Another way to interpret the relevance model is via Bayesian estimation: the relevance model could be one of a large set of urns. We know what the urns are, but don't know which one is correct, so we compute the posterior probability for each candidate urn, and combine them via Bayesian (conditional) expectation.