How can we estimate the language model of the relevant class if we have no examples of relevant documents? We play a sampling game as follows. The relevance model is an urn with unknown parameters. We draw several samples from it and observe the query. What is the probability of observing the word "w" on the next draw?