Hierarchical Text Classification of Chinese Public Security Cases Based on ERNIE 3.0 Model

This paper introduces the concept of hierarchical classification in public security cases, establishing a tree-like hierarchical structure for case labels to better address the complex relationships between multi-level labels and text. We employ the ERNIE 3.0 pre-trained model to achieve this goal and utilize a Chinese public security case dataset for training. Through the model’s self-attention mechanism, we can capture key features related to text, labels, and hierarchical structures. Upon testing, the model demonstrates a precision of $93.6 \%$, a recall of $87.1 \%$, and an $\mathbf{F 1}$ value of $\mathbf{9 0 . 2 \%}$. It exhibits excellent performance in the hierarchical classification of public security cases, providing law enforcement agencies with an efficient tool for handling case information.