Advancing Dialog
Alexander Rudnicky, Carnegie Mellon University, USA
The study of spoken dialog encompasses several different approaches that offer complimentary though occasionally incompatible views of what appear to be the same phenomena. Nevertheless the field has progressed to a point where it is possible both extract useful information from observation of human-human interaction and to build systems that (more or less) support useful interactions between humans and computers. Despite these advances we still seem to be at an early stage of mastering the intricacies of dialog processing. Learning-based approaches offer a promising direction as they promise to reduce the need for hand-crafting system knowledge. But it is not always clear how advances in learning by themselves lead to more sophisticated dialog capabilities (as opposed to better tuned ones). One path would be to develop techniques for dynamic self-modification that allow systems to monitor their behavior and control knowledge acquisition on their own. This implies addressing the problem of how dialog and communication can be integrated into intelligent agents and how dialog processing can benefit from the information available in the broader context in which such agents operate.