No black scorpion sleeps furiously on that table:

Why Chomsky, Skinner and ChatGPT got it wrong on language

Arnaud REY (Eq. MemoPsy)

When: March 31st - 11h/12h

Where: Salle des Voûtes 

Abstract: There is probably no more fascinating question than that of language. In the middle of the last century, two radically opposed approaches collided on this fundamental issue. Based on the well-established science of behavior, Skinner applied this approach to the question of language and attempted to demonstrate that it may also be a question of learning through reinforcement. At the same time, the young Chomsky put forward decisive arguments against the Skinnerian approach and showed that this approach was insufficient to account for the generativity of language, which requires an innate underlying grammar to produce the sentences of a given language. Recently, with the development of large-language models (LLMs) and more generally, with all the work resulting from the field of implicit statistical learning (ISL), it is the Chomskian approach that seems to be in a bad way. I will try to show that even if the LLMs approach seems more than promising, it also has some weaknesses in accounting for language in the human species because it does not yet take the fundamental properties of our memorization processes sufficiently into account.