<p>Aspect-based sentiment analysis task aims at predicting the sentiment polarity of a specific aspect in a sentence. Recent works have shown that attention-based and syntax-based approaches have gradually become mainstream methods. However, attention-based models may erroneously utilize unrelated context words as cues for prediction in sentences with long-range word dependency information. Besides, methods based on graph neural networks have been applied to model syntactic structure information, although great outcomes have been achieved, these methods are overly dependent on the precision of the syntactical dependency tree, which may lead to suboptimal dependencies between words and thus introduce noise. Effectively incorporating semantic relevance information and syntactic structure information remains a challenging task. To address the shortcomings referred to, we propose the dependency relationship-enhanced graph convolutional network (DREGCN) model, which utilizes a dual channel to integrate semantic relevance information and syntactic structure information. Specifically, in the syntactic channel, we preserve the original dependency tree to obtain global syntactic information, while introducing an aspect-oriented reconstruction tree to capture local syntactic information. Additionally, in contrast to previous studies where context words and aspect words were modeled separately, we propose cosine networks in the semantic channel to enhance information interaction between contexts and aspects. The experimental results show that our DREGCN model has a strong advantage on the three publicly available datasets.</p>