End-to-end Modeling for Selection of Utterance Constructional Units via System Internal States