HTCNN-Attn: a fine-grained hierarchical multi-label deep learning model for disaster emergency information intelligent extraction from social media

Résumé fourni par la source

To address the challenge of extracting fine-grained emergency information from noisy social media during disasters, we propose HTCNN-Attn, a hierarchical multi-label deep learning model. It integrates a three-level tree-structured labeling architecture, Transformer-based global feature extraction, convolutional neural network (CNN) layers for local pattern capture, and a hierarchical attention mechanism. The model employs a hierarchical loss function to enforce label consistency across three levels: binary disaster filtering (Level 1), mid-level category classification (Level 2), and fine-grained subcategory extraction (Level 3). Experiments on the Appen and HumAID datasets demonstrate superior performance, achieving an accuracy of 0.9725, a Micro-F1 of 0.7402, and a hierarchical consistency (HC) score of 0.821, outperforming state-of-the-art baselines. Ablation experiments validate the importance of hierarchical modeling, Transformer encoding, CNN layers, pre-trained embeddings, and the hierarchical attention mechanism. Cross-event generalization tests on the CrisisBench dataset demonstrate robust generalization (HC = 0.678), while its lightweight design enables efficient real-time deployment (12.0 ms latency). A case study of the 2015 Nepal earthquake validates its practical utility, where the model accurately classified tweets into hierarchical labels and routed structured information to support emergency response coordination. This demonstrates the effectiveness of the proposed model in supporting rapid, efficient, and fine-grained emergency response after disasters, thereby enhancing disaster response capabilities.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.