Messages posted on social media are a very complete and diverse source of live information. However, this kind of data is hard to exploit because of its ever-increasing volume, its heterogeneous representations and the presence of noise. We are developing tools to help an analyst extract relevant information from a stream of tweets. As a first step, the approach we propose automatically clusters similar tweets. This clustering makes it possible to categorize tweets while reducing the amount of data to be processed. Next, clusters of tweets most likely associated with events of interest are identified. One of our biggest challenges is to achieve this processing in real time, by frequently updating clusters after the publication of new tweets. This work is part of a research project in partnership with the company Thales Canada.