As a proactive defense approach, network threat hunting has become a research hotspot in modern cybersecurity due to its advantages, including strong initiative, high detection accuracy, rapid response, and continuous optimization capabilities. However, traditional subgraph matching methods based on combinatorial optimization suffer from high computational complexity, making them unsuitable for real-time threat detection in large-scale provenance graphs with millions of nodes. Although existing Graph Neural Network (GNN) approaches demonstrate superior efficiency and robustness, they often overlook the semantic information of entire behavior chains during the embedding process. To address these challenges, we propose SemantiHunt, a behavioral semantics-driven method for network threat hunting. Our method takes “behavior” as the basic modeling unit. We first use a pre-trained BERT model to perform initial semantic embedding for each behavior, and then design a behavior-centric message passing mechanism to efficiently capture and integrate the fine-grained semantic features of continuous behavior chains within the graph structure. Experimental results show that SemantiHunt can more effectively distinguish attack paths from normal operation paths in complex attack chain identification tasks. Compared with the current advanced method Provg-Searcher, the overall accuracy rate is increased by an average of about 0.81%.