Description logics have become a prominent paradigm in knowledge representation (particularly for the Semantic Web), but they typically do not include explicit representation of uncertainty. In this paper, we propose a framework for automatically learning a Probabilistic Description Logic from data. We argue that one must learn both concept definitions and probabilistic assignments. We also propose algorithms that do so and evaluate these algorithms on real data. © 2011 Springer-Verlag.