Automatic constraints generation for semisupervised clustering: experiences with documents classification