The majority of Procedural Content Generation (PCG) research has made use of human authored rules, heuristics and evaluation metrics. Machine learning techniques have gone relatively unused in PCG. We introduce a data-driven level generation approach, and apply it to the of dungeons for Zelda-like Action Roleplaying Games (ARPGs). We use Bayesian Networks (BNs) to learn distributional information about level topology. The learned networks can then be sampled to generate levels that have the same statistical properties as human authored levels.