I stumbled on a rather comprehensive taxonomy of architecture which the Getty is putting together.
This classifies houses, for example:
- by function
- by construction technique
- by occupants (including haunted)
- by designer/builder
- by form (in terms of massing, plan, or roof orientation)
This could provide a reasonable standardized vocabulary to build from, especially since it is "off the shelf" and terms are documented. It still doesn't get to the hard multiple classifications to multiple visual representations question, though. I'm personally still puzzled about how to best allow broad, top-down control without getting into the complexities of a full AI-like system.
As an end point, we have "visual representations" with at least the following important properties: some instances are selected, from a pool of candidates meeting all requirements these instances are placed, either at a specific geographic location, or relative to other resources each instance has certain properties which are fixed within a particular representation, and some which are variable of the variable properties, some are defaulted and some are data - driven of those which are data-driven, some are specified exactly and some are probabilistic
At the starting point, we have the real world (or some past or alternative future), which has a multitude of objects, and many disciplines devoted to cataloging and explaining these objects in various sometimes overlapping, sometimes conflicting classification systems. We would like to make use of the expert's classifications to drive some of the properties of our visual representation, without having to resolve conflicts in any elaborate way.
Perhaps the simplest means of doing this is to allow only a single classification, which controls all visual properties within its domain. This is what we did with the West Lake project, restricting architects and urban planners to a single typology of buildings, and this is generally what most GIS systems do, mapping attributes to colors or symbols.
A more complex, but still controllable, setup might be to allow multiple classifications, but each uniquely mapping to a single visual simulation attribute. So, a zoning map could drive buildings by use, while a separate real estate value model could be mapped to building height, for example. If the real estate value model would predict a taller building than allowed by zoning, that discrepancy would be up to the user to catch.