The goal of this PhD is to create a realistic conversational agent with a personality. Conversational agents, otherwise known as chatbots, were the subject of an impressive hype phenomenon a few years ago. Startups flourished and growth forecasts were gigantic. Since then, the fervor has subsided. Chatbots have become just another tool in the customer relationship on the Internet. Technologies have become standardized around intention detectors and database access.
It is now very easy to create a conversational agent for a given task. It is even feasible to make it accept various inputs not foreseen in advance and to make it generate a fluid, syntactically and semantically correct text. On the other hand, it is very difficult to produce a generic agent able to adapt very quickly to a new task.
The goal of this thesis is to solve this problem. The response generation model should take as input a description of the agent's knowledge about the world, about itself and about the current conversation in order to generate responses consistent with this model. This external knowledge must be easily replaced or updated. The conversational agent produced will be developed and experimented within the framework of the European project Cortex² aiming, for one of its use cases, to produce tools to facilitate the experience of online or mixed meetings.