Advanced persistent threats (APTs) pose significant challenges to the security detection of large-scale multi-host systems. To address these challenges, a novel detection framework named GE-APT was proposed. First, critical log information and features were extracted, and semantic representations were obtained using the Sentence-BERT model. A self-supervised autoencoder was then employed to compress and refine these high-dimensional semantic features. Subsequently, a system provenance graph was constructed based on system entities and their interactions, with semantic edge features introduced to enhance the expressive power of the graph structure. To model the distribution of normal behavioral patterns at the node level, an improved generative adversarial network was adopted for adversarial learning, while the discriminator was designed using a graph attention mechanism to effectively identify anomalous activities. Finally, GE-APT achieved precision rates of 90.91% and 86.86% on datasets DARPA TC and OpTC, respectively, and outperformed existing typical detection methods in accuracy, recall, F1-score, and other metrics. These results validate the effectiveness of the proposed framework in detecting anomalous behaviors associated with APT attacks.