Text Sentiment Analysis Based on ResGCNN
Chengbo Liu, Jie Qi · 2019
Sentiment analysis of text is a significant task in Natural Language Processing (NLP), and Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) are two commonly used deep learning models of NLP. RNN's variant Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) can solve the long-term dependence of traditional RNN. Their variant Bi-directional LSTM (BiLSTM), Bi-directional GRU (BiGRU) solve the problem that unidirectional LSTM and GRU can only access the above but not the below. However, the ability of the cyclic neural network to extract key features is not strong. CNN can extract local features of text vectors effectively, but ignores the meaning of text contexts. TextCNN can extract text location features by using convolution cores of multiple different windows. As the number of layers increases, deep learning can obtain more complex features, but it will cause the problem of gradient disappearance, and errors cannot be effectively back-propagated, and the residual networks can solve it. This paper proposes a ResGCNN network combining residual network, multilayer BiGRU, and TextCNN. The excellent classification performance is obtained on multiple English and Chinese data sets.