The interrogation of databases of scientific and technical articles is often characterized by the search for documents which do not come from any given specialist domain, but from multiple domains. A real interrogation can be characterized as an original combination of keywords in databases where docu ments are indexed in the form of keywords. The problem, then, is to find the right documents, while making allowances for the databases own indexing logic for documents. However, current expert systems, which permit access to documents via rules expressing the relevant field structure, make it notoriously difficult to pass from one specialist do main to another. Finally, recent work in the sociology of science has analysed the progress of knowledge in terms of 'translations' operating between one network of knowledge or know-how and another. This work has led to the elaboration of word graphs known as 'Leximappe graphs'. It is suggested that these graphs and the logic of associa tions that supports them can be used not only to interrogate databases of scientific and technical articles but also to search for new combinations at the core of these articles—with a view to advancing knowledge.