融合GPT和知识图谱的洪涝应急决策智能问答系统研究

Research on intelligent question-answering system for flood emergency decision-making with fusion of GPT and knowledge graph

为提高生成式预训练语言大模型(generative pre-trained transformer, GPT)的应急管理信息分析能力,以实现洪涝灾害应急处置过程中的在线辅助决策,提出融合GPT和知识图谱的应急决策智能问答系统(KG-GPT)。改进GPT架构以识别问题中的关键信息,利用知识图谱推理应急领域知识并生成具有逻辑性的回答;结合洪涝灾害的实际应急决策问答数据集并编制演练脚本,使用自动评估和专家评估方法将本系统与GPT进行对比实验。研究结果表明:该系统成功融合应急领域知识图谱和GPT模型,能够深刻理解问题的背景信息并生成流畅回答;与GPT相比,该系统可为决策者提供更快速准确的在线辅助决策工具。研究结果可提升洪涝灾害应急信息分析和决策效率。

To enhance the analytical capability of generative pre-trained transformer(GPT)models in emergency management information processing,aiming to achieve the online assistant decision-making during the emergency response of flood disaster,an intelligent question-answering system of emergency decision-making with the fusion of GPT and knowledge graph(KG-GPT)was proposed.The GPT architecture was improved to identify the key information in the questions,and the knowledge graph was used to reason the knowledge in the emergency field and generate the logical answers.Combining the actual emergency decision-making question and answer data set of flood disaster and compiling the drill script,the system was compared with GPT-2.0 by using the automatic evaluation and expert evaluation methods.The research results show that the system successfully fuses the knowledge graph in the emergency field and GPT model,and can deeply understand the background information of problems and generate the smooth answers.Compared with GPT-2.0,the system offers the decision-makers a faster and more accurate online assistant decision-making tool.The research results can improve the efficiency of emergency information analysis and decision-making in flood disasters.