TextCNN-based Text Classification for E-government
Wu Suyan, Entong Su, Lei Binyang, Wu Jiangrui · 2019
The proposals of CPPCC members are divided into seven areas and more than 74 sub-directions according to their themes. To classify the proposals, that is to say, to determine the subject words is the primary task of handling the proposals. In this paper, TextCNN is used to pre-classify the subject words of proposals in order to assist in the final determination of the subject words manually. According to the text characteristics of the proposal, a professional dictionary and a special stop dictionary for uncommon words are first established. The text is trained as vectors through Word2dev and used as input of TextCNN neural network. High-level features are extracted by convolution layer, dimensionality reduction by pooling layer, feature integration by full connection layer and multiclassification by softmax layer. The open source deep learning framework TensorFlow is used to realize the base. The training and design of E-government text categorization algorithm based on in-depth learning, the structure of deep convolution neural network model and the optimization of model parameters are designed and saved. The experimental results show that the accuracy is improved by 3% compared with the relative entropy classification algorithm used in the previous system. At the same time, the training model is used to restore and predict the classification in the actual system, which reduces the computational pressure of the server.