Events Calendar

CS Colloquium

Speaker: Dr. David DeVault, USC Institute for Creative Technologies (ICT)

Talk Title: Toward flexible, robust, and rapid understanding of user speech in natural language dialogue systems

Abstract: This talk presents recent research that targets two of the major limitations in current natural language dialogue systems. One limitation is that while systems face substantial uncertainty in understanding user speech, they usually have only a rudimentary ability to overcome uncertainty in their dialogues. A second and related limitation arises from the fact that human speakers are by nature highly interactive while speaking, using incremental responses such as backchannels, interruptions, and overlapping speech to signal their understanding and resolve uncertainty when it arises. However, most implemented dialogue systems have little or no support for incremental interaction.

In the first part of the talk, I will present a probabilistic approach to dialogue management, called "contribution tracking", which I developed as a way to improve the flexibility of dialogue systems in overcoming uncertainty. On this approach, when faced with an ambiguous utterance, systems can spawn multiple threads of interpretation to track the likely meanings as the dialogue proceeds. I will highlight several concrete results and benefits of this approach in an implemented dialogue system that plays a collaborative reference game. These benefits include improved robustness to clarification failure, flexible aggregation of information across utterances with probabilistic inference, and the use of successful
ambiguity resolution to automatically improve the agent's understanding models with machine learning.

In the second part of the talk, I will present more recent work, carried out within the dialogue group at the USC Institute for Creative Technologies, which has aimed to enable incremental interaction and overlapping speech in our SASO-EN virtual humans. This work has created a data-driven approach to incremental understanding and prediction of user utterance meaning during user speech. Among the results I will discuss is a prototype system that often enables a virtual human to anticipate how a user's utterance will end, and to quickly generate and utter a completion of the user's utterance for them. (Joint work with Kenji Sagae and David Traum.)


Biography: David DeVault is a Research Scientist at the USC Institute for Creative Technologies (ICT), where he is a member of the natural language dialogue group. David obtained his Ph.D. from the Department of Computer Science at Rutgers University in 2008, and was a Postdoctoral Research Associate at USC/ICT from 2008-2010. David's research focuses on the development of techniques to enable dialogue systems to respond to the inevitable uncertainties of communication in a way that is more flexible, more robust, and more human-like. His work spans the areas of natural language understanding, dialogue management, and natural language generation.


Host: Prof. Kevin Knight

Location: Seaver Science Library (SSL) - 150

Audiences: Everyone Is Invited

Contact: Kanak Agrawal