摘要
现有的自然语言大模型和搭载大模型的智能体技术主要是在通用数据集上训练的,在地理空间信息上的问答容易出现幻觉,存在回答内容相关度较低、回答不准确、缺乏实时性等问题。针对该问题,提出基于长江水利一张图的地理空间信息问答大模型智能体技术框架(GLMA),该方法主要利用智能体,理解用户的自然语言提问,驱动大模型执行长江水利一张图任务,再利用大模型融合提问和任务的返回结果,生成最终的提问回复内容。为了提高任务执行的准确性与有效性,在任务分配阶段,使用了树形的任务分配结构,可有效提高任务检索和参数生成的能力。此外,为了验证GLMA的有效性,提出了一套地理空间信息问答数据集和评估指标,在该数据集和评估指标上与最新的中文开源大模型Baichuan2、Llama3.1、ChatGLM4、Qwen2.5等进行对比研究。结果表明:GLMA在任务分配准确率和查询结果准确率等评估指标上效果最好。研究成果具有一定的扩展性,可为其他业务领域的大模型智能体研究提供参考。
The existing large models and intelligent agents powered by large models are predominantly trained on general datasets,making them prone to hallucination issues such as low relevance of answers,inaccuracies,and lack of real-time capability.To address these challenges,this paper proposes a technical framework for a large language model-based intelligent agent for geospatial information question answering,built upon the"one map"platform of Changjiang water conservancy(GLMA).This framework employs a large model agent to interpret users′natural language queries and execute corresponding tasks on the"One Map"platform.The large language model then generates final responses based on both the query and the task outcomes.To enhance the accuracy of task assignment,a tree-structured task allocation system is introduced during the task dispatch phase,significantly improving task retrieval and parameter generation capabilities.Furthermore,this paper introduces a geospatial information question-answering dataset and corresponding evaluation metrics to validate the performance of GLMA.Compared to state-of-the-art Chinese open-source models such as Baichuan2,Llama3.1,ChatGLM4,and Qwen2.5,GLMA achieves the best results in terms of task allocation accuracy and answer correctness.This research demonstrates strong extensibility and lays a foundation for future studies on large model agents in other professional domains.
作者
明晨曦
杨鹏
张志鑫
刘哲
乔延军
李杰潘
MING Chenxi;YANG Peng;ZHANG Zhixin;LIU Zhe;QIAO yanjun;LI Jiepan(Network&Information Center,Changjiang Water Resources Commission,Wuhan 430010,China;Center of Technology Innovation for Digital Enablement of River Basin Management,Wuhan 430010,China;Smart Changjiang Innovation Team of Changjiang Water Resources Commission,Wuhan 430010,China;State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430072,China)
出处
《人民长江》
北大核心
2026年第2期257-266,共10页
Yangtze River
基金
国家自然科学基金项目(42271370)
长江水利委员会流域管理数字赋能技术创新中心基金项目。
关键词
智能体
大模型
地理空间信息
长江水利一张图
AI agent
large language model
geospatial information
"one map"platform of Changjiang water conservancy
