How do interactive knowledge acquisition tools operate? Figure 3 shows the architecture of one system that integrates a wide range of knowledge acquisition methodologies. The system is written in Pascal and runs on the Apple Macintosh family of computers to provide a highly interactive and graphic knowledge acquisition environment. At the heart of the system is an object-oriented knowledge base in which knowledge is formally represented as a multiple-inheritance structure of classes, objects, properties, values, and relations. Such a structure generalizes the entity-attribute datasets used in several early knowledge acquisition systems and has proved both general and powerful in a variety of applications.
Figure 3 KSS0, an integrated knowledge acquisition system
The elicitation tools are based on Shaw's (1980, 1981) computer-based interviewing techniques extended through the use of graphical rather than numerical data entry. The visualization and direct manipulation of knowledge structures through the graphic and click-and-drag facilities of modern workstations is an important development at the user interface giving experts, knowledge engineers and clients improved access to the knowledge structures.
The visualization tools consist of an interactive interface to represent the abstractions derived from those cases in terms of hierarchical clusters using Shaw's (1980) FOCUS algorithm, and relational diagrams such as a non-hierarchical conceptual maps derived through principal components analysis (Slater, 1976, 1977). The objectives are to validate the raw domain knowledge and suggest further structure at a higher level through interactive topological induction (Rappaport & Gaines, 1988):
The group comparison tools consist of an interactive interface to represent the relations between the terminologies and conceptual systems of different experts, or experts and clients. The objectives are to determine the consensus, conflict, correspondence and contrast between different conceptual systems (Shaw & Gaines, 1988):
The inductive part consists in the derivation of constraints within the conceptual structures through logical entailment analyses (Gaines and Shaw, 1980, 1986; Quinlan, 1987; Cendrowska, 1987; Gaines, 1989a). The objective is to suggest further structure at a higher level that translates into class inclusions or rules in the expert system shell:
The generative part consists in the transformation of the knowledge analysis made by the previous tools into formalisms understandable by knowledge-based system shells such as NEXPERT (Rappaport, 1987a,b) and Babylon (Christaller, Primio & Voss, 1989):
The next section shows the use of the induction and export tools in KSS0 to create a computational knowledge base from entity-attribute datasets giving the required performance in an expert system shell. The final sections illustrates the use of the interactive interviewing and visualization tools in eliciting an adequate dataset.
gaines@cpsc.ucalgary.ca 19-Sep-95