ROAEN: Reversed dependency graph and Orthogonal-gating strategy Attention-Enhanced Network for Aspect-Level Sentiment Classification

Abstract Aspect-Level sentiment classification aims to determine the sentiment polarity of different aspects within a sentence. Many existing approaches employ Graph Convolutional Networks (GCN) in conjunction with dependency trees to model syntactic structures in sentences but often ignore its limitations in modeling long-distance words. This paper proposes a Reversed dependency graph and Orthogonal-gating strategy Attention-Enhanced Network (ROAEN), which introduces a novel Reversed dependency Graph Convolutional Network combined with Orthogonal projection (RGCN-O) to model long-distance words and syntactic structures simultaneously. Specifically, RGCN-O discards a few words with dependencies, generates vector representations for long-distance words, and models syntactic structures from its orthogonal space. The output of RGCN-O is then used to design a gating strategy that dynamically filters attention noise in a weighted manner. Additionally, we offer a novel total inter-intra loss function to enhance the compactness and distinctiveness of the intra-class and inter-class sample features. Extensive experiments on five benchmark datasets demonstrate the effectiveness of RGCN-O and the new loss function, with ROAEN achieving state-of-the-art performance.