In this paper we describe how to learn Bayesian networks from a summary of complete data in the form of a dependency network rather than from data directly. This method allows us to gain the advantages of both representations: scalable algorithms for learning dependency networks and convenient inference with Bayesian networks. Our approach is to use a dependency network as an "oracle" for the statistics needed to learn a Bayesian network. We show that the general problem is NP-hard and develop a greedy search algorithm. We conduct a preliminary experimental evaluation and find that the prediction accuracy of the Bayesian networks constructed from our algorithm almost equals that of Bayesian networks learned directly from the data.