Large language models-empowered automatic knowledge graph development based on multi-modal data for building health resilience

Abstract

Improving the health resilience of building (BHR) helps keep stable health status of both the building and its occupants under disasters. As BHR is an emerging concept, there is no structured knowledge graph to understand the whole process of BHR under disasters. Therefore, this study aims to build a structured BHR knowledge graph based on multi-modal data, providing sufficient structured knowledge for BHR enhancement. An automated knowledge graph construction approach is proposed to empower the ontology design and triple extraction by large language models (LLMs), and validation processes based on In-context Learning (ICL) prompts. A case study is conducted to construct the knowledge graph of BHR under rainstorms in Hong Kong. The performance of the proposed LLMs-empowered knowledge extraction is also validated based on natural language processing metrics and LLMs-based Evaluation (LLMs-Eval). BHR knowledge graph indicates the potential relations between disasters, factors, response actions, and the health status of the building and occupants, and provides insight to guide the BHR enhancement. The superiority of the proposed LLMs-empowered automated knowledge graph construction approach is proven, implying LLMs have great potential in knowledge graph construction, not only for BHR but also for other concepts that require structured knowledge for further explorations and analyses.