PSO-SciBERT: Particle Swarm Optimization for Multimodal Classification of Scientific Articles on Extreme Climate Events

Extreme climate events represent a growing global challenge, with severe impacts across diverse regions, demanding efficient tools for optimization and analysis of scientific knowledge. This work presents an automatic classification system for articles on natural disasters, with two contributions: the AMCLIMA-BR dataset, comprising 700 articles classified across three dimensions (disaster type, management phase, and learning paradigm), and a particle swarm optimization (PSO) approach applied to SciBERT for imbalanced data. Results demonstrate significant improvements with PSO-SciBERT, achieving gains of up to 26.4% in balanced accuracy, evidencing the robustness of the optimized model.