Aspect-level sentiment analysis aims at detecting the sentiment polarity of a given aspectual term from a fine-grained perspective. Existing approaches, by modeling the equality of different dependencies of syntactic information, ignore the weight adjustment of different dependencies by the type of relations in the convolutional network and the prediction of aspect term sentiment from a semantic perspective by the global information. In order to solve the above problems, we proposed a WRCN-CL model with two tasks: weighted relational convolutional network (WRCN) and complementary learning (CL) as an auxiliary task; WRCN adjusts the syntactic dependency tree structure based on the positional weights and the relationship types in the semantic structure; and CL searches for aspect-related semantic information from a global perspective to enhance knowledge. The experimental results show that WRCN-CL has significant improvement in Macro-F1 metrics compared to other baseline models, thus demonstrating that the proposed model can effectively solve the classification task of aspect-level sentiment analysis.