Jeffrey M. Bradshaw, John H. Boose, David B. Shema

Research & Technology, Boeing Computer Services

P.O. Box 24346, M/S 7L-64, Seattle, WA 98124; (206) 865-3422; jbrad@atc.boeing.com

Douglas Skuce, Timothy C. Lethbridge

Department of Computer Science, University of Ottawa

Ottawa, Ontario, Canada K1N 6N5; (613) 564-4518; doug@csi.uottawa.ca

From this perspective, a model is not a ‘picture’ of the problem, but rather a device for the attainment or formulation of knowledge about it (Kaplan, 1963). Indeed, sometimes the most important outcome of the modeling process may not be the model itself, but rather the insight we gain as we struggle to articulate, structure, critically evaluate, and agree to it (Moore & Agogino, 1987). Thus, the value of such an effort derives not simply from a final ‘correct’ representation of the problem, but additionally from our success in framing the activity as a self-correcting enterprise that can subject any part of the model to critical scrutiny, including our background assumptions. We must ask ourselves not only, How do we know the model is correct? (every model is an incorrect oversimplification); but also, How useful is the model (and the modeling process) in facilitating our understanding of the domain?

Concern for these issues has become the rallying cry of a number of researchers in conceptual modeling and knowledge acquisition. Recent work in knowledge acquisition has emphasized that the creation of knowledge bases is a constructive modeling process, and not simply a matter of ‘expertise transfer’ or ‘knowledge capture’ (Ford & Bradshaw, 1992). Mylopoulos (1991) describes the field of conceptual modeling as follows:

Unfortunately, until recently developers of modeling and knowledge representation tools have paid scant attention to the needs of the kind of users who are building a model with the intent of coming to understand some domain. For example, traditional data modeling tools introduce assumptions about the way conceptual schemata will be realized on a physical machine, which constrain and confound domain experts preferring to leave these considerations to others. Furthermore, the need of such users to model the ‘real world’ drives a requirement for much richer representations than modeling tools typically afford. Knowledge representation tools, on the other hand, almost always introduce the assumption that the resultant knowledge bases will be directly usable for some computational task. Considerations of efficiency and performance inevitably lead to limitations in the richness and flexibility of the knowledge representation. No less important, there are few modeling tools that have been optimized for usability. Tools are cumbersome, unnatural, and difficult to use without specialized training and extensive experience. The comprehensibility of textual and graphical notations chosen is rarely considered. As a result, many potential users abandon modeling tools altogether and use simple graphical drawing tools instead.

Elsewhere we have discussed our ideas about how the application of certain general principles of knowledge representation and user-interface design can make our tools more usable and extensible by non-experts (Bradshaw, Ford & Adams-Webber, 1991; Bradshaw, Holm, Kipersztok & Nguyen, 1992; Skuce, 1991b). In this paper, we discuss another aspect of usability for tools intended to promote communication as part of design rationale modeling: support for design and sharing of reusable ontologies.

Recently, a number of researchers have emphasized the importance of defining common ontologies to specify the concepts and terms upon which knowledge-based systems depend (e.g., Gruber, 1991, 1992; Lenat & Guha, 1990; Neches, Fikes, Finin, Gruber, Patil, Senator & Swartout, 1991; Skuce and Monarch, 1990):

The desire for large-scale knowledge sharing and reuse across different tools and applications provides a strong motivation for common ontologies. The desire for large-scale knowledge sharing and reuse across different tools and applications provides a strong motivation for common ontologies. Careful design of the ontology and the equally careful choice of terminology also benefit developers and users of the system. Our experience confirms that terminological confusion breeds conceptual confusion (and vice versa). Such confusion tends to become more probable the ‘higher’ one goes in a concept hierarchy. For example, it is much easier to reach consensus on the meaning of the term "pencil" than on terms like "object," or "entity." Hence the most general ‘ontological’ concepts are the ones most in need of standardization, yet also the most difficult to agree on. Standardization has two aspects:

At present, agreement at such high levels is probably unlikely. Our approach is therefore two-pronged:

We are currently developing ontologies that contain concepts useful in formally describing our own work in design rationale. These ontologies are being developed within a modeling framework called DDUCKS (Decision and Design Utilities for Comprehensive Knowledge Support; Bradshaw, Boose & Shema, 1992; Bradshaw, Fulton, Holm, Boose, Kipersztok & Nguyen, 1992). To assist in the definition of ontologies, we have integrated DDUCKS with an enhanced version of CODE4 (Skuce & Lethbridge 1992; Lethbridge, 1991). CODE4 provides a rich, paradigm for the definition of knowledge level concepts. A collection of integrated tools support the important and frequently overlooked aspects of conceptual, ontological, and terminological analysis (Skuce 1991b, Lethbridge & Skuce 1992b) . We are developing extensions to the representation to allow the system to make use of additional inferencing and representation facilities similar to those found in Sowa’s (1991) conceptual graphs and Gaines' (1991) KRS, which interpret taxonomic and entity-relationship structures in terms of typed formal logics. A first order logic system and a simple natural language system allow various types of syntactic and semantic checks to be performed, if desired. A comprehensive lexicon allows references to concepts to be automatically maintained and quickly accessed. We emphasize the importance of comprehensive lexical support so that terminology can be carefully chosen and subsequently controlled. Concept libraries and default inferencing mechanisms can be augmented by users employing graphical views and an integrated scripting and query language.

