How do people rapidly learn rich, structured concepts from sparse input? Recent approaches to concept learning have found success by integrating rules and statistics. We describe a hierarchical model in this spirit in which the rules are stochastic, generative processes, and the rules themselves arise from a higher-level stochastic, generative process. We evaluate this probabilistic language-of-thought model with data from an abstract rule learning experiment carried out with adults. In this experiment, we find novel generalization effects, and we show that the model gives a qualitatively good account of the experimental data. We then discuss the role of this kind of model in the larger context of concept learning.