北京第二外国语学院学报 ›› 2026, Vol. 48 ›› Issue (3): 51-68.DOI: 10.12002/j.bisu.648

• 翻译研究(AI风险治理专栏 主持人:任东升) • 上一篇    下一篇

对外型国家翻译数字资源平台构建——基于大语言模型风险治理视角

杨立学1(), 朱华2()   

  1. 1 天津职业技术师范大学外国语学院天津 300222
    2 天津外国语大学高级翻译学院天津 300204
  • 出版日期:2026-06-30 发布日期:2026-07-08
  • 作者简介:杨立学,天津职业技术师范大学外国语学院教授,300222,研究方向:国家翻译实践与翻译技术。电子邮箱:yanglixue@tute.edu.cn;
    朱华,天津外国语大学高级翻译学院,300204,研究方向:智能翻译技术。电子邮箱:zhuhua@tjfsu.edu.cn
  • 基金资助:
    天津市教委社会科学重大项目“人类命运共同体英译助建对外话语体系研究”(2023JWZD50)

Developing an External-Facing State Translation Digital Resource Platform: A Risk-Governance Perspective on Large Language Models

Yang Lixue1(), Zhu Hua2()   

  1. 1 Tianjin University of Technology and Education, Tianjin 300222, China
    2 Tianjin Foreign Studies University, Tianjin 300204, China
  • Published:2026-06-30 Online:2026-07-08

摘要:

大语言模型在提升对外型国家翻译效率的同时,也可能带来信息泄露、语料偏见等安全风险。为此,本研究构建了一套以技术治理为核心的安全保障方案。通过建立国家翻译风险矩阵,系统识别并评估数据、算法和应用3个层面中政治、文化、技术和伦理4类风险的优先级,进而提出集成化数字资源平台的治理路径,具体包括:本地部署大语言模型并经垂直领域语料微调,将其作为核心翻译引擎;整合检索增强生成(RAG)技术,关联依据专家经验构建的知识库与术语知识库,实现精准知识供给;最终通过人机交互修正输出,抑制风险表述。该方案将大语言模型能力与国家安全要求相结合,为生成式人工智能时代的国家话语安全保障与文化主权维护提供了可管控、可迭代的基础设施。

关键词: 大语言模型; 国家翻译实践; 风险治理; 数字资源平台; 微调技术; 检索增强生成技术

Abstract:

While large language models (LLMs) have significantly enhanced the efficiency of external-facing state translation, they concurrently introduce security risks such as information leakage and corpus bias. To address these challenges, this study develops a security assurance framework centered on technological governance. The research first establishes a state translation risk matrix, which systematically identifies and prioritizes four categories of risks — political, cultural, technical, and ethical — across three dimensions: data, algorithms, and application. Building upon this analysis, an integrated digital resource platform is proposed as the core governance pathway. The platform involves: (1) the on-premises deployment of an LLM, which is fine-tuned using domain-specific corpora to serve as the core machine translation engine; (2) the integration of Retrieval-Augmented Generation (RAG) technology, linking knowledge bases and terminological knowledge bases constructed from expert experience to enable precise and controllable knowledge supply; and (3) the implementation of a human-in-the-loop workflow to interactively revise the model’s outputs, thereby effectively mitigating risk-prone expressions. This proposed solution reconciles the powerful capabilities of LLMs with national security requirements, providing a manageable and iterative infrastructure for safeguarding national discourse security and preserving cultural sovereignty in the era of generative artificial intelligence.

Keywords: large language model; state translation program; risk governance; digital resource platform; fine-tuning technology; Retrieval-Augmented Generation technology

中图分类号: