Knowledge Systems Laboratory
Computer Science Department
Stanford University
The panel will focus on the following two questions:
We adopt the premise that it would be worthwhile for the KR&R community to converge on a formal definition of what it means by an "ontology". We are particularly interested in precisely relating ontologies to symbol-level artifacts and to the roles they play in KR&R so that ontologies can be distinguished from arbitrary knowledge bases, the formal properties of ontologies can be studied, reasoning methods specifically for ontologies can be developed, effective tools for building and maintaining ontologies can be developed, etc.
The following comments represent some of my personal views on these issues.
What kinds of sentences can be included in an ontology? There does not seem to be any precise way of differentiating between sentences that are "definitional" and sentences that express "contingent facts". Nor does there seem to be any rationale for prohibiting the inclusion of any sentence in an ontology that restricts the interpretations of the symbols occurring in it since any such sentence can be considered to contribute to the specification of the "meaning" of those symbols. It therefore seems that any sentence that is not a tautology and that is satisfied in the conceptualization being represented by the intended interpretation is suitable for inclusion in an ontology.
If an ontology can be any set of sentences and definitions, then what is the significance of the "ontology" notion? I suggest that the significance of ontologies is not in what they are, but in the role that they play in representing knowledge. In particular, I suggest that we consider an ontology to be an integral part of a declarative knowledge representation language. That is, consider a declarative knowledge representation language to provide a syntax, a set of inference rules, a vocabulary of non-logical symbols, and an ontology that restricts the acceptable interpretations of the symbols in the vocabulary.
If a representation language is assumed to include an ontology, then any knowledge base expressed in a given representation language will implicitly include the axioms and definitions from the language's ontology. Also, when agents agree on the representation language to be used in an upcoming interaction, they will have agreed on an ontology and therefore on a set of mutually shared assumptions about the vocabulary to be used in the interaction.
"Domain-specific" representation languages typically include a domain-specific vocabulary and therefore can naturally be considered to include an ontology. Note that even "domain independent" representation languages often include a vocabulary of relations and functions and have implicit ontologies associated with them. For example, frame languages typically includes vocabulary such as "subclass of", "instance of", "slot value type", and "slot cardinality"; and implementations of those languages assume a set of axioms (e.g., stating that "subclass" is transitive) regarding that vocabulary. Indeed, the entire frame language in our Ontolingua system [Farquhar, et al 96] is defined as an ontology added to KIF [Genesereth & Fikes 92], and KIF itself has been redefined as a core language augmented with ontologies for sets, numbers, lists, etc.
A primary difference between ontologies and arbitrary knowledge bases is that there is a much greater emphasis in ontologies on useability in multiple tasks and multiple situations since the ontology is intended to be part of a representation language. Thus, for example, the emphasis is on describing classes of objects rather than individuals and on representing general properties of relations and functions rather than on describing specific situations.
Much has been said about the need to develop reusable encoded knowledge in order to enable the development of large scale intelligent systems. (See, for example, [Patil et al 92].) Ontologies are intended for multiple uses and are therefore an appropriate focus for research on techniques for knowledge reuse.
Adopting reusability as a primary goal for ontologies has a significant impact on the tools and methodologies that are needed for ontology creation and use. For example, developers need to make their ontologies accessible and understandable to a community of use. So, new techniques are needed for translating ontologies between representation formalisms and for describing the competency of an ontology. Also, when knowledge is encoded specifically for use throughout a community, one would expect there to typically be involvement by that community in the encoding process. So, tools that support collaborative development and evolution of ontologies would appear to be important in achieving desired levels of reusability.
There is an apparently inherent tension in reuse of represented knowledge (as there is in the reuse of any software) between level of detail and breadth of applicability. The knowledge that may be appropriate to assume as part of the representation language will vary dramatically from application to application. Thus, even though it is very appealing to have large general-purpose ontologies to include in our representation languages, it seems critically important for ontologies to be available in small composable modules so that the knowledge that is appropriate to assume for a given use can be assembled. A key research challenge, then, is to address the tension between depth and breadth in knowledge reusability. A promising approach to meeting that challenge is to develop both broadly applicable ontologies containing "common sense" knowledge that can be included in general-purpose representation languages and a capability for augmenting that knowledge by assembling composable modules retrieved from online libraries.
M. Genesereth and R. Fikes; Knowledge Interchange Format, Version 3.0 Reference Manual; Technical Report Logic-92-1, Computer Science Department, Stanford University, Stanford, CA, 1992.
R. Patil, R. Fikes, P. Patel-Schneider, D. Mckay, T. Finin, T. Gruber, R. Neches; The DARPA Knowledge Sharing Effort: a Progress Report; in Proceedings of the Third International Conference on Principles of Knowledge Representation and Reasoning; Cambridge, Massachusetts; October 25-29, 1992. Also KSL Technical Report KSL 93-23.