Mis à jour
13 juin 2024

Isidora Stojanovic, François Recanati et Emmanuel Dupoux ont obtenu un financement ERC

Isidora Stojanovic, chercheuse CNRS à l'Institut Jean Nicod (IJN), François Recanati, directeur de recherche CNRS à l'IJN, drecteur des études à l'Hautes Etudes en Sciences Sociales (EHESS) et professeur au Collège de France, et Emmanuel Dupoux, directeur des études à l' EHESS et directeur de l'équipe Cognitive Machine Learning au Laboratoire de Sciences Cognitives et Psycholinguistique, ont obtenu un financement ERC Advanced Grant. 

The funding is amongst the EU’s most prestigious and competitive, providing leading senior researchers with the opportunity to pursue ambitious, curiosity-driven projects that could lead to major scientific breakthroughs. 

François Recanati, Isidora Stojanovic, Emmanuel Dupoux.

Here are the three winning projects :

"Mental Files: New Foundations" project, carried by  François Recanati 

This project is about the representation of particular objects in language and thought, a topic which has been at the forefront of philosophical attention for more than a century.

About fifty years ago, ‘descriptivism’ was demoted from its dominant position in philosophy in favour of the theory of ‘direct reference’. A similar shift away from descriptivism has been a noticeable feature of work on the representation of objects in cognitive science, where the notion of an ‘object file’ has made it possible to unify research on perception and on infant cognition. The object file construct is in many respects similar to the philosophical idea of direct (non-descriptive) grounding for thoughts about particulars, and this has given rise to a new research program: the generalization of the file idea from perception to thought. Thus the last decade has seen the development of the mental file framework, according to which nondescriptive thoughts about particulars (so-called ‘singular thoughts’), whether or not they are based on perception, involve mental files whose ‘reference’ does not depend on category information to be found in the file but on certain relations to the object the obtaining of which triggers the opening of the file. The mental file framework has attracted considerable attention not only in philosophy, but also in psychology(Perner) and linguistics (Kamp). It has also inspired work in aesthetics and the philosophy of fiction.

Successful though it is, the mental file framework currently faces what may be described as foundational crisis. According to a recurrent piece of criticism, it fails to provide appropriat identity and persistence conditions for mental files. This threatens the credibility of the framework, reduced to a convenient metaphor, and puts it at risk despite its high promises and considerable appeal. The aim of this philosophical project is to end the crisis by entirely rethinking the foundations of the framework.


"Valence asymmetries: the positive, the negative, the good and the bad in language, mind and morality"project, carried by Isidora Stojanovic

An asymmetric behavior between the positive and the negative has been evidenced in psychology, for information processing, attention, recognition and decision making, in philosophy, for judgments about morality and intentionality, and in linguistics, for a range of lexical, syntactic, semantic and pragmatic phenomena. Negative information grabs our attention, we process it more carefully, we recall it with greater precision. We easily blame others for the negative side-effects of their actions, but do not praise them for the positive ones. It takes many nice words to overthrow one nasty remark. When we say that something is "not good", we usually imply that it is bad, but by saying "not bad" we do not imply that it is good. Valence asymmetries have arisen on many horizons but have seldom been brought into correspondence, and are at odds with most theories of value. The present project is a pioneering attempt to secure the premises for a cross-fertilization between the different accounts of valence asymmetries. It will deploy methods from philosophy (argumentation and conceptual analysis), formal semantic and value-theoretic models, and experimental methodology from psycholinguistics and moral psychology. It has three main objectives: - highlight the fundamental role that valence plays beyond emotion, in particular, in value judgments and language; - examine what the different asymmetries have in common, and whether they call for a unified explanation; - show that valence asymmetries are not necessarily irrational, but often derive from a fundamental asymmetry between positive and negative value, and, as such, are a key component of our cognitive and linguistic architecture. Furthermore, we will (a) articulate the relationship among the notions of valence, value and polarity; (b) put forward a novel account of the asymmetry of negation; (c) unearth new asymmetries in the realm of morality, virtue and vice; and (d) provide an account of valence reversals.

"Why do infants learn language so fast? A reverse engineering approach" project, carried by Emmanuel Dupoux

According to a dominant hypothesis in developmental psychology, children learn their native language(s) by following an observational and predictive method: they absorb the linguistic information to which they are exposed and derive a statistical model, enabling them to guess what is going to be said from what has just been said. This is precisely the method used by AI language models such as chatGPT, Llama and Gemini, with spectacular results.  However, on closer inspection, children and AI models differ in one fundamental respect: the amount of linguistic data at their disposal. To reach a level comparable to that of a one-year-old child, AI models need to be exposed to corpora some 100 times larger, according to estimates derived from field linguistics. This gap increases for 3-year-olds, and widens further if we consider AIs trained on speech rather than text. All in all, 3-year-olds could require 10 to 100 thousand times less linguistic data than equivalent AI models. So what are the assets that enable children to learn so effectively?

Emmanuel Dupoux and his team are tackling this question with their InfantSimulator project. They are building an "infant simulator" based on the latest statistical learning techniques, which they are applying to realistic audio recordings, and comparing learning curves with those observed in children. They explore a number of hypotheses to explain the tremendous acceleration of learning in children: unlike AI, children have a body, and in particular a vocal tract enabling them to understand the organization of language sounds; children have an episodic memory enabling them to learn a word in a single exposure; children do not start out from a randomly connected brain, but from an initial state optimized in the course of evolution; children have a social environment with which they can interact, and so on. Emmanuel Dupoux's research aims to incorporate these hypotheses into his simulator, in order to quantitatively evaluate their capacity to accelerate learning. It uses AI to revisit traditional controversies about the biological, cognitive and cultural determinants of language and language learning.