This paper proposes an unsupervised training approach for SLU systems that leverages the structured semantic knowledge graphs of the emerging SemanticWeb. The approach creates natural language surface forms of entity-relation-entity portions of knowledge graphs using a combination of web search retrieval and syntax-based dependency parsing. The new forms are used to train an SLU system in an unsupervised manner. This paper tests the approach on the problem of intent detection, and shows that the unsupervised training procedure matches the performance of supervised training over operating points important for commercial applications.