The study introduces advanced methods for rhetorical role labeling (RRL) in legal documents, crucial for applications such as case summarization and semantic search. By leveraging semantically similar instances, the researchers propose inference-based and training-based approaches to significantly enhance the effectiveness of RRL.

Key Highlights:

• Develops novel inference and training methodologies based on semantically similar instances.
• Integrates prototypical learning with a discourse-aware contrastive method.
• Demonstrates cross-domain applicability and substantial improvement in macro-F1 scores.

Importance of This Work: This research provides a significant foundation for improving semantic search in legal domains, potentially revolutionizing how information is processed and retrieved in legal systems. It offers a promising avenue for further innovations in the application of AI in law, emphasizing adaptability and precision.