Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for machine learning. In one aspect, a method includes receiving a collection of facts, each fact represented as an entity-attribute-value tuple; identifying expected values for one or more individual attributes, where the identifying expected values includes, for each particular attribute: identifying facts having the attribute, calculating a value score for facts of the collection of facts having the particular attribute for each particular value, calculating a global score for all facts of the collection having the attribute, and comparing the value score to the global score such that a value is identified as an expected value if the comparison satisfies a specified threshold.