A fundamental revision of how AI acquires and processes language could result in more effective LLMs

Can a computer learn a language the way a child does? A study published in Computational Linguistics by professors Katrien Beuls (University of Namur) and Paul Van Eecke (AI Lab, Vrije Universiteit Brussel) sheds new light on this question.

“Children learn their native language by communicating with the people around them in their environment. As they play and experiment with language, they attempt to interpret the intentions of their conversation partners. In this way, they gradually learn to understand and use linguistic constructions. This process, in which language is acquired through interaction and meaningful context, is at the core of human language acquisition,” says Beuls.

“The current generation of large language models (LLMs), such as ChatGPT, learns language in a very different way,” adds Van Eecke.

“By observing vast amounts of text and identifying which words frequently appear together, they generate texts that are often indistinguishable from human writing. This results in models that are extremely powerful in many forms of text generation—such as summarizing, translating, or answering questions—but that also exhibit inherent limitations.

“They are susceptible to hallucinations and biases, often struggle with human reasoning, and require enormous amounts of data and energy to build and operate.”

The researchers propose an alternative model in which artificial agents learn language as humans do—by engaging in meaningful communicative interactions within their environment. Through a series of experiments, they demonstrate how these agents develop linguistic constructions that are directly linked to their surroundings and sensory perceptions. This leads to language models that:

• Are less prone to hallucinations and biases, as their language comprehension is grounded in direct interaction with the world.
• Use data and energy more efficiently, resulting in a smaller ecological footprint.
• Are more deeply rooted in meaning and intention, enabling them to understand language and context in a more human-like manner.

“Integrating communicative and situated interactions into AI models is a crucial step in developing the next generation of language models. This research offers a promising path toward language technologies that more closely resemble how humans understand and use language,” the researchers conclude.