Reproduce “Learning to Predict Charges for Criminal Cases with Legal Basis” on OpenResearch

About this paper

Bingfeng Luo, Yansong Feng, Jianbo Xu, Xiang Zhang, Dongyan Zhao

Researchers from Peking University and Beijing Institute of Big Data Research introduced an attention-based neural network that jointly predicts criminal charges and identifies supporting law articles from case descriptions. This model achieved a micro-F1 score of 90.21% for charge prediction and demonstrated improved generalization to informal news reports (79.12% micro-F1) compared to SVM baselines, showing that relevant articles enhance accuracy and explainability.

Key insights

• Dynamically generating context vectors for article encoding and aggregation based on fact embeddings allows the model to effectively focus on the most relevant parts of legal statutes for a given case.
• Explicitly incorporating and supervising the extraction of relevant law articles quantitatively improves charge prediction accuracy, confirming their foundational role beyond mere justification in civil law systems.
• The proposed neural network architecture, with its attention mechanisms and pre-trained embeddings, demonstrates superior ability to distinguish nuanced charges and generalize to informal language compared to traditional methods.

Summary by alphaXiv.