In recent years, the issue of document-based extractive reading comprehension has been widely studied. Multi-hop reading comprehension requires obtaining supporting facts from multiple paragraphs in the document and reasoning to get the answer. But some problems have not been well solved: during the process of extracting the answer, it is often disturbed by the non-real answer (i.e. similar answer) in the document; the lack of a related expression mechanism for multiple scattered evidences in the paragraph affects the extraction of the answer. To address these issues, this paper proposes a multi-hop reading comprehension method based on Abstract Meaning Representation (AMR) and contrastive learning, which combines graph neural network (GNN) and answer verification for multi-task joint learning to reduce the interference of confusing information and enhance the robustness of the model. Experiments were conducted on datasets including HotpotQA and MuSiQue. The results show that the method proposed in this paper performs better than some competitive baselines in terms of exact match (EM) and F1 score evaluation.