<p>Hyper-parameters settings in our model.</p>

Legal Judgment Prediction (LJP) is a core task in Legal AI systems, which aims to predict law articles, charges, and term-of-penalty from case facts. While existing deep-learning-based LJP approaches for civil law systems have achieved certain progress, they still suffer from two key limitations: (1) insufficient deep understanding and effective utilization of external judicial knowledge; and (2) the lack of effective strategies to filter out erroneous dependency information in multi-task LJP frameworks. To address these challenges, we propose a legal judgment prediction model based on knowledge fusion and dependency masking. Specifically, we first integrate a CNN-based local semantic refinement component into the existing BERT-based legal knowledge extraction method, thereby enabling the model to further extract the core knowledge embedded in judicial documents. Then, we introduce differential attention to reduce noise in conventional attention fusion methods and help the model locate key information in case facts more accurately. Furthermore, we propose a multi-task dependency information masking mechanism to accurately identify and filter erroneous dependency information for multi-task LJP methods. Experiments conducted on real-world datasets demonstrate the superiority of our proposed model. This code is available online at https://github.com/PaperCode-GNU/KFTM.

法律判决预测（Legal Judgment Prediction, LJP）是法律人工智能系统中的核心任务，旨在从案件事实中预测所适用的法条、罪名与刑罚幅度。现有基于深度学习的大陆法系法律判决预测方法虽已取得一定进展，但仍存在两大核心局限：其一，对外部司法知识的深度理解与有效利用不足；其二，多任务法律判决预测框架中缺乏有效策略以过滤错误依赖信息。为应对上述挑战，本文提出一种基于知识融合与依赖掩码的法律判决预测模型。具体而言，本文首先将基于卷积神经网络（Convolutional Neural Network, CNN）的局部语义细化模块融入现有基于BERT的法律知识抽取方法，使模型能够进一步挖掘司法文档中蕴含的核心知识。随后，本文引入差分注意力机制以降低传统注意力融合方法中的噪声，助力模型更精准地定位案件事实中的关键信息。此外，本文还提出一种多任务依赖信息掩码机制，可针对多任务法律判决预测方法精准识别并过滤错误依赖信息。基于真实世界数据集开展的实验验证了本文所提模型的优越性。相关代码已公开于https://github.com/PaperCode-GNU/KFTM。