The impact of attention mechanisms on aspect-based sentiment analysis performance using affective knowledge-enhanced GCNs

The objective of aspect-based sentiment analysis is to identify the sentiment polarity (positive, negative, or neutral) expressed toward a particular aspect in a sentence. Graph Convolutional Network-based models have been widely used in this field, typically relying on syntactic dependencies extracted from dependency trees. However, such approaches face limitations in modeling long-range dependencies and capturing the interactions between aspect terms and their contextual words. We propose an Attention-Enhanced Graph Convolutional Network (AE-GCN) that leverages affective knowledge to more accurately extract both semantic and sentiment-related dependencies. In this model, syntactic information is represented through a dependency-tree-based graph and enriched with affective knowledge. This structure is further enhanced with multi-layer attention to better capture semantic relationships. The attention mechanism improves the model’s capability to focus on aspect-relevant contextual words by assigning them adaptive weights, thereby improving sentiment prediction accuracy. The proposed model is evaluated on four widely used benchmark datasets from the restaurant and laptop review domains. Experimental results on the evaluated SemEval benchmark datasets show that AE-GCN achieves competitive performance relative to recent ABSA baselines. In our literature comparison, the model yields an average improvement of 0.44% in accuracy and 0.60% in F1-score over the next-best reported baselines. A reproduced robustness analysis further indicates that these gains remain stable across multiple data partitions within the evaluated benchmarks.