We explore a new dialogue modelling approach for assistive social robots that enables us to formalize flexible conversation flows between a robot and a human. We achieve this by introducing an expectation mechanism to handle, for example, topic change, clarification questions or misunderstandings during a dialogue. The model gave us insight into the dialogue structure and how it is shaped by several linguistic and pragmatic features. This is work in progress and in the future we will explore learning algorithms that mine the features, implement and validate the model with real conversations.