講演情報
[U15-P12]Empowering Local Resilience through Verifiable LLM Governance: Bridging the "Text-to-Action" Gap for Multi-Sector Disaster Planning
*Jie-Ru Tsai1、Po-Tsang Chen1 (1.Feng Chia University, Taiwan)
キーワード：
Large Language Models (LLMs)、Retrieval-Augmented Generation (RAG)、Disaster Governance、Knowledge Management、Actionable Decision Support、Urban Resilience.
The effectiveness of disaster governance is frequently limited by structural gaps between institutional planning and actionable field response. Existing disaster prevention protocols and standard operating procedures (SOPs) often suffer from high maintenance costs, version inconsistency, and a lack of clear execution paths for non-experts, leading to documents that are "seemingly complete but practically unusable". This study proposes an evidence-based decision support framework that integrates Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) to transform static professional texts into verifiable, actionable directives.
The core of the framework is a multi-lingual knowledge base comprising thousands of professional disaster management assets, spanning categories such as hydrometeorology, seismology, shelter operations, and business continuity management (BCM). To ensure the reliability of high-stakes outputs, we implement a "Self-RAG" architecture that enforces strict evidence-tracing and an audit-trail mechanism. This ensures that every recommendation—whether for individual Incident Action Plans (IAPs), local government plan auditing, or enterprise-level BCPs—is explicitly linked to authoritative sources and version-stamped regulations, effectively mitigating the risk of AI hallucinations.
The practical efficacy of the framework is evaluated through a stepped-wedge quasi-experimental design across diverse administrative districts (Xitun, Dali, and Wufeng). Preliminary results indicate a significant reduction in administrative workload and enhanced consistency in cross-sector disaster coordination. By lowering the professional barrier to localized planning, this research demonstrates the potential of LLMs to serve as a verifiable governance tool for advancing grassroots resilience.
The core of the framework is a multi-lingual knowledge base comprising thousands of professional disaster management assets, spanning categories such as hydrometeorology, seismology, shelter operations, and business continuity management (BCM). To ensure the reliability of high-stakes outputs, we implement a "Self-RAG" architecture that enforces strict evidence-tracing and an audit-trail mechanism. This ensures that every recommendation—whether for individual Incident Action Plans (IAPs), local government plan auditing, or enterprise-level BCPs—is explicitly linked to authoritative sources and version-stamped regulations, effectively mitigating the risk of AI hallucinations.
The practical efficacy of the framework is evaluated through a stepped-wedge quasi-experimental design across diverse administrative districts (Xitun, Dali, and Wufeng). Preliminary results indicate a significant reduction in administrative workload and enhanced consistency in cross-sector disaster coordination. By lowering the professional barrier to localized planning, this research demonstrates the potential of LLMs to serve as a verifiable governance tool for advancing grassroots resilience.