<p>Aspect-level sentiment classification (ASC) predicts sentiment polarities for specific aspect terms in written text. Although incorporating syntax information significantly enhances ASC, two primary issues remain overlooked. First, many models focus on identifying aspect-related local contexts by analyzing the syntactical relationship between aspect terms and context words. However, they neglect the interdependencies between various contexts, limiting the comprehensiveness and depth of aspect-related sentiment representations. Second, variations in sentence syntax quality pose challenges to dynamic adaptation. To address these issues, we initially decompose syntax trees into two sub-trees, facilitating the precise identification of aspect-related local contexts and the differentiation of their roles. We introduce the fusion network with syntax-based learning and semantics-based assistance (FN-SLSA). This model exploits a wider range of syntax information to learn aspect-related representations. Based on these representations, it further constructs a syntax-based sentiment representation. This representation is enriched with deeper and more comprehensive semantics derived from the interdependencies between these representations. Additionally, the model learns a semantics-based sentiment representation that complements the syntax-based sentiment representation and adapts to variations in syntax tree quality. Experimental results on four benchmark datasets demonstrate that the FN-SLSA model outperforms a range of state-of-the-art models. Moreover, the ablation experiments and case study verify the effectiveness of important components in this model.</p>