We present a model to express preferences on rhythmic structure, based on probabilistic context-free grammars, and a procedure that learns the grammars probabilities from a dataset of scores or quantized MIDI files. The model formally defines rules related to rhythmic subdivisions and durations that are in general given in an informal language. Rules preference is then specified with probability values. One targeted application is the aggregation of rules probabilities to qualify an entire rhythm, for tasks like automatic music generation and music transcription. The paper also reports an application of this approach on two datasets.