symbolic objects (Diday, 1991) of type "synthesis objects" because of the use of the iteration logic that introduces "hords" in our descriptions. Furthermore, we introduce the notion of "composite objects" (Conruyt et al., 1992) to deal with the compositional logic as explained above.

The descriptive model must also be presented in a more practical and synthetical way for the naturalist, specialist of his domain who elaborates and updates it. Its structuration is logically depicted by a tree or a graph showing parts and subparts with their own relations and characteristics. The "object" manipulation (in a computer science meaning) to create, modify, move, associate pictures to them etc, is better made in a graphical way with interactive tools that are easily used by biologists, who are not computer programmers.

One last issue, perhaps the most important practically, allows to present the descriptive model like a real observation guide : we called it "questionnaire" (Manago et al., 1992) in our developped applications because it is put in the hands of the user, under a flexible but logical navigation form between different input screens. Each screen (called a "card" because of the HyperCard tool which is used) corresponds to the acquisition of a local description, matching exactly the equivalent part in the descriptive model. Notice that the descriptive model can provide some gradation for answers' accuracy, for instance giving intervals of numerical values, and at last permit to use the answer "?" to express the absolute uncertainty. This is essential for real descriptions, where context or particular events do not allow complete descriptions.

The final descriptions, the consistency of which is ensured by complying to the descriptive model and the completion verified at the end of the data entry, can be presented in different ways too. The initial form is the one of the filled questionnaire. It can be imported again to bring corrections or further descriptive informations. But it is sometimes useful to be able to visualize a description as an instanciated subgraph of the descriptive model. This form allows to highlight the underlying structure of the description that is somehow lost sight during the questionnaire navigation. In fact, these two forms complement each other and the user must be given the possibility to switch easily from one to the other. Moreover, it is nearly necessary to be able to present the user with descriptions under a natural langage text form, as it always existed. It is not difficult here to offer a choice of several target langages. At last, for best efficiency and homogeneity, input descriptions are recorded with the same syntax representation as the descriptive model. Therefore, observable and observed facts benefit from the same well adapted formalism, that allows to use them jointly and give more consistency and power to the programs that treat them.

We will not detail here the different technical solutions which allowed us to represent the different observational mechanisms. "Frames" are applied as a structure basis. They are "objects" with their own slots (characters or attributes). Each slot can take one or several possible values (in a list, possibly in a hierarchy for nominal classified values; in an interval for numerical values); once valued, each slot expresses a descripted character or a feature. When objects correspond to subparts (but not to points of view), their stated absence is recorded as significant. The specialisation and particularisation mechanisms are expressed by "class" instanciation (in a computer science meaning) with inheritance. The iteration mechanism is delt using "variables" and a first order logic. At last, the context conditions are represented as rules or demons.

It is now possible, by using AI knowledge representation methods, to formalise such complex descriptions that are required from the "truth" of nature, without transposition bias, without resort of subjectivity, and with as little loss of information as desired.

There is a good way to make sure that obtained descriptions satisfy our quality criteria. One has only to compare such descriptions produced under their natural langage form, with those directly written by specialists. It is then very easy to estimate drawbacks of ones and others; this is independant of the fact that "conform" descriptions (to the descriptive model) have the great advantage to be comparable to each others and easily mobilizable.