Legal documents such as rulings, contracts, and statutes are typically voluminous and complex, making manual processing inefficient. This paper employs deep learning algorithms for legal document text classification, proposing two deep learning models: TextCNN and GRU. To achieve superior classification performance, a combined model TextCNN-GRU was constructed using a linearly weighted model fusion approach. The highest classification accuracy was attained when both models were fused with equal weights of 0.5. Comparisons with other single models and the high-performing CNN-HAN model demonstrate that TextCNN-GRU significantly outperforms in accuracy, macro-precision, macro-recall, and macro scores. Generalization experiments further validate the model’s performance. To explore further optimization directions, 500 misclassified samples were analyzed, revealing issues such as ambiguous category boundaries, non-standard textual expressions, and mixed narratives involving multiple events and roles. Future improvements could include refining the dataset, incorporating additional domain knowledge, and implementing secondary classification schemes to enhance classification performance.