We suggest that a successful approach to the development of sharable ontologies for design rationale representation involves the following elements:

The first four of these elements represent increasingly specialized knowledge. We anticipate that research on very general ontologies will eventually result in a consensus, although this may be many years away. More specific ontologies use knowledge from more general ontologies, therefore the closer to consensus the higher-level ontologies can be brought, the more effective will be the sharing of lower level ontologies.

The speed with which consensus can be reached is largely dependent on the fifth element above: a mechanism for sharing. Ontological research is currently shared very little, and where sharing takes place, it is usually either a) in the form of paper reports that take time to distribute and get outdated rapidly, or b) among researchers using some specific piece of not-widely-distributed research software. To facilitate development of ontologies it will be necessary for determine how divers software tools can be facilitated to exchange their knowledge. It will also be necessary to determine conceptual formats, i.e. exactly what pieces of information are necessary to convey to accurately communicate ontological concepts among a community of researchers.

An important area for of work concerns the most general categories that ontologies can have, notions like: thing, object, entity, property, attribute, event, process, state, situation, collection, relation, etc., to name a few favorites. At the moment, everyone uses these concepts and terms, but a) they probably use them in very different ways and b) they cannot tell anyone else what they mean by them. A well known example is the top of the Cyc ontology, which we find difficult to understand from the published descriptions. (Skuce, 1991a).

We believe that such notions can best be clarified by studying linguistic and psychological data. An important AI ontology based on linguistic research is the "Penman" ontology (Bateman, Kasper, Moore & Whitney, 1990). This ontology, derived from Halliday's studies of English, is well-documented, with reasonable if minimal descriptions of each of some 200 categories. Another linguistically-based ontology is Miller’s WordNet (Miller 1990). One of us (Skuce) has been working on an empirical approach to very general ontologies for a number of years. By the end of the summer, he hopes to have an initial proposal ready with approximately fifty categories of the kind listed in the previous paragraph.

In conjunction with a number of research groups interested in knowledge sharing issues, we are participating in an effort to develop specific ontologies for particular domains of interest. For example, in one project we have been developing ontological primitives to represent knowledge of the airplane design and manufacturing enterprise (Bradshaw, Holm, Kipersztok & Nguyen, 1992). In another project, we are interested in representing medical decision making knowledge (Bradshaw, Chapman & Sullivan, 1992). Others have begun projects that will result in ontologies relevant to disciplines such as process planning, software engineering, circuit layout, and device modeling (see Gruber, Tenenbaum & Weber, 1992).

While the domain ontologies are designed to be relatively neutral with respect to a particular rationale modeling framework, the modeling ontologies are meant to capture a particular theoretical or practical point-of-view. In other words, the modeling ontologies characterize the roles that domain concepts play within a particular modeling framework.

We are performing an analysis of the knowledge acquisition, decision support, and design rationale tools we have developed over the past several years to understand the concepts and assumptions, implicit and explicit, they embody. We have found that the process of developing a formal modeling ontology for our tools has greatly increased our understanding of them. As we increase our understanding of the ontological commitments of our own framework, we plan to develop and evaluate similar modeling ontologies derived from other perspectives. The elements of Lee and Lai’s DRL (1991), for example, could be represented quite naturally within a modeling ontology. MacLean et al’s (1991) questions, options, and criteria could be formally represented in the same way. The separation of domain from modeling ontologies will allow us to begin comparing and contrasting different approaches to design rationale for the same application.

Practical concerns about the engineering of large systems have led to the development of techniques for creating and operating on multiple partitions of the knowledge base (Guha & Lenat, 1990; Gruber, 1992). Each partition, or theory, constitutes a set of axioms tailored for a particular domain or purpose. Each theory is associated with a context (McCarthy, 1987) that describes, among other things, the set of assumptions made by the theory. We are exploring a layered approach to theory and context management within DDUCKS. Starting with any layer in the system, a user can produce a set of tools, models, ontologies, and representations that imports selected assumptions and axioms into the layer below.

Gruber’s work on Ontolingua (1992) currently provides the most promising mechanism for sharing ontologies between different tools and formalisms. Ontolingua extends the knowledge interchange format (KIF; Genesereth & Fikes, 1991) defined by the DARPA knowledge sharing effort with standard primitives for defining classes and relationships, and organizing knowledge in object-centered hierarchies with inheritance. Ontolingua facilitates the translation of KIF-level sentences to and from forms that can be used by various knowledge representation systems (currently LOOM, Epikit, or Algernon). We are working with Gruber to define an Ontolingua interface for CKB, the CODE4 knowledge base file format (Lethbridge & Skuce, 1992a).

The goal is to use the Ontolingua project as a basis for processing and distributing the ontologies. As good examples become available, the Sharing and Reuse of Knowledge Bases subgroup of the knowledge sharing effort can coordinate evaluations and experiments. In this way, we can pool our results with those of projects with similar goals (e.g., Gruber, Tenenbaum & Weber, 1992).

In order to facilitate sharing of ontologies a number of key research issues must be addressed.

We have chosen our own positions on the above issues, and so have others. Now the issues have been raised we anticipate a lively debate. However, we hope existing formats for sharing knowledge will begin to accelerate the exchange of ontologies.

As we have noted, an important aspect of knowledge sharing is lexical, or terminological knowledge. In addition to its other features, CODE4 provides a high level of integrated support for lexical functions such as:

We hope that our initial efforts at developing ontologies and sharing mechanisms will spur other researchers to make similar efforts, with the goal of increasing knowledge sharing and reuse within the design rationale community.