Paper Review - SAILER: A Structure-Aware Pre-Trained Model for Legal Case Retrieval(H Li, 2023)

2023-08-10 paper review

박지원 (Jee Won, Park)
박지원 (Jee Won, Park)
MSc Course
AI-based Crime Investigation System

This paper address the critical task of legal case retrieval, which is essential in intelligent legal systems.

While pre-training has succeeded in ad-hoc retrieval, effective strategies for legal case retrieval are still questioned. Legal case documents have intricate logical structures, but existing language models struggle with long-distance dependencies and key legal elements. To tackle these issues, this paper proposes SAILER, a Structure-Aware pre-trained Language Model for Legal Case Retrieval. SAILER optimizes retrieval by utilizing structural information, attending to key legal elements, and employing an asymmetric encoder-decoder architecture for pre-training. This model demonstrates strong discrimination capabilities without legal annotation data, distinguishing cases accurately. Experiments show significant improvements over state-of-the-art methods in legal case retrieval.

Challenges Addressed

  1. Long Document Structures: Legal case documents are lengthy with inherent logical structures, posing challenges for existing models in capturing long-distance dependencies.
  2. Relevance in Legal Domain: Relevance in legal case retrieval hinges on key legal elements, differentiating cases significantly. Existing models struggle with understanding these elements.




Experiments & Result Analysis

Compares with existing baseline models such as Traditional Retrieval Models, Generic Pre-trained Models, and Retrieval-oriented Pre-trained Models.

SAILER dataset


SAILER Results & Analysis

SAILER emphasizes legal terms than SEED(retrieval-oriented model)



  1. SAILER is the first model to leverage structural information in legal case pre-training.
  2. Proposing pre-training objectives that capture long-distance dependencies and logical knowledge.
  3. Extensive experiments demonstrating SAILER’s effectiveness across Chinese and English legal benchmarks.

In conclusion, “SAILER: A Structure-Aware Pre-Trained Model for Legal Case Retrieval” addresses challenges in legal case retrieval by introducing a novel framework. Through its innovative approach, SAILER significantly advances the field by emphasizing structural information and key legal elements, demonstrating its potential to enhance legal case retrieval in intelligent legal systems.