A Multiple-Method Knowledge-Acquisition Shell for the Automatic Generation of Knowledge-Acquisition Tools

The use of predefined models of problem-solving methods is receiving considerable attention from researchers in the area of knowledge acquisition. Using these models, developers of knowledge-acquisition tools are able to prescribe the roles in which knowledge is used in completing a given task. A number of method-oriented architectures based on a single problem-solving method have been developed by various research groups. Because the methods are domain-independent, method-oriented architectures are limited by the fact that knowledge roles that depend on domain-specific considerations cannot be represented using the model of problem solving. In addition, the interface between the knowledge-acquisition tool and the application expert cannot adequately convey the role of each knowledge type in the task model. PROTÉGÉ-II is a knowledge-acquisition shell that we are building to generate knowledge-acquisition tools automatically without presupposing a specific model of problem-solving. The shell manages a library of mechanisms—procedures of grain size smaller than that of problem-solving methods. Mechanisms can be combined in PROTÉGÉ-II to construct problem-solving methods and to define the roles of knowledge that depend on domain considerations. Furthermore, PROTÉGÉ-II utilizes the concept of adaptation in interfaces to allow the knowledge engineer to produce interfaces that are taskand domain-specific. In this paper, we present the PROTÉGÉ-II shell and examine the components of its architecture. We also demonstrate the use of PROTÉGÉ-II with a running example, and discuss the design techniques used to overcome the limitations of method-specific architectures. 1. Method-Oriented Architectures A recent focus of research in knowledge acquisition is the use of models of domain-independent problem-solving methods to construct knowledge-acquisition tools [McDermott, 1988]. These models of problem solving, such as heuristic-classification [Clancey, 1985] and skeletal-plan refinement [Friedland and Iwasaki, 1985], allow a knowledge engineer to develop a model of the task area at hand in terms of the abstract problem-solving method. Thus, taskand domain-specific knowledge-acquisition tools can be obtained from a taskand domain-independent model. Since the problem-solving models are independent of any knowledge-representation formalism, the modeling of tasks occurs at the knowledge level [Newell, 1982], where only the role of each type of knowledge is specified, as opposed to at the symbol level, where the representation of each type of knowledge must be described. There are several examples of knowledge-acquisition tools built from models of problem solving. ROGET [Bennett, 1985] used a specialized form of the heuristic-classification method to acquire knowledge for diagnostic tasks. SALT [Marcus and McDermott, 1989] used a propose-and-revise method for configuration tasks. There is also another category comprising tools that operate at a metalevel. These tools are able to generate knowledge-acquisition tools automatically from a model of a task. Examples of this type of tool are PROTÉGÉ [Musen, 1989a; Musen, 1989b], which is method-oriented, and DOTS [Eriksson, 1990], which does not follow a given problem-solving method. PROTÉGÉ, which was developed in our laboratory, views the problem of automating the construction of knowledge-acquisition tools as one of generating an interface to a knowledge editor that is utilized by an application expert. The components of the interface are determined by a task model developed from the method of skeletal-plan refinement [Musen and Tu, 1991]. PROTÉGÉ makes extensive use of graphical interaction modalities, such as graphical editors, which are especially well suited for capturing procedural knowledge (e.g., a flowchart diagram). Because the interfaces are based on the task model for the domain of interest, PROTÉGÉ-generated tools can guide the user through the acquisition sessions, ensuring that the knowledge captured is complete and consistent with respect to that model [Musen et al., 1987]. Like other method-oriented knowledge-acquisition tools, PROTÉGÉ suffers because its domain-independent problem-solving method cannot define the role of domain-dependent control knowledge. For example, in the domain of cancer therapy, there is no way for the knowledge engineer to use the skeletal-refinement method to declare how the effects of various specifications for altering the dose of a drug should be combined, because the concept of a dose adjustment is a domain-specific one. In these instances, a knowledge engineer must fall from the knowledge level to the symbol level to enter the required knowledge in a particular representation (e.g., by specifying a production rule or by changing the ordering of rules). Our current research aim is to identify building blocks, called mechanisms, of a grain size finer than that of problem-solving methods, that can be (1) combined to construct problem-solving methods, and (2) applied to define the role of domain-dependent knowledge [Tu et al., 1991]. Our approach is essentially empirical in that we hope to identify several of these mechanisms from the problem-solving methods being applied in our current development of various medical expert systems. After the mechanisms are identified, we will test their applicability to other tasks and domains, and their combinability into problem-solving methods different from the ones from which they were extracted. Other groups are carrying out parallel research to find ways to define the roles of domain-dependent and domain-independent knowledge at the knowledge level. Examples are Chandrasekaran’s research on generic tasks [Chandrasekaran, 1986], Steels’ work on the componential framework [Steels, 1990], and McDermott’s study of role-limiting methods [McDermott, 1988].