Neuroscientists want to know how the human brain makes sense of spoken language. This involves more than just hearing sounds. The brain has to figure out what words mean and how those meanings change depending on the situation.
Researchers are now turning to large language models (LLMs), which power tools like ChatGPT and Gemini, to learn how our minds handle these complex language patterns.
A team from Baylor College of Medicine, Rice University, UC Berkeley, and Texas Children’s Hospital recently published a study in Nature Neuroscience. They looked at how certain groups of brain cells store word meanings, taking ideas from how advanced AI systems process text.
To get accurate data, scientists recorded the brain activity of 10 people as they listened to stories. They used tiny wire electrodes, which were temporarily placed in the patients’ brains during routine epilepsy monitoring.
After gathering the data, the team used mathematical models to study the language. They compared the brain activity they recorded with the word data created by AI models that processed the same stories.
The results showed that word meanings are stored in complex patterns across many neurons in the hippocampus, a brain area known for memory. The researchers found that these patterns hold detailed information about how word meanings change with context.
Interestingly, the brain activity they recorded matched most closely with data from AI models like GPT-2. The scientists also studied polysemy, which is when one word has several related meanings.
They found that, like AI systems that use math to measure differences between data points to understand context, the human brain uses wide networks of neurons to understand spoken language.
This research offers a new way to understand how neurons in the hippocampus store word meanings during everyday conversations. In the future, this approach could help scientists use advanced AI to explore how the human brain works.