Abstract
Automated cause classification of fire accident reports (FIREAR) is crucial for enhancing public safety and developing data-driven prevention strategies. However, existing deep learning models often struggle with the unique challenges these documents present-namely their extreme length, high semantic noise, and fragmented causal information. To overcome these limitations, we propose the Fire Accident Reports Attention Mechanism (FAR-AM), a novel hybrid deep learning framework. FAR-AM first uses a large language model (LLM) to preprocess lengthy raw reports into concise, high-signal summaries. Its core architecture then employs an inter-layer self-attention mechanism to dynamically fuse hierarchical features across all encoder layers of BERT. The fused features are subsequently processed by a TextCNN for final classification. We evaluate FAR-AM on AGNews(title), AGNews(content), THUCNews, and our real-world FIREAR corpus. FAR-AM outperforms strong transformer baselines, including RoBERTa. On the FIREAR dataset, it achieves 73.58% accuracy and 70.65% F1. A comprehensive ablation study further validates the contribution of each component in the multi-stage framework. These results indicate that, for complex domain-specific tasks, specialized hybrid architectures can be more effective and robust than monolithic, general-purpose models.