Breaking the Wall of Human-like Language in Machines
Breaking the Wall of Human-like Language in Machines
Global Call 2026 Finalist Interview: Social Sciences & Humanities
Katrien Beuls is professor of computational linguistics and artificial intelligence at the UNamur Faculty of Computer Science and the Namur Digital Institute. Her work centres around computational problems involving human language, with a focus on construction grammar, language evolution and computational models of the automatic interpretation and composition of semantic structures. She develops experiments in which populations of autonomous agents take part in situated communicative interactions to co-create artificial natural languages to solve tasks in their environment. Her work combines linguistic theory with experimental and evolutionary approaches to language understanding. She has participated in several European research projects (including ALEAR, ESSENCE, ODYCCEUS and MUHAI) and has published extensively in top venues in the field. Beuls is one of the chief developers of Fluid Construction Grammar (FCG), the leading computational framework for representing, processing and learning construction grammars.
Which wall does your research or project break?
The situated, communicative and interactional aspects that characterize human language use are largely absent from today's predominant approaches to language modelling in machines. Current-generation large language models (LLMs) manage to learn highly impressive linguistic capacities even if they only ever have been exposed to purely textual input. This text-internal prediction task lies at the root of a number of important shortcomings, in particular regarding the data hungriness of the models, their limited ability to perform human-like logical and pragmatic reasoning, and their susceptibility to biases. In my research, I am working on an alternative route that models how populations of artificial agents can create and learn linguistic structures by participating in situated communicative interactions, inspired by the way in which human languages emerged and evolved. Through a selection of computational experiments, I show how the linguistic knowledge captured in the resulting models is of a fundamentally different nature than the knowledge captured by LLMs.
Carrying out a project as ambitious as this one requires bridging theory and computational implementation. Rather than modelling languages as static systems in which patterns are to be found, my work seeks to establish how patterns emerge from communication in a dynamic systems approach. Taking inspiration from the origin and evolution of human natural languages, we research and develop novel theories and algorithms for the co-creation of meaning and linguistic expressions in populations of artificial language users who engage in repeated series of “language games”. These Wittgensteinian language games, played between two agents from the population in a local fashion, have the potential to lead to a global linguistic behavior as a result of powerful processes of natural selection and self-organization.
What is the main goal of your research or project?
The evolutionary and self-organising nature of human languages gives rise to a number of unique qualities. First of all, such decentralised, self-organising systems are known to be robust and able to self-repair substantial perturbations. Second, populations of language users converge on shared conventions that remain adaptive to changes in their environment and communicative needs. Finally, the resulting languages serve as an abstraction layer above the sensory observations and internal mental representations of individual language users. While linguistic forms can be observed and shared, their meanings remain tied to each language user’s individual physical and cognitive embodiment.
Rather than modelling the learning of an existing natural language, the methodologies that I develop allow for artificial natural languages to emerge and evolve to optimally support the embodiment, environment and communicative needs of populations of artificial agents. These languages are artificial in the sense that they do not exist outside the experimental set-up, yet natural in the sense that they emerge and evolve through the same evolutionary principles as human languages do. Apart from contributing computational evidence for the cognitive plausibility of usage-based theories of language learning and corroborating the theoretical underpinnings of construction grammar theories, the techniques that result from this research pave the way for learning computationally tractable, large-scale, usage-based construction grammars that facilitate both language comprehension and production. Apart from their theoretical importance, such grammars are also highly valuable for a large range of application domains, including intelligent conversational agents (as in my work on visual dialogue systems), grounded language understanding (cf. recent work in recipe understanding and kitchen simulators), intelligent tutoring (as in my PhD thesis) and the semantic analysis of discourse (as in the opinion facilitator tool we developed - https://penelope.vub.be/opinion-facilitator/).
What impact does your research or project have on society?
My research contributes in the first place to a better understanding of how human language works, and which cultural and societal circumstances needed to be in place for languages to emerge and evolve. The models that result from my research are mechanistic in the sense that they provide a computational operationalization of a number of central processes in the emergence and evolution of languages, such as the creation of meaning from a speaker’s intention in a given situation, the shaping of meaning over time as a result of usage, and the generalization and specialization of linguistic means to express a certain meaning. Apart from informing the public through regular press releases aligned with important scientific publications, my team participates yearly in local and national science exhibitions with interactive booths. In 2026, we contributed an activity to the ‘Printemps des Sciences’ in Namur, in which primary school children were invited to watch and play a game of ‘Guess Who?’ in a language invented by a population of artificial agents who came up with their own efficient way to describe the characters in the game. Since the techniques that we develop lead to interpretable language models, participants can inspect the vocabularies of their artificial opponents to infer the meaning of words they observe. Apart from the scientific impact of my work and its dissemination to the wider public, the methods and techniques we develop are highly valuable for a large range of application domains in which robust, interpretable and adaptive language technologies are required. Examples include the above-mentioned systems for grounded language understanding applied to the kitchen domain, intelligent language tutoring systems and the frame-semantic analysis of online discourse.
What advice would you give to young scientists or students interested in pursuing a career in research, or to your younger self starting in science?
The most important advice I would give a young scientist today is to resist taking shortcuts in developing their research career. With LLMs being omnipresent, and a flood of new papers drowning us every day, it is more important than ever to focus on deep work (read Cal Newport!) and invest (a lot of) time and effort in becoming a true expert in your domain. Try to resist the fast publication of perhaps still immature work but go for a long, qualitative article in a highly visible journal in your field, as your work will have a much bigger impact on the longer term. Becoming an expert will require reading a lot, in multiple directions, and learning from experienced researchers, so you can build up your own expertise as a result of years of interactions and dialogues with other researchers and their work. Moreover, science is usually not a solitary endeavor anymore but carried out by (interdisciplinary) teams of researchers, which allows you to focus on your strengths and find compatible collaborators for aspects of the project that you are less exercised in. Finally, stay CURIOUS in all that you do, as curiosity is the motor for true scientific advancement and you will often only find out the specific topic of your own PhD research by digging deeper until you hit an interesting barrier that no one has ever overcome before.
