Integrating contextual information into text document clustering with Self-Organizing Maps

Exploration of large text collections requires suitable methods of presentation. The Self-Organizing Map has shown promising results over the past years. The topographic map approaches usually use the common vector space model for text document representation. We present here a new two stage representation which uses sentences as intermediate information units. In this way contextual information is preserved and influences the process of self-organization. We demonstrate that presence of the contextual information improves the quality of the resulting document maps. The procedure is computationally more expensive but we present modifications of the algorithm which cope with this problem.