Provided are, among other things, systems, methods and techniques for classifying a collection of documents. A term is identified based on an indication of ability of the term's presence within a given document to predict whether the given document should be classified into an identified category. A document index is then queried using the identified term and, in response, search results that define a candidate set of documents are received. Finally, a classifier is applied to documents within the candidate set to determine which of the documents should be classified into the identified category.