Graph Reasoning With Supervised Contrastive Learning for Legal Judgment Prediction.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Legal judgment prediction (LJP) involves determining law articles, charges, and penalties from case facts.
• Existing LJP models often overlook label dependencies across different prediction tasks.
• Effectively utilizing relational information among tasks and labels is crucial for LJP.

Purpose of the Study:

• To propose a novel framework, GraSCL, that integrates graph reasoning and supervised contrastive learning (SCL) for LJP.
• To address the limitations of existing methods by explicitly modeling label dependencies.
• To enhance the accuracy and efficiency of legal judgment prediction.

Main Methods:

• Transformed LJP into a node classification problem within a graph reasoning framework.
• Designed a graph reasoning network to capture dependency structures and relational learning.
• Extended supervised contrastive learning (SCL) to the node level for efficient training.
• Incorporated online hard negative mining (OHNM) to optimize SCL by focusing on challenging negative samples.

Main Results:

• The proposed GraSCL framework demonstrated significant improvements in LJP tasks.
• Graph reasoning effectively modeled inter-task and inter-label dependencies.
• Node-level SCL and OHNM enhanced model performance, especially with small batches.
• Experimental results on benchmark datasets validated the effectiveness against state-of-the-art methods.

Conclusions:

• GraSCL offers a powerful approach for legal judgment prediction by leveraging graph reasoning and contrastive learning.
• Modeling label dependencies is key to advancing LJP accuracy.
• The integration of SCL and OHNM provides an efficient and effective training strategy for complex prediction tasks.