Determining themes is disclosed. Reputation data extracted from at least one data source is received. The reputation data includes a plurality of user-authored reviews. The presence of a first keyword is detected in a first review. The presence of a second keyword that is different from but associated with the first keyword is detected in a second review. A sentiment for a theme is determined based on the detected presence of the first and second keywords. A report that indicates the sentiment for the theme is provided as output.