An AI-driven approach to extract interrelationships between disasters

Abstract

Accurately identifying interrelationships between disasters is essential for comprehensive multi-disaster risk assessment. Traditional manual methods rely heavily on expert judgment, which may lead to overlooked or inconsistently documented disaster interrelationships. To address this challenge, this study develops an AI-driven approach to automatically extract disaster interrelationships from large-scale textual data using a fine-tuned Universal Information Extraction model. First, disaster interrelationships are systematically categorized into six distinct types, considering both disaster causation and impact perspectives. Secondly, a large-scale dataset is constructed by collecting 5212 Chinese-language disaster-related paper abstracts from the China National Knowledge Infrastructure (CNKI). Among them, 267 abstracts were manually annotated to train and evaluate the model. Thirdly, the fine-tuned model is applied to the remaining 4945 abstracts to extract large-scale disaster relationship triplets, with manual validation conducted for less common triplets to ensure result reliability. Finally, disaster interrelationships are visualized using complex network graphs and matrices, providing an intuitive representation of multi-disaster interrelationships. The key contribution of this research is the development of an AI-driven approach to systematically extract disaster interrelationships from large-scale datasets, improving the accuracy and scalability of identifying disaster interrelationships. Furthermore, this study establishes a comprehensive and updatable database of disaster interrelationships, addressing limitations in previous research, such as incomplete data coverage and limited exploration of relationship types, and helps scholars to identify disaster interrelationships that may have been previously overlooked.