首页|期刊导航|计算机与现代化|面向政务服务问答的大模型检索增强生成方法

面向政务服务问答的大模型检索增强生成方法OA

Retrieval-augmented Generation Method for Large Language Models in Government Service Q&A

中文摘要英文摘要

针对政务服务问答系统面临的文档数量庞大导致文本块检索精度差的问题,提出一种基于高斯混合模型文档聚类的大模型检索增强生成方法DC-RAG.该方法包括知识库构建和答案生成2个部分.在知识库构建阶段,将所有政策文档向量化后使用高斯混合模型做软聚类,构建类别-文档-文本块的树形检索知识库.在答案生成阶段,首先对用户问题做增强并向量化,然后在知识库中按类别、文档、文本块的顺序检索与问题相关的文档与文本块,最后根据文档元数据与文本块使用大模型生成答案.爬取河南省科技厅网站上的政策、通知文档作为实验数据,通过本地部署DeepSeek-14B模型实现问题增强与答案生成,使用语义精准度、召回率和F1得分作为回答质量的评价指标.实验结果表明,与已有方法相比DC-RAG能够将3个指标平均提高9百分点左右.实验进一步分析了本文方法中不同模块和参数对结果的影响,结果表明对问题做增强及使用大参数量模型能有效提高系统回答质量.本文方法能够为政务服务问答系统的构建提供参考.

To address the challenges of low retrieval accuracy for text chunks in intelligent government service Q&A systems caused by the massive volume of policy documents,a retrieval-augmented generation method based on Gaussian Mixture Model(GMM)document clustering named DC-RAG is proposed.The method comprises two main components,including knowledge base construction and answer generation.During the knowledge base construction phase,all policy documents are vectorized and undergo soft clustering using GMM to build a tree-structured retrieval knowledge base with a hierarchy of categories,docu-ments,and text chunks.In the user query phase,the user's question is first semantically enhanced and vectorized.Then,rel-evant documents and text segments are retrieved from the knowledge base by category,document,and text chunk.Finally,an-swers are generated by a reasoning model using the retrieved document metadata and text chunks.Policy and announcement docu-ments were crawled from the website of the Henan Provincial Department of Science and Technology to serve as the experimental data.A locally deployed DeepSeek-14B model was utilized for question augmentation and answer generation.Semantic preci-sion,recall,and F1-score are employed as evaluation metrics.The experimental results demonstrate that DC-RAG improves all three metrics by an average of 9 percentage points compared to existing methods.The experiment further analyzes the impact of different modules and parameters on the experimental results,and it reveals that question enhancement and the use of large-parameter models significantly contribute to system performance.The proposed method can serve as a reference for the develop-ment of government service Q&A systems.

李哲辉;王昌海

河南省科学技术情报中心,河南 郑州 450003郑州轻工业大学软件学院,河南 郑州 450000

信息技术与安全科学

深度学习大语言模型检索增强高斯混合模型政务问答

deep learninglarge language modelretrieval enhancementGaussian mixture modelgovernment question and answering

《计算机与现代化》 2026 (9)

22-28,54,8

国家自然科学基金资助项目(62072414)河南省自然科学基金资助项目(252300420392)河南省科技攻关项目(252102211083)

10.3969/j.issn.1006-2475.2026.09.003

评论