Combining gradient axes with binary relevance filtering yields better results: irrelevant documents are first filtered out by binary judgment, and then continuous scores are used for fine-grained ranking.