In the domain of computer vision, Zero-Shot Learning (ZSL) achieves the classification of unseen class objects through the utilization of semantic information of class relationships. Acquiring richer semantic information and representation pose a significant avenue for enhancing learner performance. Existing studies of ZSL predominately address this challenge only by introducing knowledge graphs and graph neural networks, overlooking inadequacies in the original semantic information, the intrinsic hierarchical and the directional characteristics within the graph structure. This paper proposes two semantic information enhancing methods for ZSL, respectively tailored for regular datasets and large-scale datasets. Facing regular ZSL datasets, our method leverages textual knowledge within large language models, extending traditional 2-dimensional attribute annotations to a 3-dimensional space to obtain more comprehensive class-level semantic information. Addressing the large ZSL tasks, our approach combines enhanced semantic information with external knowledge graphs to simulate class relationships, employing the intrinsic structure and directionality of graphs to bolster semantic representations. We validated our approaches on four traditional ZSL datasets and the ImageNet dataset. The experimental results manifested significant improvements in ZSL performance, underscoring the potential of our methods.