Study of Deep Learning-Based Legal Judgment Prediction in Internet of Things Era

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Law

Background:

• Legal judgment prediction is a key AI application in the judicial field.
• Algorithm performance is often limited by computing resources, hindering practical deployment.
• Reducing model resource consumption and increasing inference speed are crucial for legal AI.

Purpose of the Study:

• To enhance prediction accuracy for legal decisions.
• To improve the inference speed of legal judgment prediction models.
• To reduce the memory footprint of AI models in the judicial domain.

Main Methods:

• Proposed KD-BERT, a BERT knowledge distillation-based model for legal decision prediction.
• Utilized a BERT pretraining model with lower memory requirements as the encoder.
• Employed knowledge distillation to transfer knowledge to a shallower transformer student model.

Main Results:

• KD-BERT achieved the highest F1-score compared to traditional BERT models.
• The proposed KD-BERT demonstrated significantly faster inference speeds than other BERT models.
• The model effectively reduced computational resource consumption and memory usage.

Conclusions:

• KD-BERT offers a more efficient and accurate solution for legal judgment prediction.
• The knowledge distillation approach successfully balances performance and resource efficiency.
• This model facilitates easier deployment of AI in practical judicial settings.