognitive bottlenecks and fall short of the pragmatic equivalence precision achieved by human translators. From the perspective of “human-machine collaboration,” this study preliminarily constructs a classification and multidimensional evaluation framework for cross-linguistic translation explicitation. It explores intervention pathways to optimize LLMs in processing Chinese implicit semantics, aiming to provide empirical support for the paradigm shift in digital and intelligent translation teaching in higher education, as well as the empirical support for the global dissemination of Chinese culture.