The question of machine translation has been raised since the birth of IT. It seemed accessible at the time, but anyone who now uses the automatic translators available on the Internet knows that, despite the remarkable progress made, we are still far from a translation that is still accurate. The complexity of natural language and its ambiguities are well designed to confuse the fully rational algorithms of our computers. The "neural networks" that practice "deep learning" are the latest of the many strategies deployed to talk with the machine... and make themselves understood.

Thierry Poibeau is a research director at the CNRS. He is a specialist in automatic language processing, a field of research on the borderline between linguistics and computer science.