Tagging systems can be modeled as tripartite networks consisting of users, resources, and tags. Analysis of tag co-occurrence networks reveals scale-free properties and non-trivial clustering that may encode semantic relationships between tags. A simple hierarchical model of user tagging is able to partially reproduce properties of real-world tag co-occurrence networks, providing evidence that users generally apply tags according to an underlying conceptual hierarchy. Further modeling semantic relationships could improve systems by facilitating categorization and spam detection.