Navigating toward visual targets in unfamiliar environments continues to be a core difficulty in the field of robotics. While traditional methods and learning-based approaches have been widely explored, existing systems often lack commonsense understanding of household objects and their spatial arrangements. To tackle this issue, we introduce an innovative visual navigation framework that integrates Large Language Models (LLMs) with Graph Neural Networks (GNNs) to enhance semantic reasoning. The framework extracts semantically relevant frontiers from the semantic map using language guidance and selects them as long-term goals to guide efficient exploration. Extensive evaluations conducted on the Gibson and HM3D datasets validate the superior generalization capability and navigation effectiveness of our approach over traditional map-based methods. Our ablation studies further reveal that the commonsense knowledge embedded in the LLMs significantly improves semantic exploration, thereby boosting both efficiency and accuracy. By combining visual perception, semantic mapping, and path planning, our method enables robots to understand natural language instructions and autonomously perform goal-directed navigation toward a specified object. To demonstrate the framework’s applicability, we conduct real-world experiments using the TRACER robot.