DCGCN-CPCT: Dual-Channel GCN Modules and CPCT for Aspect-Level Sentiment Classification

The task of Aspect-level Sentiment Classification (ALSC) is to predict the sentiment polarity of particular aspect words within a sentence. Researches conducted previously have demonstrated relatively promising experimental outcomes on using Graph Neural Networks to exploit sentence structure information on ALSC. But these methods do not do a good job of establishing the relations between the aspect word and the context well. In order to enhance the relationship between them, we propose DCGCN-CPCT, which mainly includes dual-channel Graph Convolutional Network (GCN) modules and Compound Point-Wise Convolution Transformation (CPCT). Specifically, we extract the syntactic information by constructing a dependency tree and transforming it into an adjacency matrix which is then employed in our GCN layer. In order to capture the semantic information, a self-attention matrix was leveraged to capture the interconnections among nodes, enabling the learning of features that the syntactic GCN module may fail to capture. Furthermore, we enhance the efficiency of feature concatenation from the dual-channel GCN modules through a Gate Mechanism. Finally, we leverage our CPCT layer to learn more complex features. Our method was evaluated on three public datasets. The experimental results show that our model DCGCN -CPCT outperforms the baseline models.