An aspect sentiment analysis model based on adversarial training and multi-attention

Abstract

Aiming at the disadvantages of the gradient vanishing and exploding of the Recurrent Neural Network in the traditional deep learning algorithm and the problem that the Convolutional Neural Network cannot obtain the global features of the classified text, a CNN(Convolutional Neural Network)-BiLSTM (Bidirectional Long Short-Term Memory) sentiment analysis method based on adversarial training and multi-layer attention is proposed to give full play to the ability of CNN to extract phrase-level features of text and the ability of BiLSTM to extract global structural information of text, and the multi-layer attention mechanism will assign higher weights to keywords, and the adversarial training can well solve the model instability problem of the current deep learning model. Using the public data set Laptop reviews and Restaurant Reviews from SemEval 2014 for verification, the results show that the accuracy of the model proposed in this paper is 1 and 1.9 percentage points higher than that of the original model on the two data sets. In contrast, the model is more efficient in aspect-level sentiment classification tasks.