24.4.1 Designing Dialogue Systems

How does a dialogue system developer choose dialogue strategies, architectures, prompts, error messages, and so on? This process is often called VUI (Voice User Interface) design. The user-centered design principles of Gould and Lewis (1985) are:

1. Study the user and task: Understand the potential users and the nature of the task, via interviews with users and investigation of similar systems, and study related human-human dialogues.

2. Build simulations and prototypes: In Wizard-of-Oz systems (WOZ) or PNAMBIC (Pay No Attention to the Man Behind the Curtain) systems, the users interact with what they think is a software system, but is in fact a human operator ("wizard") behind some disguising interface software (e.g. Gould et al., 1983; Good et al., 1984; Fraser and Gilbert, 1991). The name comes from the children's book The Wizard of Oz (Baum, 1900), in which the Wizard turned out to be just a simulation controlled by a man behind a curtain. A WOZ system can be used to test out an architecture before implementation; only the interface software and databases need to be in place. The wizard's linguistic output can be disguised by a text-to-speech system, or via text-only interactions. It is difficult for the wizard to exactly simulate the errors, limitations, or time constraints of a real system; results of WOZ studies are thus somewhat idealized, but still can provide a useful first idea of the domain issues.

3. Iteratively test the design on users: An iterative design cycle with embedded user testing is essential in system design (Nielsen, 1992; Cole et al., 1994, 1997; Yankelovich et al., 1995; Landauer, 1995). For example Stifelman et al. (1993) built a system that originally required the user to press a key to interrupt the system. They found in user testing that users instead tried to interrupt the system (barge-in), suggesting a redesign of the system to recognize overlapped speech. The iterative method is also very important for designing prompts which cause the user to respond in understandable or normative ways: Kamm (1994) and Cole et al. (1993) found that directive prompts ("Say yes if you accept the call, otherwise, say no") or the use of constrained forms (Oviatt et al., 1993) produced better results than open prompts like "Will you accept the call?". Simulations can also be used at this stage; user simulations that interact with a dialogue system can help test the interface for brittleness or errors (Chung, 2004).

See Cohen et al. (2004), Harris (2005) for more details on conversational interface design.