A Dual-channel Text Classification Model based on an Interactive Attention Mechanism

Aiming at the problem that convolutional neural network(CNN) focuses on local features and lacks the ability of text context feature extraction, In this paper, we propose a dual-channel text classification model based on Interactive Attention Mechanism(IAM). The model uses skip-gram to embed words into dense low latitude vectors and obtains the text embedding matrix, which is input into the Gate Convolution Neural Network(GCNN) and Multi-Head Attention(MHA) at the same time, and then after Pointwise Convolution(PC), the features obtained from the feature extraction layer in the two channels are calculated by an IAM, and finally, the features are fused. Compared with CNN, LSTM, and other improved models, the classification effect of this hybrid model is improved.