Aspect-Level Sentiment Analysis by Fusing Local Information with Graph Attention Networks

Aspect-based sentiment analysis is a fine-grained sentiment analysis task that requires detecting the sentiment polarity of the target aspect. To address the problems that current syntactic research focuses too much on syntactic dependencies of sentences, without noticing the correlation between sentiment polarity of target aspects and local context, and ignoring important features of local context, a relational graph attention network model incorporating local semantic information is designed. The model uses syntactic relative distance to determine the extent of local context and generates local context hidden features, followed by a graph neural network to generate syntax-aware contextual features, and combines both features for sentiment analysis. Experiments of the model on four benchmark datasets show that the proposed network model has better classification performance when compared with nine advanced baseline models. The improved accuracy of 1.43% and F1 value of 1.81% on the Restaurant dataset, and the improved accuracy and F1 value of 1.50% and 1.22% on the Laptop dataset, respectively.