Harnessing prompt-based large language models for disaster monitoring and automated reporting from social media feedback

Abstract

In recent years, social media has emerged as one of the main platforms for real-time reporting of issues during disasters and catastrophic events. While great strides have been made in collecting such information, there remains an urgent need to improve user reports’ automation, aggregation, and organization to streamline various tasks, including rescue operations, resource allocation, and communication with the press. This paper introduces an innovative methodology that leverages the power of prompt-based Large Language Models (LLMs) to strengthen disaster response and management. By analyzing large volumes of user-generated content, our methodology identifies issues reported by citizens who have experienced a disastrous event, such as damaged buildings, broken gas pipelines, and flooding. It also localizes all posts containing references to geographic information in the text, allowing for aggregation of posts that occurred nearby. By leveraging these localized citizen-reported issues, the methodology generates insightful reports full of essential information for emergency services, news agencies, and other interested parties. Extensive experimentation on large datasets validates the accuracy and efficiency of our methodology in classifying posts, detecting sub-events, and producing real-time reports. These findings highlight the practical value of prompt-based LLMs in disaster response, emphasizing their flexibility and adaptability in delivering timely insights that support more effective interventions.