Combining Verbal and Nonverbal Features to Overcome the "Information Gap" in Task-Oriented Dialogue

Dialogue act modeling in task-oriented dialogue poses significant challenges. It is particularly challenging for corpora consisting of two interleaved communication streams: a dialogue stream and a task stream. In such corpora, information can be conveyed implicitly by the task stream, yielding a dialogue stream with seemingly missing information. A promising approach leverages rich resources from both the dialog and the task streams, combining verbal and non-verbal features. This paper presents work on dialogue act modeling that leverages body posture, which may be indicative of particular dialogue acts. Combining three information sources (dialogue exchanges, task context, and users’ posture), three types of machine learning frameworks were compared. The results indicate that some models better preserve the structure of task-oriented dialogue than others, and that automatically recognized postural features may help to disambiguate user dialogue moves.