[Paper Review] Enhancing Emergency Decision-making with Knowledge Graphs and Large Language Models

Minze Chen, Zhenxiang Tao+4 more|arXiv (Cornell University)|Nov 15, 2023

Topic Modeling4 citations

TL;DR

This paper proposes E-KELL, a hybrid system integrating a structured emergency knowledge graph with large language models (LLMs) to enhance decision-making in crisis scenarios. By guiding LLMs through a prompt chain over domain-specific knowledge, E-KELL reduces hallucinations and improves reasoning, achieving high scores (9.03–9.09) in comprehensibility, accuracy, conciseness, and instructiveness from emergency responders in real-world evaluations.

ABSTRACT

Emergency management urgently requires comprehensive knowledge while having a high possibility to go beyond individuals' cognitive scope. Therefore, artificial intelligence(AI) supported decision-making under that circumstance is of vital importance. Recent emerging large language models (LLM) provide a new direction for enhancing targeted machine intelligence. However, the utilization of LLM directly would inevitably introduce unreliable output for its inherent issue of hallucination and poor reasoning skills. In this work, we develop a system called Enhancing Emergency decision-making with Knowledge Graph and LLM (E-KELL), which provides evidence-based decision-making in various emergency stages. The study constructs a structured emergency knowledge graph and guides LLMs to reason over it via a prompt chain. In real-world evaluations, E-KELL receives scores of 9.06, 9.09, 9.03, and 9.09 in comprehensibility, accuracy, conciseness, and instructiveness from a group of emergency commanders and firefighters, demonstrating a significant improvement across various situations compared to baseline models. This work introduces a novel approach to providing reliable emergency decision support.

Motivation & Objective

• Address the critical need for reliable, evidence-based decision support in emergency management, where human cognition is often overwhelmed by complexity.
• Overcome the limitations of standalone LLMs—particularly hallucinations and weak reasoning—by grounding their outputs in structured, domain-specific knowledge.
• Develop a system that integrates emergency regulations and standards into a formal knowledge graph to support accurate, traceable, and interpretable decision-making.
• Improve the interpretability and trustworthiness of AI-driven emergency recommendations through explicit reasoning chains over structured knowledge.
• Evaluate the system’s performance across diverse emergency scenarios using feedback from domain experts, including emergency commanders and firefighters.

Proposed method

• Construct a structured emergency knowledge graph using Chinese emergency standards, regulations, and incident response procedures, formalizing relationships between hazards, actions, and response stages.
• Design a multi-step prompt chain that guides the LLM to reason over the knowledge graph, ensuring each response is grounded in factual, pre-validated knowledge.
• Integrate the LLM with the knowledge graph at inference time, using retrieval-augmented generation to constrain outputs within the scope of known emergency protocols.
• Implement a modular system architecture that supports multi-turn, context-aware dialogue for dynamic emergency scenarios, such as chemical leaks or fires.
• Apply retrieval and reasoning mechanisms to align LLM-generated responses with official emergency response frameworks and real-time situational data.
• Use human-in-the-loop evaluation with emergency professionals to validate response quality across key dimensions: accuracy, clarity, conciseness, and instructiveness.

Experimental results

Research questions

• RQ1 How can large language models be effectively constrained to reduce hallucination in high-stakes emergency decision-making?
• RQ2 To what extent can a knowledge graph improve the reasoning quality and factual consistency of LLM-generated emergency recommendations?
• RQ3 Can a hybrid system of LLMs and structured knowledge outperform standalone LLMs in terms of accuracy and reliability for emergency response tasks?
• RQ4 How do emergency responders perceive the comprehensibility and practicality of LLM-generated responses when grounded in domain-specific knowledge?
• RQ5 What is the impact of structured reasoning via a prompt chain on the coherence and actionability of emergency decision support outputs?

Key findings

• E-KELL significantly reduces hallucination and improves factual consistency by grounding LLM outputs in a curated emergency knowledge graph derived from official Chinese standards.
• In real-world evaluations, E-KELL achieved average scores of 9.06 (comprehensibility), 9.09 (accuracy), 9.03 (conciseness), and 9.09 (instructiveness) from emergency commanders and firefighters.
• The system outperformed baseline LLMs (e.g., ChatGLM-6b, GPT-3.5) in generating structured, actionable, and contextually appropriate emergency response guidance.
• The prompt chain mechanism effectively guided the LLM to reason step-by-step over the knowledge graph, resulting in more coherent and logically connected responses.
• Emergency professionals found E-KELL’s outputs highly practical and suitable for real-time decision support, particularly in complex scenarios like hazardous chemical leaks.
• The integration of structured knowledge with LLMs enables traceable, interpretable, and legally defensible decision support, addressing key concerns in high-risk emergency environments.