Shape Expressions have recently been proposed as a high-level language to intuitively describe and validate the topology of RDF graphs. Current implementations of Shape Expressions are focused on checking which nodes of certain graph fit in which defined schemata, in order to get automatic typings or to improve RDF data quality in terms of completion and consistency. We intend to reverse this process, i.e., we propose to study the neighborhood of graph nodes that have already been typed in order to induce templates in which most of the individuals fit. This will allow to discover latent schemata of existing graphs, which can be used as a guideline for introducing coherent information in existing structures or for quality verification purposes. We consider that collaborative or general-purpose graphs are a specially interesting domain to apply this idea.