Social Media Sentiment Analysis Based on Dependency Graph and Co-occurrence Graph

In recent years, research of social text sentiment analysis has progressed rapidly, but the existing methods usually use the single feature for text representation and fail to make full use of potential features in social texts. Such sparse feature limits the improvement of sentiment analysis performance. Intuitively, besides the plain text feature, the features that reveal grammatical rules and semantic associations all have a positive effect on the performance of sentiment analysis. This article takes diverse structural information, part of speech, and position association information into consideration simultaneously, and proposes a brain-inspired multi-feature hierarchical graph attention model (MH-GAT) based on co-occurrence and syntactic dependency graphs for sentiment analysis. It mainly includes multi-feature fusion and bi-graph hierarchical attention. Specifically, we first design an input layer involving multiple features, such as part of speech, position, syntactic dependency, and co-occurrence information, to make up for the information lacking in conventional sentiment analysis methods. As for the bi-graph hierarchical attention mechanism, we build hierarchical graphs for each text and use a graph attention network with extraordinary aggregation ability to learn the inherent rules of language expression. Compared to the latest Att-BLSTM, Text-Level-GNN, and TextING, the sentiment analysis accuracy of the proposed model has increased by an average of 5.17% on the Chinese Weibo dataset and English SST2 dataset. The proposed MH-GAT model can effectively improve the classification performance of social short texts.