In internet chatrooms, multiple conversations may occur simultaneously. The task of identifying to which conversation each message belongs is called disentanglement. In this paper, we first try to adapt the publicly available system of Elsner and Char-niak (2010) to a French corpus extracted from the Ubuntu platform. Then, we experiment with the discursive annotation of utterances. We find that disentanglement performances can vary significantly depending on corpus characteristics. We also find that using discursive information, in the form of functional and rhetoric relations between messages, is valuable for this task.