This paper presents an original approach to modelling user's information need in text filtering environment. This approach relies on a specific novelty detection model which allows both accurate learning of user's profile and evaluation of the coherency of user's behaviour during his interaction with the system. Thanks to an online learning algorithm, the novelty detection model is also able to track changes in user's interests over time. The proposed approach has been successfully tested on the Reuters-21578 benchmark. The experimental results prove that this approach significantly outperforms the well-known Rocchio's learning algorithm.