An Aspect-Centralized Graph Convolutional Network for Aspect-Based Sentiment Classification

Abstract

Recent works on aspect-based sentiment classification have manifested the great effectiveness of modeling syntactic dependency with graph neural networks (GNN). However, these works ignore the fact of sentiment information decreasing over dependency paths due to the complex syntactic structure. To tackle the above limitation, we explore a novel solution of constructing an Aspect-Centralized Graph (ACG) for each aspect. Specifically, we directly link all words in a sentence to the aspect word and create a more effective way for the interaction between aspects and opinion words. Based on it, we also incorporate syntactic information into the new graph. To achieve this, we substitute edges of ACG with weighed values, which are calculated from the syntactic relative distance between the aspect and context words on the original dependency graph. Then we propose an Aspect Centralized Graph Convolutional Network (ACGCN) to extract aspect-specific features and effectively interact them with context representations. Extensive experiments on five benchmark datasets show that our model achieves better performance over most baseline models and extensively boosts the performance with BERT.