Legal AI aims to provide improved knowledge management services based on legal documents. Existing legal judgment prediction datasets mainly use court verdicts. However, for prosecutors, the use of indictments for judgment predictions can help detecting inconsistencies between predictions and prosecution, providing prosecutors with more accurate references to laws and charges through error analysis. In this study, we collect a dataset called TWLJP, which contains 342,754 indictments. We compared three possible messaging passing architectures among the law, regulation, and accusation cause prediction tasks, i.e. independent, topological, and interactive. The result shows that interactive message passing among the three tasks achieved the best Macro-F1 performance of 95.2 %, 79.62 %, and 65.84 % for laws, regulations, and accusation cause prediction, respectively. We further improve the prediction of accusation cause from 8.8 % macro-F1 to 62.3 % for underperformed accusation causes via Prompt-Based Learning. Finally, in view of the situation where the charge prediction are written in various ways, we adopted a lenient approach to assess the accusation and improved the accusation performance to 77.2 %.

法律人工智能旨在提供基于法律文档的改进的知识管理服务。现有的法律判决预测数据集主要使用法院判决书。然而，对于检察官来说，利用起诉书进行判决预测，可以帮助发现预测与起诉之间的不一致之处，通过错误分析为检察官提供更准确的法律和指控参考。在本研究中，我们收集了一个名为 TWLJP 的数据集，其中包含 342,754 份起诉书。我们比较了法律、法规和指控原因预测任务中三种可能的消息传递架构，即独立、拓扑和交互。结果表明，三个任务中的交互式消息传递在法律、法规和指控原因预测方面取得了最佳 Macro-F1 性能，分别为 95.2%、79.62% 和 65.84%。对于表现不佳的指控原因，我们通过基于提示的学习进一步将指控原因的预测从 8.8% 的宏观 F1 提高到 62.3% 。最后，针对指控预测写法多种多样的情况，我们采用了宽松的方式来评估指控，将指控性能提升至77.2%。