Next utterance prediction based on two kinds of dialog models