M.Sc. Thesis Defense in Computer Science - Guðný R. Jónsdóttir - A Distributed Dialogue Architecture with Learning

M.Sc. Thesis Defense in Computer Science, Friday 14th of November at 10:00 am, room K5, Kringlan 1. All are welcome.

"A Distributed Dialogue Architecture with Learning"

Guðný R. Jónsdóttir

A wide range of conversational skills is required for conducting fluid dynamic dialogue. Giving computer-controlled agent turntaking skills that function at a realtime pace in a human-like manner has been both a practical and theoretical research goal of speech technology for decades. Control of response times - knowing when it is appropriate to start responding, interrupting and the like - is one of the challenges to this end. This thesis describes the use of machine learning to aid in the realtime classifications and use of valid turntaking auditory perceptual cues. The system has been implemented and tested in dynamic interaction with itself and with humans. Results show that the agent is very close to human speakers with regards to speed. In human conversation 70% of between-speaker silences are shorter than 500 msec. The system developed here approaches this, with 60% of between- speaker silences shorter than 500. The system learns this in online training sessions in roughly 80 turns.

Thesis Commitee:

Dr. Kristinn R. Thórisson, Reykjavík University (supervisor)

Dr. Jonathan Gratch, ICT/University of Southern California, USA

Dr. Yngvi Björnsson, Reykjavik University