A Study on a RAG-Based Approach to Suggesting Revisions to Regional Disaster Management Plans Based on the Basic Disaster Management Plan

Abstract

Regional disaster management plans are revised annually in accordance with amendments to the Basic Disaster Management Plan, and this revision process requires substantial effort. Previous studies have focused on Retrieval-Augmented Generation (RAG), which generates responses based on given documents, and have developed RAG systems capable of generating answers consistent with the descriptions in regional disaster management plans. However, the feasibility of presenting revision targets and draft revisions for regional disaster management plans based on amendments to the Basic Disaster Management Plan has not been examined. Therefore, this study develops a RAG model that incorporates a preprocessing step to automatically organize the amended contents of the Basic Disaster Management Plan and uses it to present revision targets and draft revisions. Experimental results suggest that the proposed method can realize a continuous process from organizing the amended contents of the Basic Disaster Management Plan to presenting draft revisions, indicating its potential to support the revision of regional disaster management plans. In addition, under optimal conditions, the proposed method was able to present revision targets and draft revisions with accuracy generally comparable to that of existing generative AI services.