Extreme value theory applied to document retrieval from large collections