next up previous
Next: CONCLUSION Up: Describing Reusable Problem-Solving Methods Previous: Discussion


Our representation of P&R is based on the Yost's description of the VT task, which corresponds to the resolution of the problem by an expert. To validate our approach, we implemented it in Loom. Our system generated results similar to the test case presented in Yost's document.

For representing VT in our system, we reused the ontology of Motta [Motta et al., 1994] represented in OCML, which is a knowledge-level modeling language used for the Vital methodology. However, unlike our approach, the OCML's data do not represent the assumptions about the knowledge roles of P&R, such as the restrictions on the nonexistence of cycles in the parameter dependency relations, the cardinality of the relations and so on.

Fensel [Fensel, 1995] also copes with the problem of clarifying the assumptions of PSMs. He analyzes how differences in the assumptions about P&R can generate differences in the results and in the efficiency of the problem-solving method. There, the assumptions identified are described in an informal language. Our work, in contrast, presents a formal representation of the assumptions, which are grouped in the method ontology. This ontology allows a knowledge engineer to comprehend one specific implemented problem-solving method, and, at the same time, to abstract from the implementation details. Our method ontology is not a complete description of the assumptions, but it is a starting point to understand and to reuse the method.

Another work concerning the problem of specifying PSMs is the conceptual organization of assumptions proposed by [Benjamins and Pierret-Golbreich, 1996]. According to this proposition, the method ontology corresponds to the epistemological assumptions, which refer to the knowledge required by the PSM. In our future work, we intend to extend our approach for indexing PSMs by using the other kinds of assumptions proposed by [Benjamins and Pierret-Golbreich, 1996] such as teleological and pragmatical assumptions.

The KADS methodology proposes a library of inferences [Aben, 1995], in which the generic and reusable inferences are characterized in terms of operations on the base ontology. The base ontology abstracts from specific domains and provides a set of primitives to model more specific ontologies. The KADS base ontology contains the epistemological primitives of KL-ONE language [Brachman and Schmolze, 1985], such as concepts, instances, attributes, and so on. KADS's level of description makes the inferences more generic and reusable. In our work, we use the same idea of describing the inferences in terms of an ontology previously defined, but we adopted another perspective. Instead of defining the inferences in terms of our base ontology (KIF and frame ontology), we characterized the inferences in terms of operations on the method ontology. This ontology is more specific than the base ontology. Consequently, the inferences are less reusable across different PSMs. However, even if the inferences are more specific, the PSMs as a whole are made more reusable, because this level of description allows us to understand and integrate the methods for different application domains.

Ontolingua is used originally for constructing ontologies that minimize the ontological commitments and are task-independent. To minimize the ontological commitments, an ontology should make as few claims as possible about the world being modeled [Gruber, 1993a]. Following this principle, the VT Ontology [Gruber et al., 1994] represented in Ontolingua does not model all the needed knowledge roles used by the method. Motta [Motta and Zdrahal, 1995] shows that the VT task ontology of Ontolingua lacks and misinterprets the knowledge used by P&R. Thus, VT Ontolingua makes the method less reusable. In contrast, we are interested in developing ontologies that unveil as much as possible about the knowledge used by the method. Our method ontologies are specific, but they are useful to understand a method and can contribute to its reusability.

Another work whose research's goals are in line with ours is the Protégé-II system [Musen et al., 1994], which defines a complete environment for modeling knowledge-based systems from reusable components. Our ontology is similar to Protégé-II method ontology [Gennari et al., 1993], which specifies the data requirements for a particular method. The significant advantage of Protégé-II is the automatic generation of knowledge acquisition tools specific for each PSM. Our approach differs in that our method ontology allows to define declaratively the inferences of the method. From this perspective, we can obtain definitions of inferences that are precise and, at the same time, separated from the implementation details. We believe that the method ontology can serve as a specification for the construction of task-specific knowledge acquisition tools that can be easily adaptable for different domains, but we have not yet explored this aspect.

next up previous
Next: CONCLUSION Up: Describing Reusable Problem-Solving Methods Previous: Discussion

Eliana Coelho
Wed Sep 4 15:57:17 EDT 1996