Reinforcement approaches for online learning of virtual vocal conversational agents

Seminar by Fabrice Lefèvre, Université d’Avignon

Room C207 9:30 – 11:00


Abstract: Stochastic techniques, supervised or not, or by reinforcement, are now largely spread on every modules intervening in academic systems of human-machine dialog. Current researches strive to improve even further their quality to reach satisfying and acceptable global performance. A major default remains the cost of porting such interaction systems to new domains, tasks or languages, which implies collection and annotation of huge corpora. During the talk, after a brief overview of the state-of-the-art techniques for academic systems, I will expose the recent work carried out at LIA in the Vocal Interactions Group targeting a dramatic reduction of the models initialisation effort and allowing the tuning of the inner components of the interfaces in usage conditions with real users.