Extracting, Locating and Visualizing Geospatial Rescue Information in German-language Social Media

Abstract

Timely extraction of rescue-related data from social media is vital for emergency response, with event extraction and geolocation playing a key role. This paper presents a demo system that leverages Large Language Models (LLMs) and Knowledge Graphs (KGs) to identify rescue-related data from social media streams and integrate this information into a continuously updated KG, with a focus on the German city of Hamburg. Our approach utilizes an LLM to process unstructured social media text, accurately identifying events and relevant location references. LLMs in combination with in-context learning are applied for event extraction as well as geoparsing. The extracted and linked information is stored in a KG, which is both queryable for further analysis and supports downstream applications such as interactive map-based visualizations, providing real-time awareness for emergency services. Specifically, our geoparsing methods bridge the gap in the German setting, achieving state-of-the-art performance on the benchmark dataset MobIE.