6533b7d4fe1ef96bd1261a89

RESEARCH PRODUCT

Langage et Apprentissage en Interaction pour des Assistants Numériques Autonomes - Une Approche Développementale

subject

description

The rapid development of digital assistants (DA) opens the way to new modes of interaction. Some DA allows users to personalise the way they respond to queries, in particular by teaching them new procedures. This work proposes to use machine learning methods to enrich the linguistic and procedural generalisation capabilities of these systems. The challenge is to reconcile rapid learning skills, necessary for a smooth user experience, with a sufficiently large generalisation capacity. Though this is a natural human ability, it remains out-of-reach for artificial systems and this leads us to approach these issues from the perspective of developmental Artificial Intelligence. This work is thus inspired by the cognitive processes at work in children during language learning.First, we propose a language processing module, which relies on semantic comparison methods to interpret the user’s natural language requests. The variability of a user speech is indeed one of the main difficulties of these learning assistants. We provide them with a generalisation tool to continuously adapt to the user language. Another challenge for these learning agents is their ability to transfer their knowledge to new objects and contexts. We propose a series of architectures for Deep Reinforcement Learning agents that learn to perform tasks expressed in natural language in various environments. By exploiting language as an abstraction tool to represent tasks, we show that in structured environment, these agents are able to transfer their skills to new objects.Finally, we develop a use case in a home automation environment. We propose a learning assistant that integrates the systems mentioned above.