Reproduce “Overview of CAIL2018: Legal Judgment Prediction Competition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haoxi Zhong, Chaojun Xiao, Zhipeng Guo, Cunchao Tu, Zhiyuan Liu, Maosong Sun, Yansong Feng, Xianpei Han and 3 more
The CAIL2018 competition introduced a large-scale public dataset of over 2.6 million Chinese criminal cases to benchmark Legal Judgment Prediction (LJP), establishing three subtasks for predicting law articles, charges, and prison terms. It attracted significant participation, achieving high F-micro scores for article (0.958) and charge (0.962) prediction, while highlighting the greater challenge in estimating prison terms.
Key insights
- Advanced deep learning architectures (e.g., CNNs, RNNs, HANs) combined with sophisticated word embeddings and data balancing techniques proved highly effective for LJP, particularly for high-frequency legal categories.
- Multi-task learning approaches and the integration of additional explicit legal attributes or factual features were crucial for enhancing overall prediction accuracy and modeling inter-task dependencies, especially for prison term estimation.
- Predicting prison terms remains the most challenging subtask, consistently showing lower performance compared to article and charge prediction, highlighting the need for methods that incorporate a broader range of explicit legal and non-textual factors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.