面向政务服务的知识增强问答系统研究与实现

Research and Realization of Knowledge-Enhanced Q&A System for Government Services

摘要: 针对政务领域知识检索准确性不足、回答缺乏事实依据以及大语言模型专业知识能力有限等问题，本研究构建了一种融合领域知识图谱与大语言模型的政务智能问答系统(GOV_QA)。首先，对政务服务数据进行知识建模与结构化组织，构建政务领域知识图谱；其次，提出分层知识召回方法，通过实体匹配构建局部子图，将图谱信息转化为知识单元并筛选高相关知识；最后，将召回知识与用户问题共同输入大语言模型，约束回答生成。实验结果表明，GOV_QA的BERTScore精确率、召回率和F1值分别达到0.7600、0.8557和0.8025，主观评价平均得分为4.49分，其中76%的回答获得5分评价，整体性能优于Doubao、DeepSeek和Qwen等基准模型，为政务智能问答提供高可靠的技术方案。

Abstract: In view of the insufficient accuracy of knowledge retrieval, the lack of factual support for answers, and the limited domain knowledge of large language models in government services, this study develops an intelligent question answering system named GOV_QA, which integrates a domain knowledge graph with a large language model. First, government service data are modeled and structurally organized to construct a government service knowledge graph. Second, a hierarchical knowledge retrieval method is proposed to construct local subgraphs through entity matching, transform graph information into knowledge units, and select highly relevant knowledge. Finally, the retrieved knowledge and user questions are jointly input into the large language model to constrain answer generation. Experimental results show that GOV_QA achieves BERTScore precision, recall, and F1 values of 0.7600, 0.8557, and 0.8025, respectively. The average subjective evaluation score is 4.49, with 76% of the answers receiving a score of 5. The overall performance is superior to that of the Doubao, DeepSeek, and Qwen baseline models, providing a reliable technical solution for intelligent question answering in government services.