In the recent past, several spoken-language dialogue applications have been implemented. In most of the cases, the implementations focus on one particular task such as Air Travel Information Service (ATIS) (see, e.g., [Ward1994]) or hotel reservation and travel information ([Constantinides et al1998]). In some cases ([Ferrieux and Sadek1994]), a shift towards task-independent implementations can be observed, leading to a principle-based implementation of a task-oriented dialogue system [Sadek et al.1997], taking advantage of the structural similarity in task-oriented dialogues of different domains. Most of the above-cited applications have in common that they are able to perform a limited set of operations (such as hotel reservations) and that, in order to perform these operations, the user needs to specify a certain amount of information (such as arrival date). Put simply, the task of the natural language understanding component in these implementations is to determine the operation the user wants to perform, and then obtain the information necessary to perform the operation.
On the other hand, there are implementations of dialogue toolkits (see, e.g., [Sutton et al1996] ) aiming at providing a platform to design dialogue systems without the need to take recourse on linguistic specifications. Approaching the problem of task-independent dialogue strategies from the other side, these systems typically offer an implementation of a template dialogue system bare of any task-specific knowledge at the expense of less sophisticated models of dialogue structure. When instantiating the system for a particular task, the system designer typically has to specify the flow of the dialogue, for example in form of a finite state automaton. Disadvantages of this approach are the stiff information flow following the specification and the fact that complementary information sources such as results from database requests can only be integrated with difficulties.
: A part of the type hierarchy and its appropriateness
conditions used in the map application. The
least specific type is at the bottom of the tree. Information
increases from the bottom to the top. Two sub-domain models, called date & time and reservations
are merged with the application-specific declaration of the types of the goals. The part of the hierarchy
declaring the speech acts is domain-independent. This is a simplified presentation of the domain model
actually used in the system.
The work presented in this paper aims at combining advantages of the first type of system - such as natural dialogue structure - with the key advantage of the second type of system, namely easy deployment for new tasks. We assume that the behavior of a dialogue system can be sufficiently described by answering the following questions: (i) What are the entities, properties and actions the user and the system may refer to during dialogue? (ii) What kind of information is sufficient for the system in order to perform the action the user intended the system to perform? and (iii) How should the system perform the intended actions? Consequently, we are interested in separating domain-independent and domain-dependent knowledge in order to simplify as much as possible the specification for new systems. We show how the behavior of the natural language processing component in a dialogue system can be specified using declarations answering the three questions above, namely specification of a domain model, a task model and clauses describing the systems' behavior. In each instance, the specifications consist of a set of domain-dependent and a set of domain-independent specifications.
From a processing point of view, we describe a system in which the way of determining information to be exchanged is domain-independent whereas the exchanged information itself may be domain-dependent. As a result, we arrive at a specification of a dialogue system in which domain-specific and domain-independent knowledge are orthogonal.
The system has been implemented in a travel information booth setting. Currently the system is capable of performing hotel and restaurant reservations and generating path descriptions to sites of touristic interest.
