A Hybrid Semantics and Syntax-Based Graph Convolutional Network for Aspect-Level Sentiment Classification

Aspect-level sentiment classification seeks to ascertain the sentiment polarities of individual aspects within a sentence. Most existing research in this field focuses on individually assessing the importance of contexts on individual aspects, disregarding the negative impact of imbalanced relations between aspects due to their mutual influence. This paper presents a hybrid semantics and syntax-based graph convolutional network (SS-GCN) for aspect-level sentiment classification. This model addresses the imbalanced limitation by creating aspects-based balance relations between the strengths and weaknesses of different aspects through an auxiliary task. Furthermore, the multi-head self-attention mechanism utilizes position-enhanced encoding to identify the most relevant aspects of the current word. Extensive experiments demonstrate that SS-GCN outperforms other baselines in terms of classification performance. Compared to state-of-the-art methods, SS-GCN significantly improves 0.39–1.66% in accuracy and 0.43–1.92% in Macro-F1 on the SemEval 14-15 and MAMS datasets.