Fusion of syntactic enhancement and semantic enhancement for aspect-based sentiment analysis

The graph neural networks mainly focus on the syntactic structure when they are used to model the syntactic dependency tree of a sentence for aspect-based sentiment analysis (ABSA). However, these methods neglect the study of dependency tree itself and semantic information, and syntactic information in unstructured text can bring additional noise. To address these issues, a graph convolutional network (GCN) model integrating syntactic and semantic enhancements was proposed. Attention mechanism was used to extract aspect information and global information, and enhance semantic information through gating mechanism. Simultaneously, dependency tree was reshaped to increase the weight of opinion words and enhance syntactic information by utilizing external emotional knowledge and the distance features of words. Finally, dynamic feature fusion was performed on semantic and syntactic information. Extensive experiments were conducted on three public datasets. The experiment results show that the accuracy and macro-F1 values are better than the compared models, which indicate the effectiveness of the fusion of syntactic enhancement and semantic enhancement for ABSA.