CIA's view of the world and what neural networks learn from it:
A comparison of geographical document space representation metaphors

Dieter Merkl and Andreas Rauber

Institut für Softwaretechnik, Technische Universität Wien
Resselgasse 3/188, A-1040 Wien, Austria
{dieter, andi}@ifs.tuwien.ac.at

Abstract:

Text collections may be regarded as an almost perfect application arena for unsupervised neural networks. This because many operations computers have to perform on text documents are classification tasks based on noisy patterns. In particular we rely on self-organizing maps which produce a map of the document space after their training process. From geography, however, it is known that maps are not always the best way to represent information spaces. For most applications it is better to provide a hierarchical view of the underlying data collection in form of an atlas where starting from a map representing the complete data collection different regions are shown at finer levels of granularity. Using an atlas, the user can easily ``zoom'' into regions of particular interest while still having general maps for overall orientation. We show that a similar display can be obtained by using hierarchical feature maps to represent the contents of a document archive. These neural networks have a layered architecture where each layer consists of a number of individual self-organizing maps. By this, the contents of the text archive may be represented at arbitrary detail while still having the general maps available for global orientation.