The Knowledge Modelling Paradigm in Knowledge Engineering

ion Heuristic Match Solutions Refinement Solutions Abstractions Fig. 1. Clancey’s heuristic classification model. Clancey’s analysis of rule-based systems provided an important milestone in knowledge engineering. He showed that a knowledge-level analysis [2] makes it possible to focus on what a system actually does, rather than how it does it. In other words we have a shift from an emphasis on the symbol level (Mycin is a rule-based system that reasons by means of backward-chaining) to the knowledge level (Mycin carries out medical diagnosis using a heuristic classification approach). Moreover, 6 Handbook of Software Engineering and Knowledge Engineering by showing that the heuristic classification model is generic, Clancey uncovered the principle of role differentiation, which has subsequently informed much knowledge engineering research [4, 21, 22]. Role differentiation means that it is possible to describe problem solving agents in terms of generic models, which impose specific problem solving roles on the domain knowledge. For example, domain structures in different application domains, such as diseases and book classes, actually play the same role (e.g. solution abstraction) when a heuristic classification model is used to describe the problem solving behaviour of a pulmonary infection or a book selection system. 2.2. Knowledge acquisition as modelling Wielinga and Breuker [23, 24] were among the first to apply to knowledge acquisition the lessons drawn from the work carried out by Clancey. In particular, they argued that the so-called bottleneck was caused by the fact that “the mapping between the verbal data on expertise and the implementation formalisms is not a simple, one to one correspondence”. Therefore, in developing the KADS methodology [4], they proposed an approach in which expertise modelling and design are clearly separated. First, “in an analysis stage, the knowledge engineer develops an abstract model of the expertise from the data ...this model is (then) transformed into an architecture for the KBS” [25]. Thus, they made the case for the development of conceptual modelling frameworks, addressing the issue of characterizing expertise at a level independent from implementation. A similar approach was also taken by my colleagues and I working on the KEATS project [26], in which we distinguished between modelling “overt behaviour” (i.e. understanding problem solving behaviour) and “internal representation” which was concerned with the realization of this behaviour on a computer system. Other researchers [21, 27] set to the task of putting the role differentiation principle into practice, by developing knowledge acquisition tools based on taskspecific, but application-independent problems solving models. Of course there are differences between the approaches followed by all these researchers. Nevertheless, it is possible to group all these efforts around a common paradigm, which considers knowledge acquisition and engineering as a modelling activity. Below I list the main features of the modelling paradigm. • Knowledge engineering is not about cognitive modelling (i.e. ‘reproducing’ expert reasoning) but about developing systems which perform knowledgebased problem solving and which can be judged on task-oriented The knowledge modelling paradigm in knowledge engineering 7 performance criteria. • There are enough similarities between classes of applications, which make it possible to build generic models of problem solving. • Knowledge acquisition should not be characterized as a process of mapping expert knowledge to a computational representation, but it is a modelbuilding process, in which application-specific knowledge is configured according to the available problem solving technology. In the words of Ford et al. [28], “The mining analogy notwithstanding, expertise is not like a natural resource which can be harvested, transferred, or captured, but rather it is constructed by the expert and reconstructed by the knowledge engineer”. • It is useful to describe such a model of problem solving behaviour at a level which abstracts from implementation considerations (the knowledge level). This approach has the advantage of separating problem solving from implementation-related issues. • Given that i) knowledge acquisition is about model construction and that ii) models can be application-generic, it follows that these generic models can be used to provide the interpretation context for the knowledge acquisition process (i.e. the knowledge acquisition process can be model-based). In this scenario, much of the knowledge acquisition task can be reduced to acquiring the domain knowledge required to instantiate generic problem solving roles [29]. Table 2 characterizes the modelling approach according to the same template used to describe the mining approach. In particular, the table shows a paradigm shift from an implementation-oriented to a knowledge-oriented view of knowledge acquisition. Multiple levels of descriptions are introduced and, as a result, the choice of implementation-level formalisms becomes less important. The knowledge categories are characterized at a conceptual, rather than computational level. The goal is no longer to emulate an expert by means of some kind of ‘expertise mapping’, but to acquire the domain knowledge required to configure a generic problem solving model. Thus, the knowledge acquisition process becomes less amenable to the cognitively-motivated criticisms aimed at the mining approach. Researchers subscribing to the modelling approach no longer make claims of building rule-based cognitive models of experts and acquiring expertise by ‘direct transfer’. The cognitive paradigm underlying the modelling approach can be 8 Handbook of Software Engineering and Knowledge Engineering characterized as a pragmatic one, which is based on a functional view of knowledge. Knowledge is functionally described as whatever an observer attributes to an agent to explain its problem solving behaviour [2]. A knowledge-level description characterizes knowledge neither as a symbol-level data structure, nor as ‘stuff’ in the mind of an expert: it is simply what enables a knowledge-based system