1 2 3 4 5 6 7 8

the one to be used.

2.1.6

Contextual conditions

The characters are generally dependent from one another. Rather than distorting the reality with an independence hypothesis (too rarely verified), it would be better to get the best of the information brought by these relations.

Co-existence and exclusive relations appear frequently in descriptions. They give respectively a condition of presence or absence of a character depending on the "context" made by other characters. For example, in the Mammals' classification, there is the fact that some have a placenta and others don't (distinction between Placentalia and Aplacentalia); it is obvious that this should not be observed on male individuals; if a bull is described, it is "not relevant" to know if it is gravid, or to know the number of dugs carried on its udders. One can notice that, as for the "value «unknown»", it is non-sense to speak about the "value «not relevant»" unless one needs to fill empty boxes in data matrices : if the sex of the bull is male, this carries all the information related to the "non relevance" of the gravid character, and shows the general fact of exclusion between masculinity and pregnancy. Nature is so made.

One can easily imagine co-existence relations, when the presence of a character is deduced "automatically" from the context. Such relations are sometimes perceptible only by specialists, and that constitute their expertise. We will take a real example from the diagnostic of plant diseases : the expert notes a withering of leave extremities and will focus on the most inexpected part of the plant (the collar at the base of the stem) to see if there is not a "canker" or a tumor that stops sap circulation. He thus uses this way a co-existence relation, and more precisely here a cause to effect relation.

Because of the variety of nature, dependencies between characters are not absolutely marked. For instance, some witherings are not due to a collar canker, and Nature does not like "rules" or "laws" without exceptions. Thus it is important, not only to consider dependency relations, but also to specify their applicability conditions, that is to say exceptions and related contexts.

In a lot of situations, a part that should be theoretically observable is not; or on the contrary, a local description is only possible under some conditions. This can be turned into contextual rules, for instance : if the dog is nasty, then don't observe its teeth; or : if the bird is flying, then describe the marks that are under its wings. These conditions are common sense knowledge and can be well used to guide cleverly the observations.

2.2

Structured representation with a descriptive model

For a given domain, the descriptive model is created by the expert. He must represent all what is observable as a structured scheme.

Furthermore, the major goal of the descriptive model is to be transposed in an observation guide to help the user to describe. It must be a way to translate without constraints the set of mechanisms or observational logics shown up precedently. So it is a representation of the set of all the observable knowledge, well suited for acquiring the observed knowledge.

The descriptive model can take several aspects equivalently, depending on the target user. In depth, it is represented under a data processing aspect adapted to observable knowledge bases; one can find objects like "frames", lists, matrices, rules, pictures etc, written with a syntax that translates as exactly as possible the different observation mechanisms and the "background knowledge" of the domain. This form is not to be read by the naturalist; it is only a technical representation, used as input and/or output to the different modules of description treatments.

The processing model must of course follow a formalism that can be transcribed immediately to a mathematical plan, in order to be able to process knowledge with symbolic data analysis programs, inductive ones or others. Our individuals (or subjects) are represented as boolean