|
Realtime multimodal turntaking| old_uid | 4930 |
|---|
| title | Realtime multimodal turntaking |
|---|
| start_date | 2008/05/27 |
|---|
| schedule | 13h30 |
|---|
| online | no |
|---|
| summary | Getting computers to respond in a human-like manner in realtime dialogue presents numerous challenges. One of the larger -- but often overlooked -- issues is the highly dynamic nature of such interaction. This includes how conversants decide, in fact negotiate, how to "divide the work" of talking, using various kinds of multimodal behaviors for orchestrating it. Over the last decade research into multimodal communication has brought some progress to this issue but significant questions remain with regard to the perception-action loop, in particular, what kinds of perception, planning and prediction is involved. I will present some results from an effort to build holistic artificial dialogue systems with reaction times and capabilities similar to those of human-human dialogue. For a deeper understanding of dialogue, and indeed general cognitive issues related to communication, I argue that we need to approach the issue as the *design of a complex system* (a system whose behavior cannot be simply inferred from the behavior of its components). Looking at the problem holistically we need to consider what kinds of architectures are likely to produce the kinds of behaviors exhibited by people in natural communication. I will describe the latest results from CADIA in architectural modeling of natural dynamic turntaking and dialogue, and present data from models using classical as well as hybrid neural information approaches. |
|---|
| responsibles | Lœvenbruck, Welby |
|---|
| |
|