A new evidence extraction architecture called ATT-MRC improves the recognition of evidence entities in judgement documents by treating it as a question-answer problem, resulting in better performance than existing methods.

Transformer models like RoBERTa outperform traditional machine learning models in legal question answering tasks, achieving significant improvements in performance metrics such as F1-score and Mean Reciprocal Rank.

The paper proposes L-HetGRL, an unsupervised approach using a legal heterogeneous graph and incorporating legal domain-specific knowledge, to improve Legal Document Similarity Measurement (LDSM) with superior performance compared to other methods.