In this paper, we propose a multi-task legal document analysis framework for feature extraction and case type prediction. Our method integrates Transformer text encoders, hierarchical feature aggregation, and two-branch supervision to model fact statements, procedural records, legal citations, and party information in legal documents. A corpus of 12,000 Chinese legal documents covering eight case categories was constructed, with 9600 samples for training, 1200 samples for validation, and 1200 samples for testing. Experimental results show that the proposed model achieves 93.4% micro f1 value and 91.8% macro f1 value in feature extraction, and achieves 92.6% accuracy and 91.1% macro f1 value in case type prediction. The average inference time per document was 38 ms. Under the type constraint correction, the boundary agreement rate of the system reaches 92.7%, and shows a good overall cross-class discrimination stability. The framework can provide highly accurate structured results for legal document indexing, content retrieval and case intelligent analysis, and can also support downstream applications such as evidence organization and semantic recommendation in information systems.
