Text Classification Based on BertRCNNATT Hybrid Model

The Recurrent Convolutional Neural Network (RCNN) has achieved excellent performance in text classification. However the RCNN ignores the context of the information and the use of the max-pooling will causes the loss of feature information. Therefore, We proposed a text classification method based on the BertRCNNATT hybrid model to solve these problems. First, the BERT pre-training is used to generate semantic vectors rich in context information. Then, we use the RCNN and the max-pooling attention that considers the weight distribution problem for further classification. Our model has an accuracy rate of 94.656% on the SST-2 data set, which outperforms the state-of-the-art classification model.