An Improved Text Classification Model Based on Memory Convolution Neural Network

This paper proposes a text classification model, called improved memory neural network model, which is used to process large-scale training data. In this model, the optimized transformer feature extractor is used to replace the memory neural network which can not be parallelized. At the same time, the multi-level void convolution matrix is designed to replace the convolution neural network, so as to extract more accurate semantic unit features. Finally, in order to reduce the model parameters, each level of the convolution network pooling layer and the full connection layer are eliminated, but the global average pooling layer is instead used. The experimental results on THUCNews dataset and Twitter dataset show that the proposed method achieves competitive results in the accuracy, model parameters and convergence rate.