Contextualized Early Detection of Online Firestorms: A Sequential LLM-Based Approach

arXiv cs.AI · 6 hours ago

Researchers developed an LLM-based system to detect online firestorms on Reddit by analyzing discussion threads in two modes: retrospective classification of complete threads and sequential early warning detection. The early warning mode processes threads as they develop and flags escalating discussions after only a small number of comments and contributors, using indicators like negativity share and escalation level. This approach enables earlier intervention in negative discourse escalation compared to existing detectors that rely on volume signals or sentiment scores alone.