Conclusions

In this paper we have provided an account on the feasibility of using unsupervised neural networks in a highly important task of information retrieval, namely text classification. As an experimental document collection we used the description of various countries as contained in the 1990 edition of the CIA World Factbook. For this document collection it is rather easy to judge the quality of the classification result. For document representation we relied on the vector space model and a simple $tf \times idf$ term weighting scheme.

We demonstrated that both the map and the atlas metaphor are highly useful for assisting the user to find her orientation within the document space. The shortcoming of the map metaphor, however, is that each document is shown in one large map and thus, the borderline between clusters of related and clusters of unrelated documents are sometimes hard to find. This is especially the case if the user does not have sufficient insight into the contents of the document collection.

The atlas metaphor overcomes this limitation in that the clusters of documents are clearly visible because of the architecture of the neural network. The document space is separated into independent maps along different layers in a hierarchy. The user thus gets the best of both worlds. The similarity between documents is shown in a fine-grained level in maps of the lower layers of the hierarchy while the overall organizational principles of the document archive are shown at higher layer maps. Since such a hierarchical arrangement of documents is the common way of organizing conventional libraries, only small intellectual overhead is required from the user to find her way through the document space.