1 Introduction

The continually increasing quantity of textual data requires constant effort in order to update storage and access methods so that the totality of information is easily accessible. Scientific publications are no exception. Astronomy is a good example in view of the enormous mass of data collected by modern satellites and large ground-based facilities, and the numerous scientific articles which result from such data.

The Strasbourg Data Centre (CDS) has the role of collecting and organising different types of astronomical information (Egret et al. 1995; Genova et al. 1996). In particular, the CDS has the charge to build and offer on-line access to some bibliographic data from Astronomy and Astrophysics and its Supplement Series.

This article presents work ongoing at CDS on the use of an artificial intelligence technique applied to the classification of scientific articles. This technique is Kohonen's self-organizing feature map (SOM; Kohonen 1995). The way this particular type of neural network works will be described first. Then we will see how we have built a system appropriate for the classification, retrieval and consultation of documents. A description of the use of this search system, which we could call a bibliographic map, will conclude this article.

Information retrieval by means of ``semantic road maps'' was first detailed by Doyle (1961). The spatial metaphor is a very powerful one in human information processing. As we will see in this paper, the spatial metaphor also lends itself well to modern distributed computing environments such as the World-Wide Web (WWW). Semantic road maps are not necessarily based on neural networks. In fact there are quite varied approaches to visual information retrieval interfaces (Pörner 1995). Nor is the SOM approach the only neural network approach which could be used: for example Zavrel (1996) favours an adaptive grid (growing cell structures network). However we find the Kohonen approach to be a highly effective method. This method is validated on real data in this paper. This method is computationally tractable and produces stable results of high quality. It is compatible with widely-available software - for instance, no Web browser add-ons are required contrary to the case of other visual interfaces.