We use supervised learning. For the system to work, you have to first give it a labeled training corpus. It should consist of historical user queries with the correct answers or intents assigned to them. Slots shall be identified and labeled.