This paper investigates a new approach for unsupervised and semisupervised learning. We show that this method is an instance of the Classification EM algorithm in the case of gaussian densities. Its originality is that it relies on a discriminant approach whereas classical methods for unsupervised and semi-supervised learning rely on density estimation. This idea is used to improve a generic document summarization system, it is evaluated on the Reuters news-wire corpus and compared to other strategies.