Information filtering is one of the most useful and challenging tasks for effective information access. It is concerned with dynamically adapting the distribution of information where both evolving user's interests and new incoming information are taken into account. In this paper, we present an innovative approach to text filtering based on the novelty detection principle. This approach relies on a specific learning model which allows both accurate online learning of user's profile and evaluation of the coherency of user's behaviour during his interaction with the system. We empirically analyse our approach and present experimental results on the Reuters-21578 benchmark. The obtained results bring out a significant enhancement of performance as compared to the widely used Rocchio's learning algorithm.