A knowledge-enhanced graph Transformer network (K-GTNet) was proposed in order to address the reasoning limitation of existing aspect-based sentiment analysis method caused by insufficient feature interaction and lack of knowledge utilization. A logical framework of “prior knowledge guidance–local syntactic modeling–global semantic reasoning” was constructed to achieve a deep understanding of complex textual sentiment. External sentiment knowledge and internal syntactic structure were fused to construct a graph representation with edge dependency weight. Neighborhood information was aggregated on this graph through a graph neural network, and local sentiment feature centered on aspect term was captured. The extracted local-aware feature sequence was fed into a Transformer. Global contextual dependency was built through self-attention mechanism, and robust aspect sentiment polarity classification was achieved. Experiments conducted on five benchmark datasets demonstrated that K-GTNet achieved average accuracy of 84.95% and F₁ of 77.96%, outperforming mainstream baseline model.