Scholarlink

<p>Civil Case Judgment Prediction (CCJP) aims to automatically determine the extent of court support for a plaintiff’s pleas, constituting a pivotal task in the field of Legal Artificial Intelligence (LegalAI). Civil litigation often involves intricate factual determinations, multifaceted legal relationships, and substantial judicial discretion. Current models, however, typically treat complex civil cases as linear text, which leads to suboptimal modeling of logical dependencies and a lack of interpretability in the reasoning process. To address these limitations, we propose SJR-GN, a Graph Neural Network framework grounded in Structured Judicial Reasoning. We designed a Dual-level Structured Reasoning Mechanism: first, global topological reasoning, which utilizes a semantic-aware Graph Attention Network (GAT) to perform information fusion on the adjacency matrix of the merged heterogeneous graph, capturing long-range dependencies; second, plea-guided focused reasoning, which employs a cross-attention mechanism to dynamically aggregate information pertinent to the plaintiff’s pleas. Experimental results on a real-world “private lending dispute” dataset demonstrate that SJR-GN outperforms robust sequence-based baselines. Moreover, attention visualizations provide a reasonable interpretability interface, validating that the transition from sequence modeling to graph-structured reasoning effectively enhances both the predictive performance and interpretability in complex civil case outcomes.</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

Simulating judicial reasoning: a plea-guided explainable graph neural network for civil judgment prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月25日

查看全文
  • Xiang Lu,
  • Danyang Xu,
  • Juan Shan,
  • Xiaopeng Zheng,
  • Tianyue Huang

摘要

Civil Case Judgment Prediction (CCJP) aims to automatically determine the extent of court support for a plaintiff’s pleas, constituting a pivotal task in the field of Legal Artificial Intelligence (LegalAI). Civil litigation often involves intricate factual determinations, multifaceted legal relationships, and substantial judicial discretion. Current models, however, typically treat complex civil cases as linear text, which leads to suboptimal modeling of logical dependencies and a lack of interpretability in the reasoning process. To address these limitations, we propose SJR-GN, a Graph Neural Network framework grounded in Structured Judicial Reasoning. We designed a Dual-level Structured Reasoning Mechanism: first, global topological reasoning, which utilizes a semantic-aware Graph Attention Network (GAT) to perform information fusion on the adjacency matrix of the merged heterogeneous graph, capturing long-range dependencies; second, plea-guided focused reasoning, which employs a cross-attention mechanism to dynamically aggregate information pertinent to the plaintiff’s pleas. Experimental results on a real-world “private lending dispute” dataset demonstrate that SJR-GN outperforms robust sequence-based baselines. Moreover, attention visualizations provide a reasonable interpretability interface, validating that the transition from sequence modeling to graph-structured reasoning effectively enhances both the predictive performance and interpretability in complex civil case outcomes.