This paper deals with a prospective way to allow the communication between heterogeneous agents in the framework of multimedia services. The focus is set on the interaction between the user of a service and his/her virtual assistant. The interaction takes the form of a dialogue where the assistant tries to identify the tasks that the current user wants to perform. After a short description of dialogue systems, we propose a way to adapt them for discovering the user's goals. Our approach to control the dialogue relies on Markov Decision Processes, which are particularly suitable to handle uncertainty that occurs at many points of the communication. We moreover collect knowledge about the users in order to build profiles that will be reused for future sessions.