Integration issue in knowledge acquisition systems

• Authors: K. Morik
• Venue: ACM SIGART Bulletin - Special issue on knowledge acquisition
• Year: 1989

Ever since knowledge acquisition systems have been applied to real world problems, the issue of integration has become important (Gaines 88). Most often, the following types of integration are considered:• Systems are to be integrated: the knowledge acquisition system is more closely linked with an expert system shell (Eshelman et al. 87), or a database system is linked to the knowledge acquisition system in order to read data of a domain.• Various sources of knowledge are to be integrated: text files, data files, statistics, rules and facts all contain knowledge about a domain and should be handled by the same system (Gaines 88).• The represented knowledge of various experts is to be integrated either into one consistent domain model or into a model which shows the conflicting views of the domain (Shaw 88).• Diverse knowledge sources with their respective representations are to be integrated, e.g. a taxonomy of domain concepts, possible values of attributes, well-formedness conditions of facts and rules.• Diverse tasks of knowledge acquisition are to be integrated: declaring epistemic primitives, defining concepts, adding facts and rules, deducing new facts from rules, dealing with inconsistencies, changing the terminology, changing facts and rules, grouping facts or rules together, presenting possible operations, showing views of the represented domain model, indicating consequences of an operation to the user. A particular topic of task integration is the integration of machine learning into a knowledge acquisition system.In this paper, we want to discuss only the last two integration problems and present the solutions we implemented in the BLIP system, currently under development at the Technical University of Berlin1. First, we give a short overview of knowledge acquisition tasks and indicate which tasks some prototype systems can handle. As we will see, the integration of machine learning into knowledge acquisition is still not yet well achieved. Second, we discuss the integration of tasks. In particular, the integration of machine learning into knowledge acquisition is discussed in some detail. The architecture of BLIP illustrates the paradigm of cooperative balanced modeling of both system and user. Third, we investigate the integration of knowledge sources. The integrity between diverse knowledge sources, the propagation of consequences of the operation to all relevant knowledge sources, and the interpretability of a component's results by other components are the 3 issues there. In BLIP, the knowledge needed for the learning task is also integrated into the domain knowledge. This gives BLIP the power of closed-loop learning (Michalski 87). Fourth, we describe the integration of BLIP's learning into knowledge revision in more detail.