11 Dec 2001

Bettina Braun & Tiina Karsikas (Saarland University)

Dialogue systems in which humans communicate with machines will become more and more common. This is especially true in in situations in which speech interaction is easier than manual operation (e.g. in car environment). Current dialogue systems, however, still have a very rigid dialogue structure and inappropriate speech output, which prevents and many users from supporting these systems.

Within the large field of research in man-machine interaction I am mainly interested in canned speech output and in how cooperation can be conveyed/simulated with this technique. Thus, I want to find out which behaviour is judged as cooperative.

Under the assumption that humans behave cooperatively in dialogues, a pilot study with human operators was performed. They were told to interact with real users but were restricted to a set of utterances (to simulate a real dialogue system). The analysis of the data showed that only a few human operators behaved as expected. Reasons for this may partly be found in the experimental setup. I will discuss this difficulty of evoking cooperative utterances in dialogue in a controlled way in slightly more detail.

The last part of my talk will be about the problem of assessing cooperation in dialogue systems. Even if we had a maximally cooperative way of reacting to different sitations it is still questionable if this behaviour would be appropriate for machines. A perception study aims at ratings of operator behaviours in two conditions (once under the assumption that human-human interaction is to be judged, once under the assumption of man-machine interaction). For this part only preliminary ideas can be presented.

[back to PWorkshop Archives]

<owner-pworkshop@ling.ed.ac.uk>