Methods, systems, and apparatus, including computer program products, for generating or using augmentation queries. In one aspect, subject phrases for detection in content are identified. Each phrase has a corresponding cardinality of terms. First hash sets for each of the subject phrases are generated, each first hash set including first hashes of bigram term subsets for each of the phrases. Sub-phrase scores for each of the hashes based on the cardinality of each phrase are assigned. The sub-phrase scores a used to detect the subject phrases in hashes of portions of received content. Other implementations of this aspect include corresponding systems, apparatus, and computer program products.