Documents, like computer programs have a logical structure defined by syntaxical and semantical rules. This fact allowed the use of software engineering knowledge to define the foundations for representation and handling of structured documents. However, documents and programs are different in many aspects. In this thesis, we consider two characteristics of documents : Programs are intended to be runned while documents are intended for multiple purposes : storage, reading, exchange, etc. Each use corresponds to a particular kind of document defined by a document model. Translating a document to another one often requires difficult structure transformations, and the number of document models prevent the development of specific converters. We aimed our study on prototype languages which allow the description of objects by differential copy. We have adapted this principle for the representation of document models. In this way, document models are not longer isolated entities. Each of them have at least one link which another model. By successive linking, we collect the set of transformations to move from a model to another one and we infer the modifications to be done on documents. The second feature comes from the reading process. Like the execution of a program, the reading of a document causes either an action or a change in the receiver. The singularity of the reading process is that the result depends not only on the content of the document, but also on the reader's profile. Adapting documents to readers, or more generally to a certain context would be desirable. We realize the design of context-controlled documents by breaking the generating process in two stages: the design of a gauge which brought together principles common to the design of a set of documents, and an automatic phase of document generation depending on a client profile. A gauge is represented by an abstract data structure in which some parts can be selected among a set of candidate structures. In order to ensure the correctness of generated documents, we use a gauge model obtained by a semantic extension of the document model.