LegalCap: a model for complex case discrimination based on capsule neural network

In recent years, artificial intelligence, especially deep learning techniques, has provided an excellent theoretical and technical basis for the development of intelligent court. Given that case description usually contains multiple charges, charge prediction can be regarded as a task of multi-label text classification. We utilize the gated recurrent unit (GRU), a variant of recurrent neural network (RNN), to extract the chronological features of sequence. Due to the lack of retaining the positional relationship of entities when using RNN to extract features from text, a fraction of information will be lost in the context of multi-label classification. To address the drawbacks aforementioned, capsule neural network is utilized to extract the positional relation. Considering the basic and positional feature of instance, a composite model LegalCap, which combines capsule neural network with GRU, is proposed to model the complicated cases. Extensive experiments prove that the proposed model outperforms the selected baselines in the task of charge prediction. To overcome the unsatisfactory performance of directly conducting deep learning in charge prediction on the minority cases, we relieve the impact of training case imbalance by means of re-sampling. Experimental results demonstrate that after modifying the data distribution, the LegalCap model has a significant improvement on models’ bias on labels, i.e., better predictions on minority cases.