Use a self-defined feature which took word's different occurrence rates in different classes into consideration to train comments data for books, hotels and mobile phones.
For more details, see /docs/nlp_doc.pdf