Advancing cross-domain emergency classification with multi-view adversarial learning

The growing volume of natural and man-made emergency data requires effective real-time classification across various emergency domains on social media. However, current Unsupervised Domain Adaptation (UDA) methods for emergency data classification face two key challenges: predominant grounding in natural disaster contexts that limits generalizability, and difficulty handling domain shifts caused by heterogeneous distributions, linguistic variations, and emotional expressions. To overcome these challenges, we propose Multi-View Adversarial Neural Networks for Robust Unsupervised Domain Adaptation (MARDA), a novel framework that integrates adversarial domain adaptation with multi-view feature learning. First, a cross-view processor consisting of semantic and emotional processors, along with interactive integrators, is designed to generate rich and comprehensive multi-view feature representations. Second, an adaptive weighted domain enhancer is developed to dynamically balance contributions from multiple views, effectively aggregating discriminative information in various domains. Third, an adversarial cross-view optimizer is proposed that employs a minimax game and feature consistency regularization, thereby enhancing cross-domain generalization. Experimental results on four real-world emergency datasets with 24,008 samples show that MARDA outperforms advanced baselines by 7.39% and exceeds large language models by 1.59% in average F1-Score, demonstrating its effectiveness as a generalized solution for cross-domain emergency event classification.