There has been considerable interest in transforming unstructured social tagging data into structured knowledge for semantic-based retrieval and recommendation. Research in this line mostly exploits data co-occurrence and often overlooks the complex and ambiguous meanings of tags. Furthermore, there have been few comprehensive evaluation studies regarding the quality of the discovered knowledge. We propose a supervised learning method to discover subsumption relations from tags. The key to this method is quantifying the probabilistic association among tags to better characterise their relations. We further develop an algorithm to organise tags into hierarchies based on the learned relations. Experiments were conducted using a large, publicly available dataset, Bibsonomy, and three popular, human-engineered or data-driven knowledge bases: DBpedia, Microsoft Concept Graph, and ACM Computing Classification System. We performed a comprehensive evaluation using different strategies: relation-level, ontology-level, and knowledge base enrichment based evaluation. The results clearly show that the proposed method can extract knowledge of better quality than the existing methods against the gold standard knowledge bases. The proposed approach can also enrich knowledge bases with new subsumption relations, having the potential to significantly reduce time and human effort for knowledge base maintenance and ontology evolution.