Aspect-based sentiment analysis (ABSA) consists of two fundamental subtasks: aspect term extraction (AE) and aspect-level sentiment classification (ASC). Existing research inevitably leads to severe error propagation, and the model cannot effectively capture the relationships between words. To tackle these problems, in this paper, we propose a novel joint learning model: MRC-GCN: Machine Reading Comprehension-Graph Convolutional Network. Firstly, aspects are extracted through MRC combined with pre-training models. In order to enhance the feature representation of the query, we utilize Long Short-Term Memory (LSTM) combined with self-attention mechanism and concatenated the original sentence and query using certain rules before inputting them into the model. Subsequently, due to better capture and aggregation of words and the relationship between words being effective for ASC, we combine the multiple latent information graph structures of sentences to form a text graph and use GCN with an attention mechanism to achieve sentiment classification. Experimental results on three public datasets show that our model is more effective than the previous models, demonstrating that our method can significantly empower ABSA models. In addition, we have designed ablation experiments for specific modules to verify their effectiveness.