Something like a Hidden Markov Model might be better then? That way you can can keep the information about word order.