User created playlists can be analyzed to create a statistical language model indicating the likelihood that a particular sequence of content attributes will be found in a playlist created by a user, as well as the likelihood of any sequence of one or more content attributes following a playlist or partial playlist created by a user. The language model can be used to generate a recommended content attribute sequence based on a partial playlist of one or more content items. A recommended content item sequence that will be pleasant to a user when added to the partial playlist can be selected based on the recommended content attribute sequence.