Dual channel sentiment classification model based on grammar rules and multi attention

Abstract

As convolutional neural networks (CNN) and recurrent neural networks (RNN) have achieved excellent results in the field of Chinese text sentiment analysis. More and more researchers are extracting features of text based on the advantages of CNN and RNN in extracting features. However, the current scholars fail to make full use of sentiment language resources such as sentiment words, negatives and degree adverbs when they adopt deep learning methods. For implicit texts without explicit sentiment words, they cannot fully identify the differences between words and sentiment tendency. At the same time, it fails to consider the grammatical structure of the text, which leads to poor classification effect for some turning sentences or summary sentences. In addition, most of the models are mainly input in the form of word vector. For English, it is very convenient to segment words through spaces between words, but for Chinese, there may be inaccurate word segmentation, which will reduce the accuracy of classification. To solve these problems, a dual channel sentiment classification model based on grammar rules and multi attention (DCGA) is proposed. Firstly, the text with clear sentiment tendency is obtained according to the grammar rules, and the local features of the text are extracted by CNN channel. Considering that the grammar rules may ignore the context information, bi-directional long short term memory network (Bi-LSTM) channel is used to extract the global features containing the context information, and attention pooling is used to improve the sentiment information extracted by CNN channel. Then, the part of speech attention mechanism is used to mine the implicit sentiment features in the sentiment text to solve the poor classification effect caused by the existence of implicit text. Finally, the local features obtained by CNN channel and the global features obtained by RNN channel are fused for classification. Experimental results on four Chinese comment text datasets show that the proposed model outperforms most existing methods in accuracy.