Automatically Classifying Chinese Judgment Documents Using Character-Level Convolutional Neural Networks

Abstract

Judgment is a decision by a court or other tribunal that resolves a controversy and determines the rights and obligations of the parties. Since the establishment of the China Judgments Online System, more and more judgment documents have been stored online. With the explosive growth of the number of Chinese judgment documents, the need for automated classification methods is getting increasingly urgent. For Chinese data sets, traditional word-level methods often bring extra errors in word segmentation. In this paper, we proposed an approach based on character-level convolutional neural networks to automatically classify Chinese judgment documents. Different from traditional machine learning methods, we hand over the work of feature detection to the model. Throughout the process, the only part that requires human labor is labeling the category of each original documents. In order to prevent overfitting when the amount of training data is not very large, we use a shallow model which has only one convolution layer. The proposed approach does well in achieving high classification accuracy based on 7923 pieces of Chinese judgment documents. In the meanwhile, the effectiveness of our model is satisfactory.