Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for semantic document analysis. In one aspect, methods include the actions of segmenting a document into segments; generating semantic representations, each corresponding to one of the segments; determining a corresponding segment score for one or more of the segments based on the corresponding semantic representation, such that each segment score represents a change in the corresponding segment; comparing each segment score to a threshold score, such that the threshold score represents an expectation of change for the document; and identifying segments having segment scores that indicate the change in the corresponding segment deviates from the expectation of change for the document based on the comparison.