<p>Aspect-based sentiment analysis (ABSA) focuses on accurately identifying the sentiment polarity of specific aspect terms in a text, offering critical support for a deeper understanding of user opinions. However, current ABSA models face two major challenges: (1) They often rely solely on syntactic dependency trees, making it difficult to fully capture both the semantic and syntactic structures of sentences. (2) They struggle to effectively recognize sentiment expressions in complex sentences, which limits overall classification performance. To address these challenges, this paper proposes a Semantic-Guided Fusion Network with Abstract Meaning Representation (AMR) and Dependency Trees (SGFNet) model for the ABSA. To overcome the first challenge, this paper constructs a multi-level semantic-syntactic fusion graph that integrates global semantic paths from AMR with local structural features from syntactic dependency trees. This design enables joint modeling of semantic and syntactic information, leading to more comprehensive sentence representations. To address the second challenge, this paper introduces a Semantically-Guided Relation Enhancement mechanism to enhance the expressive power of sentiment features and accurately extract aspect-related sentiment cues, thereby improving sentiment classification in complex sentences. Extensive experiments on four public datasets demonstrate that the proposed SGFNet consistently achieves better accuracy and F1 scores than competitive baseline models.</p>