Conference: Knowledge co-construction environments, large language models (chatGPT, Flaubert) and small discrete models (JeuxDeMots, Hélix, Idéfix)

As part of the Master of Computer Science graduation ceremony, the FdS Computer Science Department is hosting a conference for the general public.

Friday, June 2 - 2:30 pm
Amphi Dumontet

Conference moderated by Mathieu Lafourcade, Lecturer at Lirmm

It is interesting to develop environments for the co-construction of knowledge between human experts and one or more artificial intelligences (AIs). The human experts discover knowledge with the help of the AIs, and the AIs learn (essentially from common sense). A conversational interface, such as a chatbot, seems relevant because it is intuitive and versatile. We look at the use of chatGPT (based on GPT 3.5) and another language model, Flaubert, and attempt to determine its performance in this context using the lexical-semantic WordGames.

We also present a discrete neural model that is an overlay of the JeuxDeMots lexical network, which can be seen as an illustration of a different approach. This model implements forward-forward learning and seems a biologically credible and effective alternative to gradient backpropagation. The architecture of this model is not fixed a priori, and evolves as learning progresses thanks to the dynamic addition of neurons.

To illustrate our point, we present the OACS project, which aims to automatically determine whether certain clauses in a consumer contract are unfair or not.