Van Heijst et al. [van Heijst et al., 1996] discuss a number of ways to categorize and organize ontologies, and what role they play in the knowledge engineering process. In the categorization of Van Heijst et al., the PHYSSYS ontology is a knowledge-modelling ontology, while the OLMECO conceptual database schema would count as an information ontology derived from the former. They point to the need to explicate ontological commitments early in the process. As a method to modularize and organize ontologies, they suggest, first, to single out basic concepts (such as patient, disease, therapy) on the basis of `natural categories' of the field to construct some widely usable base ontologies; to specialize these concepts with respect to various relevant (here, medical) subdomains; and then add method-oriented extensions. These are steps needed in achieving modularity of ontologies, which is seen as a key principle in ontology library organization (see especially their section 3). In this section we consider the impact of our work in this regard.
There is in our opinion no doubt that modularity is indeed a key success factor to ontology library construction. In any large-scale application we face what Van Heijst et al. call the hugeness problem: the enormous amount of domain knowledge that is involved in expert tasks. However, as these authors point out, concepts involved come in different levels of generality, and this gives a handle on organizing an ontology library. This is clearly visible in the structure of the GAMES-II core library, and we have deployed a very similar approach, as depicted in Figure 1. What Van Heijst et al. call `generic concepts' is very akin to our reusable `super'theories. They claim that partitioning should be based on two considerations: (i) definitions are to be centered around available `natural categories' of concepts that belong together, and (ii) the number of theory inclusions must be kept to a minimum.
Although we are generally in agreement with these views, our PHYSSYS and OLMECO work offers some different perspectives as well as extensions. As we have argued in the previous section in relation to the work of [Gruber, 1995], minimization of theory inclusions to achieve minimum ontological commitment is a phrase that is in danger of simplifying the real picture in applications. Rather, we would phrase it as piecemeal ontological commitment: starting from (indeed) the minimal side, one needs to incrementally build up the ontological commitments until the right degree of commitment for the particular application is achieved. The organization of an ontology library must be modular in such a way that this can be realized.
In [van Heijst et al., 1996], the proposed `natural categories' are groups of concepts that naturally belong together, reflecting the social consensus of a certain expert community. We share the observation that such natural categories do exist (although they may be so self-evident to a community that they are implicit for outsiders), and that they provide a good handle for partitioning ontologies. This has also been a major structuring principle in PHYSSYS. In our case, we have called these natural categories viewpoints [Top & Akkermans, 1994], because our ontology organization has been based from the start on coherent categories of properties of the same class of real-world objects, rather than on categorizing different real-world objects. In modeling and simulation, engineers view the same object (say, a hospital heating system under design) sometimes as a collection of connected components, or as a collection of interacting physical processes, etc., depending on the type of information that is to be extracted or decisions that are to be made.
Talking about partitioning and modularity, naturally leads us to the question how ontological modules can be connected again to meaningful assemblies. Here, our work offers an important new extension. The standard mechanism for configuring ontologies is theory inclusion, as it is used in most ontology work including [van Heijst et al., 1996]. We have found in our applications that richer and more flexible means for linking ontologies together are necessary, that go beyond relatively simple inclusion, specialization and extension operations. To this end we have developed what we call ontology projections: connections between two different ontologies realized by a mapping, that is highly knowledge-intensive itself and therefore assumes the form of an ontology in its own right. A good example is our specification of the connection between physical process knowledge and mathematical theory concepts.
Finally, we want to emphasize that using explicit ontologies yields benefits for a much wider range of information systems than KBS only. The PHYSSYS ontology has provided the basis for the QuBA modeling assistant [Top & Akkermans, 1994] and its successor IMMS, the KBS 007 for automated model revision [Pos et al., 1996b, Pos et al., 1996a], and the OLMECO library of reusable mechatronic models reported in this article. The latter is, at least in its implementation, a conventional database. Explicit ontologies are helpful in two ways here. First, they support and even enforce a sharply defined conceptualization of the information in the system in a way that is natural to the user (some might perhaps want to view this as a formal and high-quality `data dictionary'). Highly important is the experience that ontologies are a great help in clarifying the many tacit and implicit aspects involved. Second, from the modular organization of ontologies the modular structure of the information system itself quite easily follows. This proved to be strongly beneficial in the OLMECO library work, resulting in a design and demonstrator system that was appreciated by end users. Hence, also in a conventional implementation setting, this approach leads to a knowledge-oriented system design that reflects the way users view their world during task execution.