Sentiment Analysis Of Online Users Reviews Using Graph Convolutional Network

For sentiment analysis of user review,models based on traditional machine learning have been unable to meet people's requirements for accuracy of results.In deep learning-based methods,these sequential-based learning models often pay more attention on the local semantic information of the text,thus ignoring the influence of text structure information on the representation of emotional features,such as transition word and syntactic information.In the related research of neural network,compared with sequential-based learning models,the graph-based learning models can capture the structural information of data.As one of the representatives of graph learning models,graph convolutional network has achieved fine results on graph tasks.Using text graph to represent the text,the graph convolutional network is applied to the text graph,which can filter node features that is not related to the central node,and retain the global sentiment information of the text,thereby improving the accuracy of the model on sentiment analysis tasks.The main research contents as follows:(1)Attention mechanism based on CNN and RNN models usually perform unsatisfactory because of ignoring the connection between grammatical related words.In this paper,we propose a graph convolutional network named SCGCN for sentiment analysis task to explore the information hidden in the sentence parsing trees.Firstly,the model obtains the dependency relationship between words among syntactic dependency trees,and utilizes the bidirectional long short-term memory to obtain the sentence representation from text.Secondly,following the dependency relationship between words,we design the graph convolution network to encode the sentence representation to get the node representation,where we combine the attention mechanism to redistribute the sentiment weights of the sentence representation.Finally,the output representation generated is fed into a softmax layer to calculate the probability scores of sentiment polarity.The experimental results show that the SCGCN model has high accuracy and fine generalization performance when dealing with the two classification and multi classification annotation tasks of user comments.(2)The sequential-based learning models include CNN and RNN usually have poor performance in sentiment analysis task because of ignoring discontinuities and long-distance semantic information of the text.Besides,when using static vectors to represent polysemous words,the context of the words will be ignored.In this paper,through introducing the pre-training language model BERT,we propose a sequential graph based graph convolutional network model named SGCNSA for sentiment analysis task to explore the information hidden in the sentence structure.Firstly,SGCNSA uses the BERT to segment the text and generate dynamic word representation and sentence embedding.Secondly,a text graph containing text structure information is contructed by setting the co-occurrence window to traverse the sentences.Moreover,utilizing the bidirectional long short-term memory from word representation to obtain the sentence representation.Following the text structure information,we design a graph convolutional network model to encode sentence representation to get the node representation and concat it with sentence embedding to generate the output representation.Finally,the output representation is fed into a softmax layer to calculate the probability scores of sentiment polarity.The experimental results show that the accuracy of SGCNSA is better than other contrast methods.