Feng-Jen Yang, Jung Hee Kim, Michael Glass, and Martha Evens, Illinois Institute of Technology, USA
Turn planning is not well-studied in natural language dialogue systems. The current version of CIRCSIM-Tutor attempts to imitate human tutorial dialogue. It generates comprehensible conversation but sounds unnatural. We propose to add a turn planner to improve intra-turn coherence by overall planning of variable references, discourse markers, acknowledgements, and softeners. This paper specifies the architecture of the turn planner, which we believe will improve the rhetorical structure to a reasonable level.