Children inspire smarter AI: VUB research shows a new path in language technology

“Children learn their mother tongue by communicating with the people around them in their daily environment. While playing and experimenting with language, they try to interpret the intentions of their conversation partners. Step by step, they learn to understand and use language constructions. This process, in which language is acquired through interaction and meaningful context, is at the heart of human language acquisition,” says Katrien Beuls.

“The current generation of large language models (LLMs), such as ChatGPT, learn language in a very different way,” adds Paul Van Eecke. “By observing in vast amounts of text which words frequently occur together, they learn to generate texts that are often indistinguishable from human writing. This results in models that are extremely powerful in many forms of text generation – from summarising or translating texts to answering questions – but which also show a number of inherent limitations. They are prone to hallucinations and biases, often struggle with human forms of 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 in the same way humans do: by taking part in meaningful communicative interactions within their environment. In a series of experiments, they demonstrate how these agents develop language constructions directly linked to their environment and sensory perceptions. This leads to language models that:

• Are less prone to hallucinations and biases, because their understanding of language is grounded in direct interaction with the world.
• Use data and energy more efficiently, leaving a smaller ecological footprint.
• Are more firmly rooted in meaning and intention, enabling them to understand language and context in a more human-like way.

“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 towards language technologies that come closer to how humans understand and use language,” the researchers conclude.