Extracting the 'Why': Flood Cause Annotation and LLM Benchmarking from Multi-Platform Crisis Data

Abstract

We present a framework for extracting and classifying flood cause information from crisis-related online media data. By harnessing data across multiple online platforms from both social media and online news sources, we can understand causes of floods ranging from environment factors like heavy rainfall to human-related factors like infrastructure failure. This task is operationally critical not only for disaster response but also for public accountability, yet it remains a challenging task for automated systems.  Our key contributions are:  (1) a comprehensive flood cause annotation protocol that operates at two levels of granularity i.e., document-level and sentence-level, with strict exclusion criteria for distinguishing causal statements from impacts; and (2) initial benchmarking experiments to evaluate the performance of zero-shot large language models (LLMs) on the classification task.  Although the underlying corpus—covering 70 distinct flood events from 2022–2025 across News, X (formerly Twitter), YouTube, and Bluesky is not publicly released, we describe the data construction methodology in full to support replication on comparable data. Our initial analysis on a stratified 10-event sample demonstrate that LLMs struggle: they have high recall but suffer from low precision in flood cause classification.