Scholarlink

<p>In large-scale judicial scenarios, document-level legal relation extraction plays a pivotal role in enabling efficient case text analysis and structured legal information organization, thereby facilitating downstream tasks such as case recommendation and judgment prediction. To address the problems including legal document encoding, imbalanced relation distribution, and noise in training dataset, the Document-Level Legal Relation Extraction (DLRE) model is proposed to enhance the performance in semantic understanding and relation extraction of legal documents. Firstly, Lawformer and BiGRU are integrated for document encoding in DLRE model, effectively leveraging a reconstructed training corpus to learn hierarchical textual patterns. Furthermore, the prior knowledge derived from the training samples is incorporated, and a label distribution-sensitive margin loss is introduced in DLRE to address the problem of data imbalance to improve the model’s ability to distinguish between the types of minority and majority relations. Finally, to reduce noise interference and improve generalization, a dual self-knowledge distillation framework is introduced in the DLRE model with: (1) feature distillation, which ensures semantic consistency across model layers, and (2) target distillation, which leverages soft labels from a teacher model to guide the student model. Experimental results on the legal datasets show that the proposed DLRE model outperforms baselines in extracting legal relations, indicating its effectiveness and application potential in legal document analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Document-level legal relation extraction based on domain feature and dual self-distillation

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Yuming Wang,
  • Guangsheng Lin,
  • Kanghua Zhang,
  • Yuxin He,
  • Jingpei Dan,
  • Yingfei Wang

摘要

In large-scale judicial scenarios, document-level legal relation extraction plays a pivotal role in enabling efficient case text analysis and structured legal information organization, thereby facilitating downstream tasks such as case recommendation and judgment prediction. To address the problems including legal document encoding, imbalanced relation distribution, and noise in training dataset, the Document-Level Legal Relation Extraction (DLRE) model is proposed to enhance the performance in semantic understanding and relation extraction of legal documents. Firstly, Lawformer and BiGRU are integrated for document encoding in DLRE model, effectively leveraging a reconstructed training corpus to learn hierarchical textual patterns. Furthermore, the prior knowledge derived from the training samples is incorporated, and a label distribution-sensitive margin loss is introduced in DLRE to address the problem of data imbalance to improve the model’s ability to distinguish between the types of minority and majority relations. Finally, to reduce noise interference and improve generalization, a dual self-knowledge distillation framework is introduced in the DLRE model with: (1) feature distillation, which ensures semantic consistency across model layers, and (2) target distillation, which leverages soft labels from a teacher model to guide the student model. Experimental results on the legal datasets show that the proposed DLRE model outperforms baselines in extracting legal relations, indicating its effectiveness and application potential in legal document analysis.