Legal Information Retrieval and Entailment Using Transformer-based Approaches.

Area of Science:

• Legal informatics
• Artificial intelligence
• Natural language processing

Background:

• Information overload in the legal domain presents significant challenges.
• The COLIEE competition addresses these challenges through tasks like case law and statute law retrieval and entailment.
• Developing effective systems for legal information processing is crucial.

Purpose of the Study:

• To describe methods for Task 1 (case law retrieval) and Task 4 (statute law entailment) in the COLIEE 2023 competition.
• To evaluate the performance of different natural language processing models for legal tasks.

Main Methods:

• For Task 1, a sentence-transformer model generated numerical representations of case paragraphs, and similarity histograms were used to build a binary classifier.
• For Task 4, a pre-trained DeBERTa large language model (LLM) was fine-tuned on SNLI and MultiNLI datasets.
• Performance was compared against knowledge distillation and ensemble methods (Random Forest, Voting).

Main Results:

• The sentence-transformer approach provided a method for case law retrieval.
• The DeBERTa LLM-based method for Task 4 achieved third place among eight teams in the COLIEE 2023 competition.
• Comparative analysis of DeBERTa against other models was conducted.

Conclusions:

• The developed methods show promise in addressing legal information overload.
• Fine-tuning LLMs like DeBERTa is effective for legal entailment tasks.
• Further research can build upon these approaches for enhanced legal information systems.