Predicting Document Coverage for Relation Extraction