Law text classification using semi-supervised convolutional neural networks

With the developments of internet technologies, dealing with a mass of law cases urgently and assigning classification cases automatically are the most basic and critical steps. Convolutional Neural Networks (CNNs), has been shown to be effective for text classification. To better apply CNNs into law text classification, this paper presents a new semi-supervised Convolutional Neural Networks (SSC) framework. Our method combines unlabeled data with a small labeled training set to train better models, and then integrates into a supervised CNN. More specifically, for effective use of word order for text categorization, we use the feature of not low-dimensional word vectors but high-dimensional text data, that is, a small text regions is learned based on sequences of one-hot vectors. To better improve the prediction accuracy of the scheme, we seek effective use of unlabeled data for text categorization for integration into a supervised CNN. We compare the proposed scheme to state-of-the-art methods by the real datasets. The results demonstrate that the semi-supervised learning model can get best text classification accuracy.