Enhancing Aspect-Based Sentiment Classification with Local Semantic Information

Abstract

AbstractThe research goal of aspect-level sentiment analysis is to analyze the sentiment polarity expressed by a given sentence according to its specific aspect. The Graph Convolutional Network (GCN) model based on attention mechanism performs better among the existing solutions. This kind of method uses syntactic dependency information and semantic information to adjust and optimize attention. Still, the experimental results show that this kind of model does not perform well in complex sentences. We consider that attention mechanism is good at capturing global feature information, while Convolutional Neural Networks can effectively utilize local features. In this work, we study how to use attention mechanism and Convolutional Neural Networks to extract sentiment features better and propose a Graph Convolutional Network Model (DWGCN) of Depthwise separable convolution. Moreover, the basic design idea is to obtain syntactic dependency information utilizing graph convolution network learning and use the corresponding attention mechanisms to interact syntactic features with contextual information about word order to get semantic information. Then, we extract the emotional statement about the given Aspect-term from the semantic information. Therefore, we design a Feature extraction module based on a Depthwise separable convolution network and GLU (FMDG). To verify the effectiveness of the model, we test it on five benchmark datasets. The experimental results show that the proposed model outperforms the current relevant work in classification accuracy and generalization ability.KeywordsAspect-level sentiment analysisGraph Convolutional NetworkDepthwise separable convolution network