Basic Principles of Spoken Dialogue Systems

Spoken dialogue systems can help users with a wide range of tasks, from monitoring their health to choosing a nice restaurant. In order to design a good dialogue system, we need a good model of the relevant task, a good understanding of typical user goals, and a model of how users will interact with the system. For example, if we want to build a dialogue system that helps users monitor their blood pressure, we need to ensure that users enter both systolic pressure and diastolic pressure. If users just wish to enter their readings, then that is enough. But if they also wish to monitor lifestyle factors, the system also needs to ask about factors such as weight and amount of exercise taken.

Many dialogue systems try to control the interaction with the user as much as possible. At every stage of the dialogue, the system offers only a very limited number of options. Since this in turn greatly limits user responses, it is much easier for the system to understand what the user has to say. The more initiative the user has, the more difficult it becomes for the system to understand what the user wants and means.

An example of interaction with a very directive (system-initiative) system might be:

System: Welcome to the blood pressure symptom management system. Please say the number that corresponds to your systolic blood pressure.
User: One hundred ten.
System: Thank you. I have recorded a systolic blood pressure of one hundred and ten. Please say the number that corresponds to your diastolic blood pressure.
User: Eighty.
System: Thank you. I have recorded your blood pressure as one hundred and ten over eighty. Please say your weight in kilograms.
User: Sixty five.
System: Thank you. I have recorded your weight as sixty five kilograms and your blood pressure as one hundred and ten over eighty. If those values are correct, please say yes.
User: Yes.
System: Thank you for using the blood pressure symptom management system. Good-bye!

The following interaction shows a mixed-initiative system that gives the user the maximum possible leeway. This time, the user's utterances are far more complicated. In order to process them, dialogue systems need to be quite sophisticated.

System: Welcome to the blood pressure symptom management system. Please enter your data.
User: Well, when I stepped on the scales this morning, I couldn't believe my eyes: seventy! All those pies over Christmas! It's good that I did my blood pressure before I weighed myself - it would have been even higher afterwards. It was a hundred and ten over eighty, which is OK, I think.
System: Thank you. I have recorded your blood pressure as one hundred and ten over eighty, and your current weight as seventy kilograms. Anything else?
User: Not that I can think of right now, but I'll check again tomorrow.
System: Thank you for using the blood pressure symptom management system. Good-bye!