Raisonnement sur les incertitudes et apprentissage pour les systèmes de dialogue conventionnels

  • Romain Laroche

PhD Thesis: PhD Thesis at Universite Paris VI, France |

More and more industries need dialogue applications. In customer relationship management, they range from home shopping to after-sales service, order tracking, directory enquiries. . .The hotlines of these vocal services are very costly and often overcrowded. In the end, the proposed service is expensive and has a low quality level. Confronted with this problem, industry and research struggle to converge. On the one hand, industrial dialogue systems are designed with decision-making automata describing the dialogue logics. These automata are reputed simplistic, hard to design and suboptimal. On the other hand, scientists focus on advanced techniques that only experts are able to implement and that remain sorely monitorable. Grounded in the system global architecture, this PhD thesis endeavoured to reconcile research with industry by enclosing the scientific advances into the industrial process. This work led to the definition of a new model for reasoning on uncertainties, to the definition of a new non-Markovian decision process and to the implementations and optimisations of plug-and-play algorithms dedicated to the problem. The advanced functionalities developed in this thesis enable to improve robustness, to guarantee the optimality of design choices and to have the project managers receive an easy-to-comprehend usage feedback. These results have motivated the implementation of the world première of commercial dialogue application incorporating online reinforcement learning.