From knowledge graph construction to retrieval-augmented generation: a framework for comprehensive earthquake emergency support

Effective decision-making during earthquake emergencies requires rapid access to accurate, structured, and context-specific knowledge. However, existing knowledge resources in this domain are fragmented, heterogeneous, and largely unstructured, causing decision-makers to rely heavily on intuition or scattered textual materials, which often results in delayed, inconsistent, or suboptimal emergency responses. To address these challenges, this study proposes a structured framework integrating large language models (LLMs) with knowledge representation techniques to systematically construct domain-specific knowledge graphs (KGs) tailored explicitly for earthquake emergency scenarios. The framework comprises three primary stages: (1) developing an ontology that encompasses the complete earthquake emergency management cycle – prevention, preparedness, response, and recovery – as well as earthquake-specific measures, models, terminology, and attributes; (2) guiding LLMs with structured prompts to extract entities, relationships, and attributes from unstructured data; and (3) employing a knowledge fusion strategy to resolve ambiguities and consolidate information across the graph. From a corpus of 2682 professional documents, including emergency plans, technical standards, and specialized books, the framework extracted 284,801 entities and over 80,000 unique relationship types, subsequently consolidated into approximately 1000 meaningful categories. The final KG, refined through entity fusion and clustering, comprises over 268,000 nodes and 833,000 relationships. To effectively utilize the constructed KG, we developed an Improved Hybrid Retrieval-Augmented Generation (HybridRAG) application framework, integrating symbolic retrieval from the KG with semantic similarity-based retrieval from a vector database. This dual retrieval approach enables LLMs to generate responses that are both semantically coherent and deeply grounded in operational knowledge. Comparative experiments conducted on a newly constructed dataset of 150 earthquake-specific questions demonstrated that the Improved HybridRAG method significantly outperforms standard methods – including LLM-only, semantic vector-based retrieval, and purely symbolic retrieval – in accuracy, clarity, comprehensiveness, conciseness, and relevance. These findings validate the advantage of combining structured and semantic knowledge retrieval, illustrating the framework’s capability to provide reliable, contextually aligned, and actionable insights for decision-makers in earthquake emergency scenarios. Future work will focus on improving attribute completeness, refining entity alignment techniques to capture complex semantic nuances, and exploring multimodal data integration to further extend the KG’s utility. This study underscores the potential of systematically combining structured KGs and LLMs to significantly enhance decision-making capabilities in disaster management contexts.