Knowledge-enhanced model with dual-graph interaction for confusing legal charge prediction.

The rapid development of natural language processing (NLP) technologies has enabled the emergence of legal intelligence assistance systems, with legal charge prediction (LCP) being a critical technology. The automatic LCP aims to determine the final charges based on fact descriptions of criminal cases. LCP assists human judges in managing workloads and improving efficiency, provides accessible legal guidance for individuals, and supports enterprises in litigation financing and compliance monitoring. However, distinguishing between confusing charges in real-world judicial practice remains a significant challenge. Most exist works cannot effectively capture complex relationships and discern subtle differences in fact descriptions while ignoring the legal schematic knowledge. In order to improve confusing LCP performance, we propose a novel knowledge-aware model for legal charge prediction that leverages Graph Neural Networks (GNNs) to capture complex relationships within criminal case descriptions. Specifically, the model constructs structural and semantic graphs from fact descriptions and integrates information from both through a dual-graph interaction process. A legal knowledge transformer generates key knowledge representations at schema and charge levels, while a knowledge matching network incorporates hierarchical charge knowledge into facts. Besides, we also propose two real-world datasets namely Criminal-All and Criminal-Confusing, containing 203 different charges and 86 confusing charges, respectively. To the best of our knowledge, these datasets are the first well-organized datasets for confusing LCP task. Experimental results demonstrate that the proposed model outperforms baselines and significantly improves the distinction of confusing charges, providing valuable support for intelligent legal judgment systems. • A legal schematic knowledge-aware model for confusing charges. • Dual-graph interaction to integrate information from structural and semantic graphs. • A hierarchical transformer to obtain legal knowledge representations. • A deep matching network for incorporating domain knowledge into the fact. • Our model improvement outperforms baselines and achieves a significant. [ABSTRACT FROM AUTHOR]