to handle complexity. Such knowledge can be represented in different ways e.g., as plain text, in some logical formalism, as a set of rules, but the representation should not be confused with the knowledge itself (i.e., the competence expressed by a knowledge model of problem solving is not a function of the chosen representation). The advantage of this approach is that it makes it possible to characterize knowledge modelling as a distinct technology, which focuses on knowledge-based behaviour per se , independently of cognitive or machine-centred biases. In other words, as researchers in knowledge modelling (and I dare say, in knowledge engineering) we are not interested in transferring expertise, we are interested in building models of intelligent behaviour. Knowledge Categories Differentiation is driven by generic knowledge roles KA Methodology Acquisition is driven by a knowledge-level model of problem solving, which is independent of the chosen computational platform Levels of Descriptions Multiple (e.g. knowledge vs. symbol level) KA Paradigm Model construction Cognitive Paradigm Functional view of knowledge Reusable Components Generic task, generic problem solving model, generic domain model Table 2. Characterization of the modelling approach The adoption of a knowledge modelling paradigm introduces a number of important research avenues. • The emphasis on knowledge-level modelling and the separation between knowledge-level and symbol-level analysis opens the way to structured development processes in knowledge engineering, characterized by distinct analysis and design phases, each supported by different languages, tools and guidelines. In the next section I will show how these ideas have shaped The knowledge modelling paradigm in knowledge engineering 9 the CommonKADS methodology, which relies on multiple models to break down the complexity of a knowledge management or engineering project. • We have seen that Clancey’s analysis uncovered the existence of generic, knowledge-level models of problem solving. This raises interesting issues with respect to both the theory and practice of knowledge engineering. From a theoretical point of view, interesting questions concern the space of these reusable models, “What are the main classes of reusable models?”, “Is it feasible to hypothesise that future knowledge engineering handbooks will provide comprehensive lists of reusable models?”. To answer these questions we have to devise sound typologies of reusable problem solving models, based on clear theoretical foundations. The engineering questions concern the use of these models in a specific project: “What tools do we need to support model-based knowledge acquisition?”, “Is it feasible to imagine automatic configuration of these reusable models for specific domains?”. In Section 5 I will briefly discuss ongoing work on the IBROW project, which is developing a range of advanced technologies to support the specification and the reuse of library components. Thus, an important aspect of the knowledge modelling paradigm is that it creates a framework in which new research issues can be addressed, which concern robust knowledge engineering. In other words, while knowledge-based system development used to be seen as “an art” [16], the adoption of a knowledge modelling paradigm allows us to construct what can essentially be seen as a practical theory of knowledge engineering: an enquiry into the space of problem solving behaviours which are of interest to researchers in knowledge engineering. The domain is intelligent problem solving, the approach is knowledge-level analysis. In the next section I will show how these ideas have been used to devise a comprehensive methodology for knowledge engineering and management. 3. The CommonKADS Methodology The modelling paradigm in knowledge engineering informs many approaches to knowledge-based system development, e.g., KADS [4], CommonKADS [6], VITAL [30, 31], Mike [32], Generic Tasks [22], Components of Expertise [33], Role-limiting Methods [21] and Protégé [34]. Clearly, there is not enough room in this paper for a detailed survey and the reader is referred to the given references for more details on the various approaches. Here I will briefly illustrate the main features of the CommonKADS methodology, which is the best established and the 10 Handbook of Software Engineering and Knowledge Engineering most comprehensive of all current approaches to knowledge engineering. The CommonKADS methodology has evolved over almost two decades. It was originally conceived as the KADS approach to knowledge acquisition [23] and it has been extended, adapted and revised over the years until its most recent formulation by Schreiber et al. [6]. The Common KADS methodology provides a comprehensive framework in which both ‘traditional’ knowledge engineering projects (i.e., projects whose main goal is the development of a performance system) and ‘modern’ knowledge management projects can be situated. Indeed, an important contribution by the CommonKADS authors is the provision of an integrated framework for knowledge management, which also encapsulates knowledge engineering activities. In [6] the authors explicitly stress this point, by stating that “knowledge engineering has several different applications. The construction of knowledge systems is only one of them, albeit an important one. In all applications of knowledge engineering the conceptual modelling of knowledge at different levels of details is a central topic” (pp. 82). In other words, Schreiber et al. emphasize that the role of a methodology based on the knowledge modelling paradigm is to provide a suite of techniques to support knowledge analysis in an organization, in a wide range of scenarios. In some cases, the goal may be to develop and integrate a knowledge-based system in the organizational workflow; in other cases it may be simply to develop a competence model of an organization, or to provide a knowledge management solution to support knowledge creation and sharing. An important aspect of the CommonKADS methodology is its reliance on multiple models to address the complexity of a knowledge management or knowledge engineering project. These models are briefly illustrated in the next sub-