Modeling the Score Distributions of Relevant and Non-relevant Documents