基于Ernie胶囊网络的多标签依赖法条推荐

Multi-Label Dependency Statute Recommendation Based on Ernie Capsule Network

摘要: 针对法律文书中存在多标签法条预测的问题，为了提高法条推荐任务的准确性，本文提出了一种基于标签依赖关系的法条推荐模型。该模型利用Ernie中的词感知预训练任务、句子感知预训练任务、语义感知预训练任务增强文本语义表示能力，利用胶囊网络将低层胶囊的特征传递到高层胶囊，利用动态路由更新获取层次丰富的语义特征，再将文本向量输入到BiLSTM获取上下文特征信息。最后，在法条标签预测层，通过加权求和每一个时间步的法条标签向量，捕获法条标签之间的依赖关系，解决多标签法条推荐问题。实验结果证明，在CAIL公开数据集上，本文取得了82.3%的F1值，有效地提高了法条推荐的准确性。

Abstract: In order to improve the accuracy of law recommendation task, this paper proposes a law recommendation model based on label dependency. The model uses the word perception pre-training task, sentence perception pre-training task and semantic perception pre-training task in Ernie to enhance the semantic representation ability of text. Capsule network is used to transfer the features of low-level capsules to high-level capsules, dynamic routing update is used to obtain hierarchical semantic features, and text vector is input into BiLSTM to obtain contextual feature information. Finally, in the law label prediction layer, the dependence relationship between law labels is captured by weighted summing the law label vector of each time step, and the multi-label law recommendation problem is solved. The experimental results show that 82.3% F1 value is obtained on the CAIL open data set, which effectively improves the accuracy of the recommendation.