FAR-AM: A hybrid attention framework for fire cause classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Automated classification of fire accident reports (FIREAR) is vital for public safety and prevention.
• Existing deep learning models face challenges with long, noisy, and fragmented fire reports.

Purpose of the Study:

• To develop a novel hybrid deep learning framework, FAR-AM, to address limitations in classifying fire accident reports.
• To improve the accuracy and robustness of automated cause classification for domain-specific documents.

Main Methods:

• Utilized a large language model (LLM) for preprocessing lengthy reports into concise summaries.
• Developed an inter-layer self-attention mechanism to fuse hierarchical BERT features.
• Employed TextCNN for final classification of fused features.

Main Results:

• FAR-AM achieved 73.58% accuracy and 70.65% F1 score on the FIREAR dataset.
• Outperformed strong transformer baselines like RoBERTa on various benchmarks.
• Ablation studies confirmed the effectiveness of each component in the FAR-AM framework.

Conclusions:

• Specialized hybrid architectures like FAR-AM are more effective for complex, domain-specific NLP tasks than general-purpose models.
• FAR-AM offers a robust solution for analyzing challenging fire accident report data.
• The proposed method enhances public safety through improved data-driven prevention strategies.