In a lexical analyzer, the scanning of a whole dictionary using an editing distance is of very high computational cost. We present a lexical analyzer designed to focus on a very limited sub-set of the whole dictionary. The system is based on a Self-Organizing Feature Map which maps the dictionary on a two-dimensional space. The neighborhood relationships on this space are then used to define a short list of hypotheses. We introduce a multi-staged pyramidal network to speed-up the access and we present the performance of the system. these results are then interpreted.