Using controlled query generation to evaluate blind relevance feedback algorithms