面向AIGC幻觉治理的档案资源高质量数据集建构研究
Research on the Construction of Archives High-Quality Datasets for AIGC Hallucination Governance
摘要
生成式人工智能(AIGC)广泛渗透于社会生活与科学研究的同时,其实际应用常常受"幻觉"问题的困扰.生成事实性谬误、逻辑矛盾及虚构内容等失真答案,严重削弱了模型的可信度与应用价值.而幻觉根植于大语言模型的统计生成范式,难以通过算法切换完全消除.档案资源凭借其本质属性与有机整体逻辑构成的理论内核,以及在规模性储备、体系化供给与生态化赋能三个维度上的现实支撑,与高质量数据集建设形成高度契合.本研究基于档案资源的独特属性,从顶层规划与技术落地两个层面,系统设计了其向高质量数据集转化的框架与实施路线,重点突破非结构化手写文本智能识别、全宗感知数据集范式构建、内容可信度风险预警三大难点,并将档案的时间、内容与利用三维属性嵌入RAG增强检索流程,为可信人工智能的知识基础设施建设提供理论参考与实践路径.
AI-Generated Content(AIGC)is deeply integrated into social life and scientific research,yet its practical applications are often plagued by the issue of"hallucination."The generation of distorted answers such as factual fallacies,logical contradictions,and fabricated content severely undermines the credibility and application value of models.Hallucination is rooted in the statistical generation paradigm of large language models and cannot be entirely eliminated through algorithmic switching.Archival resources,with their theoretical core formed by inherent attributes and organic holistic logic,along with practical support in three dimensions—extensive storage,systematic aggregation,and ecological empowerment—align closely with the construction of high-quality datasets.This study leverages the unique attributes of archival resources to systematically design a framework and implementation roadmap for transforming them into high-quality datasets.It focuses on overcoming three major challenges:intelligent recognition of unstructured handwritten text,the construction of a fonds-aware dataset paradigm,and risk early warning for content credibility.Additionally,the three-dimensional attributes of archives—time,content,and utilization—are embedded into the RAG-enhanced retrieval process,providing theoretical references and practical pathways for the knowledge infrastructure development of trustworthy artificial intelligence.
年,卷(期):
2026, (9)
分类号:
G270.7
机标分类号:
TP391
F294
G206.3
在线出版日期:
2026年10月21日
基金项目:
2419:上海市档案科技项目CSI25B04:中国索引学会规划项目,