Knowledge Representation: Covlan

Medin and Ross (1990) report that experts qualitatively represent large knowledge bases in memory. Doing this for visual information requires an ontology of visual primitives. I am designing Covlan, a language to describe visual analogs. It can represent diagram-like images, relationships between them, and changes to them. I will also design a non-visual language, based on the SBF modeling language, but the details of this non-visual representation will not be important for my theoretical claims.

I will make choices of what to put into Covlan based on the following constraints:

1. I have some data from experiments run by Dr. David Craig. The details of these data can be found in appendix A. Diagrams made by experimental participants in the Craig study provide information about what abstractions are useful for representing the systems in question. I will say more about these data in the theory evaluation section.

2. Primitives from research that suggest a visual vocabulary such as geon theory (Beiderman & Cooper, 1991). As this is a cognitive theory, psychological research will also constrain it.

3. Certain choices in what will go into the theory will be determined by what is needed to get the program to function correctly. For example, to represent the fortress/tumor problem's solution, I needed to break the solution procedure into steps, so that they could be transferred. The steps chosen should be general enough such that they might be useful for other problems. Thus the choices of transformations, for example, were constrained by what was needed. I will make similar decisions with the examples from the Craig data.

Covlan consists of the following kinds of entities: S-images, transformations, elements (primitive and complex), miscellaneous slot values, and relations. I will describe these in turn.