A Methodology and Tool for Knowledge Acquisition

Because knowledge is a critical factor affecting the performance of an expert system, knowledge acquisition is a priori a fundamental task of any knowledge engineering enterprise. A number of methodologies, tools and techniques have been proposed in order to address the problem of knowledge acquisition. In this paper we describe knowledge engineering from a functional point of view which emphasizes the stages involved. In particular, we stress the need for tools supporting empirical data analysis as well as facilities for allowing a smooth transition from data analysis to domain conceptualization and finally to the system implementation. The strengths and weaknesses of our earlier system, KEATS-1, have led us to embark upon the design and implementation of a new system, KEATS-2, which provides a more coherent and integrated framework for performing both bottom-up and top-down knowledge acquisition. This paper introduces the first of our redesigned KEATS-2 modules, entitled Acquist, and describes how it provides support for data analysis and domain conceptualization. Acknowledgement: This research is supported by a grant from British Telecommunications, plc. John Domingue participated in the discussions which led to this paper. Steven Rose and Mike Stewart of the Open University's Brain Research Group provided valuable domain expertise. 1. THE PROBLEM OF KNOWLEDGE ACQUISITION The most popular principle in knowledge based systems states that the performance of an expert system critically depends on the amount of knowledge embedded in the system [1]. Therefore the knowledge engineer usually spends a great deal of time eliciting knowledge from domain experts and even more trying to make sense of the data acquired. This combined activity of eliciting, interpreting and organizing the knowledge acquired from the expert is called 'knowledge acquisition', and is often described as a lengthy and painful process. A number of factors related to the problem being tackled, the techniques used for knowledge elicitation and the personality of the domain expert can make this process extremely difficult, although clearly this is not inherently so, as illustrated by the nonproblematic cases itemized in Figure 1. • There is a well established knowledge elicitation technique that suits the current task. For instance, the Repertory Grid method works very well for tackling problems of simple data classification. • The knowledge engineer already has a model of the task that can drive the knowledge elicitation sessions. • The structure of the domain and the problem solving strategies are trivial. This is true for a number of mundane domains, for which expert systems are nowadays being built. Figure 1. Knowledge acquisition is not a problem when any of these conditions prevails. Nevertheless, problems can arise, due to the complexity of the domain and to possible mismatches between the knowledge elicitation technique and the structure of the problem (e.g. techniques such as protocol analysis work very badly for domains which are best represented declaratively). Moreover, recent empirical studies have shown that the efficacy of a particular knowledge elicitation technique can also be affected by the personality traits of the domain expert [2]. Much research has tackled the problem of devising tools [3] and techniques [4] that could support, speed up and eventually automate the knowledge acquisition process. Although all of these proposals attempt to tackle the same problem, they vary a great deal depending on the approaches, functionalities, underlying models and assumptions. In this paper we address some of the issues concerning knowledge acquisition, trying to clarify the nature of the knowledge acquistion process, and to point to the kind of software support the knowledge engineer can be provided with. The rest of the paper is organized as follows. In section 2 we outline a functional model of knowledge engineering that will provide both the basis for clarifying the typology of the knowledge acquisition task and the theoretical framework for devising a knowledge acquisition tool. In section 3, we describe how these considerations led to the implementation of data analysis and data conceptualization tools in the original KEATS system [5], hereafter referred to as KEATS-1. The strengths and weaknesses of KEATS-1 have in turn led us to the development of a new system, Acquist, that is the kernel of the knowledge acquisition facilities being provided as part of our ongoing KEATS-2 project. Section 4 describes Acquist's structure, and its support for bottom-up and top-down approaches to knowledge acquisition. Finally, section 5 provides conclusions and a restatement of our current view of the state of knowledge acquisition. 2. A STAGE-ORIENTED MODEL OF KNOWLEDGE ACQUISITION