Natural Language Dialogue group

We are performing research to enable incremental dialogue capabilities in our virtual humans. Recently, we have focused on developing techniques for the incremental interpretation and prediction of utterance meaning in dialogue systems. These techniques are motivated by the high-level goal of enabling virtual humans to react more quickly, and in more human-like ways, to real-time user speech. These quick reactions will reflect the virtual human's developing understanding of the user's speech, and may include responsive overlap behaviors such as verbal and non-verbal back-channels, interruptions, or even having the virtual human complete the user's utterance once it has been understood.

Our research has shown that relatively high accuracy can be achieved in understanding of spontaneous utterances before utterances are completed (Sagae et al, NAACL 2009). We have also developed a further machine learning approach that enables a virtual human to determine when it has reached a point of maximal understanding of an ongoing user utterance (DeVault et al, SigDial 2009). We have used this model to develop a prototype implementation that allows this virtual human to strategically initiate a system completion of some user utterances (DeVault et al, SigDial 2009; Sagae et al, NAACL-HLT 2010).