Beyond Structured Dialogues: Incorporating Clark’s Models of Language Usage

In this paper, we discuss the use of Clark and Wilkes-Gibbs’ model of collaboration (1986) and Clark and Schaefer’s model of contributions (1989) in using structured models of dialogue. Rather than completely specifying the behavior of a dialogue system by a structured dialogue model, we advocate that the dialogue structure should only specify the behavior of the system with respect to high-level contributions. The realization of these contributions and the associated grounding that accompanies them would be handled by a general dialogue manager based on Clark’s models. 1 FINITE STATE DIALOGUE MODELS In building spoken dialogue systems, the designer needs to ensure that the system is able to properly understand what the user said. This is difficult because the system might misrecognize what the person said, or might not properly interpret it. One technique for overcoming this obstacle is to use a structured dialogue approach. This approach involves keeping initiative solely with the system, and letting it direct the user through a series of pre-ordained steps to accomplish the task. The steps are described as a finite state dialogue model. Each state has an associated system prompt, and a set of possible responses that the user can make. The determination of the next state is a function of both the current state and the user’s response. Figure 1 gives a simple structured subdialogue that allows the user to specify information for finding train information, namely the destination city, the initial city, and the time that the train leaves. The advantage of structured dialogues is that the system prompts encourage the user to say something from a limited set of possible responses. This simplifies speech recognition and almost eliminates the need for natural language processing. It also simplifies the dialogue management component, which is completely specified by the specification of the transitions. This technology has proved very popular for building working systems, and the Center for Spoken Language Understanding at Oregon Graduate Institute even provides a toolkit for building such systems with this approach (Novick and Sutton, 1996; Sutton et al., 1996). Although such systems seem overly restrictive for the user, Walker et al. (1997) found that users preferred a system-initiative system over a mixed-initiative system, probably due to the former giving more reliable performance. 2 LIMITATIONS OF STRUCTURED DIALOGUE When a designer specifies a dialogue by means of a finite state dialogue model, the designer must account for all dialogue behavior between the system and the user. As such, he must hand-code the grounding behavior in the dialogue: the process by which participants ensure what they have said is mutually understood (Clark and Schaefer, 1989). In a structured dialogue, this is often achieved by asking the user to verify the recognition and understanding results of the system after each piece of information