地震地质 ›› 2026, Vol. 48 ›› Issue (5): 1552-1569.DOI: 10.3969/j.issn.0253-4967.20260047

• 新技术应用 • 上一篇    

基于大语言模型的文献驱动构建活动断层知识库的方法--以京西北为例

王瑛奎1),2)(), 吴熙彦1),*(), 任俊杰2)   

  1. 1) 中国地震局地质研究所, 地震动力学与强震预测全国重点实验室(中国地震局地质研究所), 北京 100029
    2) 中国地质大学(北京), 地球科学与资源学院, 北京 100083
  • 收稿日期:2026-03-31 修回日期:2026-05-15 出版日期:2026-11-11 发布日期:2026-11-11
  • 通讯作者: 吴熙彦
  • 作者简介:

    王瑛奎, 男, 2003年生, 现为中国地质大学(北京)地理学专业在读硕士研究生, 主要从事活动构造与地震研究, E-mail: 。

  • 基金资助:
    国家重点研发计划项目(2024YFC3012900)

A LITERATURE-DRIVEN METHOD FOR CONSTRUCTING AN ACTIVE FAULT KNOWLEDGE BASE BASED ON LARGE LANGUAGE MODELS: A CASE STUDY OF THE NORTHWESTERN BEIJING AREA

WANG Ying-kui1),2)(), WU Xi-yan1),*(), REN Jun-jie2)   

  1. 1) State Key Laboratory of Earthquake Dynamics and Forecasting, Institute of Geology, China Earthquake Administration, Beijing 100029, China
    2) China University of Geosciences(Beijing), School of Earth Sciences and Resources, Beijing 100083, China
  • Received:2026-03-31 Revised:2026-05-15 Online:2026-11-11 Published:2026-11-11
  • Contact: WU Xi-yan

摘要:

活动断层研究及其知识库的构建对地震灾害防御具有重要的作用。传统活动断层数据库构建主要依赖人工收集整理的方式, 存在资料整理效率低、数据更新周期长等问题。近年来, 大语言模型在自然语言处理领域展现出强大的能力。将其引入活动断层知识库构建, 有望在文献信息自动提取与多源数据整合方面, 实现效率上的突破。文中提出了一种基于大语言模型的文献驱动构建活动断层知识库的方法, 并以京西北地区为研究区进行验证。研究设计了包含断层基本参数、构造背景、研究方法、古地震研究和历史地震记录5个模块的标准化报告框架, 构建了涵盖13条主要活动断层的文献库, 开发了基于Qwen3.5大语言模型和LangGraph多智能体框架的智能处理系统。系统通过专家设计的提示词, 对单篇文献进行信息提取, 对多篇文献进行信息识别与融合, 同时辅以人工校验环节, 以保障提取信息的质量, 最终实现从文献库中提取断层知识并生成断层报告。文中通过构建一套基于正确率与完整性2个核心维度的评价体系, 对生成的京西北地区13条断层报告进行评估。评价结果显示, 13条断层报告的平均正确率为77.58%, 平均完整性为80.74%。结果表明, 该方法能够有效提高活动断层信息收集与整合效率, 验证了基于大语言模型的人机协同模式在专业地质文献处理与知识库构建中的可行性与有效性。

关键词: 大语言模型, 活动断层, 知识库, 京西北地区

Abstract:

Active fault research and the construction of its knowledge base play an important role in earthquake hazard prevention. Traditional active fault database construction has mainly relied on manual collection and organization, suffering from low compilation efficiency and long data update cycles. In recent years, large language models(LLMs)have demonstrated powerful capabilities in natural language processing. Introducing them into active fault knowledge base construction is expected to achieve breakthroughs in efficiency for automatic literature information extraction and multi-source data integration. This study proposes a literature-driven method for constructing an active fault knowledge base based on large language models, validated in the northwestern Beijing region as the study area.
This study designed a report framework comprising five modules: Basic fault parameters, tectonic setting, research methods, paleoseismic studies, and historical earthquake records. A literature database was constructed covering 13 major active faults and containing over 140 related publications. An intelligent processing system was developed based on the Qwen3.5 large language model and the LangGraph multi-agent framework. Through expert-designed and refined prompts, the system extracts information from individual publications and performs information identification and fusion across multiple publications, supplemented by manual verification to ensure extraction quality. Ultimately, fault reports are generated by extracting fault knowledge from the literature database. This study established an evaluation system based on two core dimensions-accuracy and completeness-to assess the generated reports for the 13 faults in the northwestern Beijing region. The evaluation results show an average accuracy of 77.58% and an average completeness of 80.74%. Notably, generation quality varies with literature scale and parameter type: faults with scarce literature exhibit lower accuracy due to limited contextual support, whereas faults with excessive literature may suffer from increased fusion difficulty due to contradictions and redundancy, leading to decreased completeness. Modules dominated by qualitative descriptions generally demonstrate more stable extraction quality than numerical modules, which impose higher requirements on numerical precision and unit conversion.
The results confirm that this human-machine collaborative model, while ensuring information quality through standardized prompts and manual verification, significantly reduces the time and labor costs of literature compilation. Through automated single-publication extraction and multi-publication fusion, the system compresses work traditionally requiring several weeks into a much shorter cycle, and the mandatory citation mechanism effectively mitigates the risk of model hallucinations. However, this method still has several limitations: the current workflow relies on a single small-to-medium scale model, which occasionally falls short in processing professional geological terminology and complex logical reasoning in long texts; furthermore, faults with low literature coverage have their generation quality intrinsically constrained by insufficient input information.
Future work will focus on three directions: deploying larger-scale models or integrating multiple models to enhance professional terminology recognition and contradiction reasoning capabilities; establishing a dedicated geological terminology database to standardize entity extraction; and expanding the scope of literature collection to improve generation robustness for less-studied faults. The method proposed in this paper provides an actionable pathway for transforming unstructured domain literature into structured knowledge bases, and offers reference value for other geoscience branches facing similar information integration challenges.

Key words: large language models, active faults, knowledge base, northwestern Beijing area