Conversational agents are increasingly used in customer relations activities.Indeed, they can automatically carry out various tasks during a conversation with a human user in natural language. The number of communications that these agents are handling is constantly increasing, and the domains in which they are applied are diversifying. However, current methods are not yet able to adapt efficiently to new domains, as they are based on a mix of deep learning and business logic, and thus require a large amount of training data and specific modifications of business code in order to be able to handle new cases effectively. Besides, these approaches often don’t make use of the interactions that agents have once deployed, which could be a valuable source of training data for continuous improvement.

Thus, the goal of this thesis is to propose a distributed approach, coupling the multi-agent paradigm with machine learning, in order to build adaptive and learning conversational agents, with minimal need for supervision signals.