Grounded concepts in multi-word utterances: A methodology for the emergence of compositional languages

Défense de mémoire de Xantin Préser

The recent and impressive advances in artificial intelligence, particularly with Large Language Models (LLMs), demonstrate a remarkable ability to replicate certain human behaviours, especially in generating and interpreting text or images. However, LLMs also suffer from several limitations, such as a tendency to hallucinate, a lack of human-like logic and pragmatic reasoning, and a massive need for data. They also do not, or at least not easily, shed light on the
mechanisms behind the emergence of human language. To better understand these mechanisms, other approaches and frameworks exist, such as the language games, although their capabilities
remain limited. Greater attention to these alternative approaches could, in addition to helping the linguistic community to understand how languages emerge, offer more effective artificial
intelligence models for certain tasks using mechanisms closer to those employed by humans.

In this master’s thesis, we build on a method that enables the decentralised emergence of a linguistic convention grounded in the real world that is communicatively effective, robust and
adaptable to change, and we add the ability for agents using this convention to produce sentences composed of several words. This method uses each word in a sentence as a filter that
is applied to the world around the agents, allowing them to identify one entity being discussed among others. Adding this capability does not diminish the effectiveness of the original method
and preserves its ability to give rise to a robust and adaptive linguistic convention. Above all, it reduces the amount of vocabulary needed to accomplish this same identification task, making
it more efficient.

Keywords: language emergence, compositional language, emergent communication, multiagent systems, autonomous agents, compositionality

Contact : Isabelle Daelman - isabelle.daelman@unamur.be
Télecharger : vCal