Hierarchical-enhanced graph convolutional networks leveraging causal inference for aspect-based sentiment analysis

Abstract

Aspect-based sentiment analysis (ABSA) aims to determine the sentiment polarity of a particular aspect in a sentence. Existing research focuses on shortening the distance between opinion words and aspect words, resulting in spurious correlations. At the same time, the use of different dependent tools will bring different types of noise, destroying the effectiveness of the model. To address these issues, we propose a causal model of hierarchically augmented graph convolutional networks (CausalGCN). Specifically, we subdivide the language features into four relationships and then construct their corresponding mask matrices based on different relationships. At the same time, we introduce an instrumental variable to eliminate the confounders generated by the tool. Our model then combines the resulting mask matrix with localized attention at multiple levels. We treat the relationships between words and adjacent tensors as nodes and edges respectively, resulting in a multi-channel graph. Finally, we utilize graph convolutional networks to enhance relationship-aware node representations. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed model.