Our framework is what we call `intentional approach' of dialogue. An agent is represented by its intentional state (or mental state): belief, mutual belief, intention and action are represented by modal operators in a dynamic doxastic logic. We study the dynamics of beliefs in cooperative task-oriented human-machine dialogues. We suppose that during the dialogue each participant can mistake a piece of information, forget it or simply change his mind. We introduce the topic concept in the aim to describe the dynamics of beliefs.