Some embodiments provide a method for determining a relatedness of content items to categories. The method identifies a particular content item, a relevancy score associated with the particular content item, and a set of categories to which the particular content item is classified as related. Based on a set of glossaries associated with the set of categories, the method calculates a set of affinity scores that each represents a degree of relevancy between the particular content item and a category in the set of categories. The method modifies the relevancy score associated with the particular content item based on the calculated set of affinity scores.