Abstract: In the existing aspect-level sentiment analysis methods, there are problems that the optimal text representation cannot be obtained, and the general graph convolutional network cannot be used to extract the deep structural information in the dependency graph. For these problems, this paper proposes a graph convolutional network (DDGCN) model based on deep BiLSTM and dense connections. Firstly, the deep semantic information between aspect words and context words is obtained by deep BiLSTM. Secondly, dense connections are added to the original graph convolutional network to generate a dense graph convolutional network that can extract the deep structural information. Then, the improved graph convolutional network is used to capture the structural information on the dependency graph. Finally, the text representation combining the two deep information is used for sentiment classification. Experimental results on three datasets show that DDGCN model improves both accuracy and F1 compared with the comparative model.

Key words: aspect-level sentiment analysis, graph convolutional network, deep BiLSTM, dense connection