Reverse engineering early language learning by Emmanuel Dupoux by (Ecole des Hautes Etudes en Sciences Sociales, Laboratoire de Sciences Cognitives et Psycholinguistique)
Decades of research on early language acquisition have documented how infants quickly and robustly acquire their native tongue(s) across large variations in their input and environment. The mechanisms that enable such a feat remain, however, poorly understood. The proposition, here, is to supplement experimental investigations by a quantitative approach based on tools from machine learning and language technologies, applied to corpora of infant directed input. I illustrate the power of this approach through a reanalysis of some previous claims made regarding the nature and function of Infant Directed as opposed to Adult Directed Speech (IDS vs ADS). I also revisit current ideas about the learning of phoneme categories, a problem that has been long thought to involve only bottom-up statistical learning. In contrast, I show that a bottom up strategy does not scale up to real speech input, and that phoneme learning requires not only the joint learning of phoneme and word forms but also of prosodic and semantic representations. I discuss a global learning architecture where provisional linguistic representations are gradually learned in parallel, and present some predictions for language learning in infants.
