Dialogue Patterns and Misunderstandings

Disruptive errors are common in many human-computer (HC) dialogues. We manually applied initiative and dialogue act annotations to HC dialogues in the travel domain in an effort to find patterns that are predictive of misunderstandings. While we found some interesting patterns of dialogue acts, we also found that a detailed understanding of the misunderstandings in our data required us to perform more in-depth analysis than is possible just by examining dialogue acts. Our hope is that analyses such as these will inform the design of HC dialogue systems, so that systems may predict problematic situations in order to deal with them more effectively.