Specification and Generation of Custom-Tailored

Knowledge-Acquisition Tools?

Henrik Eriksson

Medical Computer Science Group

Knowledge Systems Laboratory

Stanford University School of Medicine

Stanford, California 94305-5479

U.S.A.

Abstract

Domain-oriented knowledge-acquisition tools provide efficient support for the design of knowledge-based systems. However, the cost of developing such tools is high, especially when their restricted scope is taken into account. Developers can use metalevel tools to generate domain-oriented knowledge-acquisition tools that are custom tailored for a small group of experts, with considerably less effort than is required for manual tool development. An epistemic obstacle to creating such metatools is the specification model for target knowledge-acquisition tools. The metatool dots is based on an abstract-architecture approach to the specification and generation of knowledge-acquisition tools. Dots is domain and method independent, because it is based on an architectural model of the target knowledgeacquisition tool.

1 Introduction

Many knowledge-acquisition tools are unsuitable for their tasks because they are adapted neither to the application domain, not to the requirements of individuals, such as developers and experts. Researchers in knowledge acquisition are experimenting with knowledgeacquisition tools custom tailored for specific domains [Gale, 1987; Musen et al., 1987]. Usually, such domainoriented knowledge-acquisition tools are more useful than are general knowledge-acquisition tools, because custom-tailored tools can meet the requirements of the particular knowledge-acquisition situation.

Simultaneously, traditional knowledge engineering and expertise transfer is being replaced gradually by methodologies where developers assemble problem solvers for knowledge-based systems from reusable

?The preparation of this manuscript has been supported in part by grants LM05157 and LM05305 from the National Library of Medicine, by gifts from Digital Equipment Corporation, and by scholarships from the Swedish Institute, from the Fulbright Commission, and from Stanford University. The development of dots has been supported by the Swedish National Board for Industrial and Technical Development (nutek).

method components that accomplish subtasks [Chandrasekaran, 1986; Steels, 1990]. In McDermott's [1988] approach, developers use method-specific knowledgeacquisition tools to acquire the domain knowledge required by the methods. Method-oriented knowledgeacquisition tools, however, are not domain oriented per se; they must be adapted to specific domains and individuals.

Developers of knowledge-based systems wishing to use domain-oriented knowledge-acquisition tools face several barriers: It is difficult and laborious for developers of knowledge-based systems to adapt existing knowledgeacquisition tools, and to implement new domain-oriented knowledge-acquisition tools for new domains. Another barrier is that the investment of developing and maintaining domain-oriented tools cannot always be justified within the budget of a single application project. Many of these barriers can be eliminated by tools that enable developers to generate new domain-oriented knowledgeacquisition tools from high-level descriptions. Such metatools can simplify the task of developing domainoriented knowledge-acquisition tools, and can reduce significantly the work required to implement these tools. Thus, metatools can make domain-oriented knowledgeacquisition tools feasible in situations where these tools could not be used previously. The knowledge-acquisition tools generated can then support the development of the target knowledge-based systems.

The design of metatools presents several epistemological and technical challenges. The way developers view knowledge acquisition affects the way that they design knowledge-acquisition tools. The developer's view of target knowledge-acquisition tools determines the appropriate specification strategy for target knowledgeacquisition tools in metatools. Automatic generation of such knowledge-acquisition tools requires a high-level description|or metaview|of the target tools [Eriksson and Musen, in press]. For example, prot?eg?e [Musen, 1989] is a metatool that generates a domain-oriented knowledge-acquisition tool from an instantiation of a generic problem-solving method. The drawbacks of method-oriented metatools, such as prot?eg?e, are that the problem-solving method supported cannot be replaced easily, and that such metatools do not handle combined methods well.
We have formulated a metaview, the abstract-