Most existing negative sampling methods for knowledge graph embedding models involve randomly selecting entities to replace the head or tail entities of factual triples to generate negative triples, resulting in lower-quality negative triples that affect the feature learning capability of entities and relationships. In order to solve this problem, after studying the factors affecting the quality of negative sampling, a knowledge graph connection prediction method based on the negative sampling of adversarial network and entity-level information is proposed. First, after inputting triple samples, model the semantic hierarchy of triples to increase hierarchical constraints. Then, based on the features of entities that are similar within the same hierarchy, select similar entities and relationships in a polar coordinate system and provide them to the generator in the adversarial network. This is used to generate high-quality negative triples for the discriminator, ultimately obtaining an ideal discriminator model, thus improving the performance of knowledge graph embedding models. Results from comparative and ablation experiments demonstrate significant improvements achieved by the proposed method in the link prediction task, thereby confirming its effectiveness.