Providing Topically Sorted Access to Subsequently Released Newspaper Editions or: How to Build Your Private Digital Library

Abstract

Self-organizing maps are a popular neural network model for presenting high-dimensional input data on a two-dimensional map, providing a particularly useful interface to electronic document collections. However, as the size of the training data increases, both the necessary computational power as well as the training time required exceed tolerable limits. Still more important, not all training data may be available in one central location but may rather be collected and managed at different repositories or released in subsequent periods of time.

This paper describes an approach for combining independent, distributed self-organizing maps to build a higher order map, allowing the creation and maintenance of scalable, independent map systems, which can be built to suit the needs of individual users. This is achieved by training higher order maps using the trained lower order maps as input data. We demonstrate this approach by creating an integrated view of subsequent releases of a newspaper archive.