Agreed. For many many applications, bag of words/tf-idf/support vector machines are enough but with much better training time.