Aspect-based sentiment analysis via dual residual networks with sentiment knowledge

Sentiment analysis has recently garnered significant interest in natural language processing, especially in aspect-based sentiment classification, which focuses on identifying and analyzing sentiments associated with specific aspects of a sentence. Previous approaches primarily utilizing graph convolutional networks (GCN) to construct syntactic or semantic graphs have shown promising results in sentiment polarity evaluation. However, these approaches typically rely on two-layer graph structures to aggregate neighboring nodes, leading to insufficient extraction of sentiment information and limited exploration of sentiment knowledge. To overcome these limitations, we introduce a novel syntax–semantic dual GCN (SSD-GCN) for aspect-based sentiment analysis. Our method enhances the dependency graph between sentiment words and aspect terms by integrating sentiment knowledge with syntactic dependency trees. Additionally, it incorporates a semantic relationship graph to further strengthen the model. To more effectively aggregate information from neighboring nodes, we designed a dual residual network that preserves original features while thoroughly capturing sentiment characteristics within the context. Our systematic evaluations on five widely used datasets, namely Laptop, Restaurant14, Twitter, Restaurant15, and Restaurant16, demonstrate the strong competitiveness of our model, with performance scores of 82.56%, 87.75%, 77.35%, 86.01%, and 91.60%, respectively. This comprehensive analysis confirms the effectiveness of our approach. The code can refer to https://github.com/ZCMR/SSD–GCN