Multi-label charge predictions leveraging label co-occurrence in imbalanced data scenario

Charge prediction is to predict associated charges based on fact descriptions and plays a significant role in legal aid systems. It is a fundamental and challenging task to automatically predict charges in the multi-label classification paradigm, which is fit to real applications. Existing works either focus on balanced data scenario and multiple charges or few-shot charges with a single label. Moreover, previous models utilize special initialization with label patterns to improve the performance of the multi-label classification task, which is only applicable when there is less training data, resulting in poor robustness. To this end, a multi-task convolutional neural network combined with bidirectional long short-time memory leveraging label co-occurrence framework, called CBLLC, is introduced to predict multiple charges with article information on imbalanced data occasion. We develop a new learning mechanism to train the framework of charge and article patterns when there is a lot of training data, increasing its robustness. In CBLLC, the data preprocessing process serves to aid the training in a more generalized manner and reduce overfitting. A salient word annotation is introduced to deal with few-shot charges. A better classification result is obtained with processed data and improves the generality of the model. Experimental results of Chinese AI and Law Challenge test set show the superiority of our proposed method compared with the state-of-the-art methods. In particular, a macro-F1 score of 92.9% for charges and 86.6% for articles is achieved with co-occurrence of charges and patterns of articles.