<p>Knowledge reasoning based on knowledge graph embedding relies on negative sampling to enhance training quality. Traditional random negative sampling methods often suffer from semantic mismatch, where the generated negative samples are too distant from the positive ones, resulting in insufficiently informative training signals. Although recent approaches based on generative adversarial networks(GAN) have attempted to produce negative samples that are more semantically relevant and more challenging, they still fail to adequately capture the conceptual context of entities. To address these challenges, we propose a knowledge reasoning framework based on a concept-guided generative sampling paradigm (CGSP). The framework introduces conceptual knowledge at multiple levels to guide adversarial negative sampling. First, it constructs a concept-level knowledge graph by grouping entities that share the same concept, which helps reduce conceptual ambiguity. Then, a generator based on adversarial learning produces candidate negative samples with strong semantic relevance, and these candidates are filtered using the concept graph to ensure consistency within the conceptual domain. Meanwhile, distinct sampling strategies are designed for unique and non-unique relations to further mitigate the false negative problem. Finally, the generated high-quality negative samples are used to train knowledge graph embedding models. Extensive experiments conducted on five representative models and six real-world datasets demonstrate that the proposed method substantially improves both the performance and stability of existing approaches.</p>