Several text classification methods have been adopted and extensively tested in the past, yielding promising results. Related research has concentrated on the calibration of the main attributes of each method. However, no essential effort has been made to improve the discrimination capacity of the learning space, prior to classification, using text-data mining techniques. We propose a two level text classification model. The first level consists in clustering the classes using a self-organizing map. Subsequently, a text classifier completes the task, by surveying a reduced space, but with enhanced discrimination capacity.