Large margin training for hidden Markov models with partially observed states