Talking to a Computer Is Not like Talking to Your Best Friend

When developing natural language interfaces, it is important to know the characteristics of the language used. We have developed a set of tools for conducting human-computer natural language dialogue simulations (Wizard of Oz experiments) and for analyzing the data obtained. We report methods used and results obtained from a series of such experiments. The focus of the study was on occasional users and on the structure of the dialogue. Three different background systems were used; one database and two advisory systems of different complexity. The results point to the need for mechanisms for handling connected discourse in interfaces for this user group. They also indicate that there are different classes of dialogue situations requiring discourse representations of different computational complexity.