I introduce NTT's open domain conversation project. For understanding open domain utterances, we use predicate-argument structure (PAS) analysis. I describe how we made our PAS analyzer for dialogues, with a special emphasis on zero-anaphora resolution. We also use PASs for utterance generation. I describe how our system generates utterances by using PASs collected from a large number of blogs. I also describe how we use Twitter data for utterance generation. At the end of the talk, I will show some evaluation results of our dialogue experiments using novice users.