大语言模型对中国形象类块状语英译能力评测与优化研究——跨文化传播视角下的人机协同翻译路径探索

A Study on Evaluating and Optimizing the Capability of Large Language Models in Translating Chunks with Chinese Images into English—Exploring Human-Machine Collaborative Translation Pathways from an Intercultural Communication Perspective

摘要: 大语言模型技术的快速发展为中国形象的跨文化传播提供了新的技术路径。本研究通过构建三维量化评价体系，自建三大领域共450组四字格块状语语料库，对三款国产主流大语言模型进行系统评测与对比分析。研究整合COMET自动化评分与专家人工校验机制，并运用Few-shot学习和Chain-of-Thought等提示工程技术开展优化实验。研究发现：(1) 三款模型在语义准确性与表达流畅性上表现相对均衡，但文化传递力普遍偏低，呈现出“形似而神异”的翻译特征；(2) 四字格块状语因其高度凝练的认知格式塔结构，对模型的文化解码能力提出更高要求，模型普遍存在“去语境化”倾向；(3) 基于思维链的显性提示策略能够有效激活模型的隐性文化知识，使其从“概率匹配”转向“逻辑推理”翻译；(4) 人机协同介入机制在改善中国形象对外话语传播中具有不可替代性，提示工程本质上构成了一种认知补偿机制，为模型搭建了外部的认知支架。本研究为大语言模型在中国形象对外传播中的应用提供了科学的诊断依据与技术优化方案。

Abstract: The rapid development of large language model (LLM) technology provides new technological pathways for the cross-cultural communication of China’s national image. This study systematically evaluates and comparatively analyzes three mainstream domestically-produced LLMs by constructing a three-dimensional quantitative evaluation framework and building a self-compiled corpus of 450 sets of four-character chunky expression across three domains. The research integrates COMET automated scoring with expert manual verification mechanisms, and conducts optimization experiments by employing prompt engineering techniques such as Few-shot learning and Chain-of-Thought. The findings reveal: (1) The three models demonstrate relatively balanced performance in semantic accuracy and expressive fluency, but universally exhibit low cultural conveyance capacity, manifesting translation characteristics of “formal similarity with semantic divergence”; (2) Four-character chunks, due to their highly condensed cognitive Gestalt structure, impose higher demands on models’ cultural decoding capabilities, while LLM models universally displaying a tendency toward “decontextualization”; (3) Explicit prompting strategies, based on chain-of-thought, can effectively activate models’ implicit cultural knowledge, enabling a shift from “probabilistic matching” to “logical reasoning” translation; (4) Human-machine collaborative intervention mechanisms possess irreplaceable value in improving the external discourse communication of China’s national image, because prompt engineering essentially constitutes a cognitive compensation mechanism that constructs external cognitive scaffolding for models. This study provides scientific diagnostic evidence and technical optimization solutions for the application of large language models in the external communication of China’s national image.