Relevance feedback is an effective and widely accepted method in information retrieval to improve performance. Relevance feedback generally uses an adaptive learning method to estimate the userís information need. In this research, we propose an alternative two-stage sampling method to obtain an unbiased estimate of the userís information need. Our estimate shows not only improved retrieval performance, but also better prevention of query drift, which troubles traditional relevance feedback. We also give theoretical justification and empirical support for this method.