To inhibit the spread of rumorous information, fact checking aims at retrieving relevant evidence to verify the veracity of a given claim. Previous work on fact checking typically uses knowledge graphs (KGs) as external repositories and develop reasoning methods to retrieve evidence from KGs. Domain knowledge structure, including category hierarchy and attribute relationships, can be utilized as discriminative information to facilitate KG based learning and verification. However, in previous fact checking research, category hierarchy and attribute information was often scattered in a KG and treated as the ordinary triple facts in the learning process like other types of information, or was utilized in a limited way without the consideration of category hierarchy or the combination of category hierarchy with the learning process. Thus to better utilize category hierarchy and attribute relationships, in this paper, we propose an end-to-end knowledge structure driven prototype learning and verification method for fact checking. For improving intra-category compactness and inter-category separation, we develop a hierarchical prototype learning technique that jointly learns a prototype for each sub-category to enhance entity embeddings and optimize embedding representations using highlevel category. For further enhancing embedding learning, we propose a graph attention network to aggregate information from neighboring attribute nodes. We construct a real-world dataset on food domain, and experimental results on the benchmark datasets and our domain dataset show the effectiveness of our method compared to both previous fact checking methods and representative KG reasoning methods.& nbsp;(c) 2021 Elsevier B.V. All rights reserved.