| 摘要: |
| 针对开源情报分析中网页信息提取问答问题，提出一种融合生成式大语言模型（Large Language Model，LM）、XPath与检索增强生成（Retrieval-Augmented Generation，RAG）的方法，涉及动态模板化提示策略与多粒度语义检索。动态模板基于情报类型生成领域知识约束提示，提升实体提取精度；多粒度检索构建文档-段落-实体三级体系，结合BERT-Top玨算法优化长文本信息定位。通过OpenKG知识库对齐实体构建属性-关系-事件三维网络，增强复杂事件逻辑分析。该方法在ClueWeb22与TAC-KBP2022数据集上的提取率为0.85，回答准确率为0.78，相比传统RAG，性能提升18%~31%。实际应用中，热点事件简报关键事实准确率达92%，综合成本仅为GPT-4的12%。 |
| 关键词: 开源情报分析 网页信息提取；生成式大语言模型；检索增强生成 |
| DOI：10.20079/j.issn.1001-893x.250204001 |
|
| 基金项目: |
|
| An Open Source Intelligence Analysis Method Based on Generative LLM |
| CHENG Leifeng,LUO Ji,WANG Lei,ZHU Min,TAO Sitong |
| (1.College of Computer Science,Sichuan University,Chengdu 610065,China;2.Southwest China Institute of Electronic Technology,Chengdu 610036,China;3.China Telecom Digital Intelligence Technology Co.,Ltd.,Beijing 100001,China) |
| Abstract: |
| The authors propose a method integrating generative large language models(LLMs),XPath,and retrieval-augmented generation(RAG) for web page information extraction in open-source intelligence analysis.Key innovations include a dynamic templated prompting strategy and multi-granularity semantic retrieval.The dynamic templates generate domain-constrained prompts based on intelligence types(events/persons/organizations),enhancing entity extraction accuracy.The multi-granular retrieval establishes a document-paragraph-entity hierarchy optimized by the BERT-Top玨 algorithm for fragmented long-text information.By aligning entities with OpenKG,a three-dimensional attribute-relation-event network is constructed to strengthen complex event analysis. |
| Key words: open source intelligence analysis web information extraction generative large language model retrieval-augmented generation |
