<p>E-commerce sites have become valuable sources of information for both customers and businesses. To gain a more profound knowledge of the public’s perceptions of some entities, sentiment analysis seeks to identify the sentiments of the reviews. Aspect-based sentiment analysis is a detailed approach that aims to detect sentiments towards specific aspects of an entity expressed in reviews. Recently, applying graph neural networks to dependency trees has become a prevalent trend for aspect-based sentiment analysis. However, these methods often rely solely on a single type of linguistic representation, ignoring deeper phrase-level dependencies. Thus, in this paper, we design a dual-graph model (Bissa-DGCN) that integrates word-level dependency representations from the dependency graph, phrase-level representations from the constituency graph, and the strength of associated contextual affective words to provide richer representations. Two fusion strategies are introduced to efficiently combine these representations: parallel fusion, which employs an attention-based weighting mechanism to assign global importance to each graph and adaptive fusion, which employs feature-wise integration to balance dependency-based and constituency-based information at each hidden dimension. The proposed dual-graph architecture is efficient and supports scalable sentiment analysis. Experiments conducted on benchmark datasets, SemEval restaurant 2014, restaurant 2015, restaurant 2016, and laptop 2014, demonstrate that the proposed model with adaptive fusion achieves consistently strong and competitive performance compared to existing graph-based baselines.</p>