Automatic de-identification of electronic medical records using token-level and character-level conditional random fields