Improving the Quality of Labels for Self-Organising Maps Using Fine-Tuning

Vector representation of legal documents is still the best way for computing classification clusters and labelling of its contents. A very special problem occurs with self organising maps: strong clusters tend to dominate neighbouring smaller clusters in terms of their weight vector structure, which influences the labels extracted from these. This unwelcome side-effect can be overcome efficiently with a dedicated fine-tuning phase at the end of the training process, in which the neighbourhood radius of the training function is set to zero. Experiments with our text collection have shown the high improvement of the quality of labelling.