LLMs for Causal Relation Extraction in Disaster Social Media

A recent research article introduces a framework grounded in expert evaluation to determine if Large Language Models can identify causal relationships in social media posts related to disasters. This study tackles the difficulty posed by the informal, fragmented, and context-sensitive nature of these posts, which often recount personal experiences instead of clear causal connections. The researchers analyze causal graphs produced by LLMs against reference graphs derived from disaster-specific reports and assess whether the identified relationships are backed by evidence from after the event or merely reflect the models' inherent biases. The results underscore both the opportunities and dangers of employing LLMs for extracting causal relations in systems that support disaster decision-making.