Frequent pattern mining is an important research field and can be applied to different labeled data structures ranging from itemsets to graphs. There are scenarios where a label can be assigned to a taxonomy and generalized patterns can be mined by replacing labels by their ancestors. In this work, we propose a novel approach to generalized frequent subgraph mining. In contrast to existing work, our approach considers new requirements from use cases beyond molecular databases. In particular, we support directed multigraphs as well as multiple taxonomies to deal with the different semantic meaning of vertices. Since results of generalized frequent subgraph mining can be very large, we use a fast analytical method of p-value estimation to rank results by significance. We propose two extensions of the popular gSpan algorithm that mine frequent subgraphs across all taxonomy levels. We compare both algorithms in an experimental evaluation based on a database of business process executions represented by graphs.