面向大模型的人机协作翻译与译后编辑研究——基于科技著作实践经验分析

Human-Machine Collaborative Translation and Post-Editing for Large Language Models—A Translation Studies Analysis Based on Practical Experience

摘要: 生成式人工智能(ChatGPT、DeepSeek和UPDF AI等)的快速发展，推动机器翻译进入以大模型为核心的新范式，人机协作已逐渐成为实现高质量翻译的重要路径。相较于传统的多人协作翻译模式，这一新范式在语篇连贯性、语义生成能力、术语一致性以及逻辑结构呈现等方面优势显著。但在专业性较强的技术类著作翻译中，大模型仍面临知识精度不稳定、长文本语篇一致性欠佳等问题。针对生成式人工智能引发的翻译范式变革，本研究立足于大模型专业著作的翻译实务，构建“LinguaPro翻译大模型 + 译后编辑 + 交叉审校”的协同翻译路径，系统总结大模型在长篇技术文本翻译中的表现特征，重点梳理译后编辑在术语预设、句段修订、章节审校及整体质量控制中的核心要求，并深入剖析人机协作过程中角色分工、流程设计与质量校验标准的重构机制。研究结果有助于完善以大模型为核心的人机协作翻译模式，为复杂技术文本的高质量、可控化翻译提供可操作的实践方案。

Abstract: The rapid development of generative artificial intelligence (such as ChatGPT, DeepSeek, and UPDF AI) has propelled machine translation into a new paradigm centered on large language models (LLM), in which human-machine collaboration is increasingly regarded as a key pathway to achieving high-quality translation. Compared with traditional multi-translator collaboration models, this paradigm demonstrates clear advantages in discourse coherence, semantic generation capability, terminological consistency, and the presentation of logical structure. Nevertheless, in the translation of highly specialized technical monographs, LLM continues to face challenges, including instability in knowledge precision and insufficient discourse consistency in long texts. In response to the paradigm shift in translation triggered by generative AI, this study is grounded in the practical translation of professional works concerning large models, and develops a collaborative translation pathway integrating the LinguaPro translation large model, post-editing, and cross-review. The study systematically summarizes the performance characteristics of LLM in long-form technical translation, with a particular focus on the core requirements of post-editing in terminological predefinition, sentence- and paragraph-level revisions, chapter-level review, and overall quality control. It further analyzes the restructuring mechanisms for role allocation, workflow design, and quality assurance standards in human-machine collaboration. The findings contribute to the refinement of LLM-centered collaborative translation models and provide a practical, actionable pathway for achieving high-quality and controllable translation of complex technical texts.