Legal Core Element Recognition Based on XLNet with Correlation Matrix
Xiaolong Liu, Fuhui Sun, Xiaoyan Wang, Tanfeng Sun · 2025
Accurate identification of core legal elements helps improve the accuracy and efficiency of case judgments. However, the accuracy of existing deep learning methods is limited by case information complexity and typically cannot effectively extract contextual relationships from factual descriptions. The XLNet-Cor model integrates the Extra-Long Network (XLNet) with a Correlation Matrix, applying it to the identification of core legal elements. This XLNet-Cor model was trained and tested using the element recognition task dataset from CAIL2019, achieving optimal F1 scores of 74.64%, 75.71%, and 73.75% for divorce, labor, and loan cases respectively, improving by 18.0, 13.6, and 12.6 percentage points compared to traditional multi-label decision tree models. Experimental results demonstrate that this model performs well in multi-label binary classification tasks for legal core elements.