Detection and Analysis of Fake Alerts During Climate and Crisis Events Using NLP

Crisis misinformation often circulates as short “alerts” during extreme climate events (floods, storms) and humanitarian or health emergencies. Such messages can trigger panic, disrupt response operations, and reduce trust in official communication channels. Unlike general misinformation, crisis alerts are typically short, urgent, highly time-sensitive (“today”, “tonight”) and location-specific (“in Cameroon”), making them especially vulnerable to false confirmation when systems rely on generic web evidence retrieval. Contribution. This deposit provides the Semester 1 research prototype and technical report for the project “Detection and Analysis of Fake Alerts During Climate and Crisis Events Using NLP” (ITMO University, MSc Information Security). The prototype implements a hybrid verification pipeline combining: Hoax-risk linguistic analysis (urgency/panic, conspiracy framing, virality prompts, vagueness signals), and Evidence-based verification using trusted sources and institutional feeds. Crucially, the system enforces: Time-window recency constraints, inferred from the message text (e.g., strict evidence window for “today/now/tonight”), and Location consistency enforcement, requiring retrieved evidence to match extracted geographic entities when available. This directly addresses a major failure mode of naive verification systems: falsely supporting a claim due to outdated or geographically irrelevant online reports.