Events

Human-human conversation is highly interactive, on a very rapid time-scale, with participants providing a range of overlapping and low-latency responses and signals, as speech is occurring, to facilitate a rapid and successful communication process.  My research aims to advance our understanding of how such fluid turn-taking and interaction can be modeled and replicated computationally.  I am also interested in the impact that specific system behaviors and interaction styles can have on user perceptions, user disclosure of information to automated systems, and overall user satisfaction. In this talk, I highlight two recent empirical studies in which hundreds of human interlocutors have conversed with spoken dialogue systems and virtual human systems developed at the University of Southern California.  Both studies compare human-human interaction to

human-agent interaction, and provide insights into how improved turn-taking skills can facilitate more human-like interaction and improved performance in state-of-the-art dialogue systems. The first study focuses on a virtual human interviewer that conducts automated spoken interviews of users who may be suffering from psychological distress conditions such as depression and PTSD.  The study illustrates how turn-taking and dialogue policy decisions can be customized to help build rapport and encourage disclosure of sensitive personal information by the system's human users.

 

The second study explores the extent to which fluid, low-latency turn-taking can be achieved in an automated system designed to participate in a negotiation roleplay scenario with students of negotiation.  This project relies on a combination of human-human negotiation data, Wizard-of-Oz data, and statistically trained language processing components in an effort to bring a more fluid and natural interaction style to an automated negotiation system.