The predictive brain: Neural correlates of word expectancy align with large language model prediction probabilities

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Psycholinguistics

Background:

• Predictive coding theory posits the brain anticipates words for efficient language processing.
• Neural mechanisms underlying this predictive processing in naturalistic speech remain largely unknown.

Purpose of the Study:

• To investigate the neural correlates of word predictability in naturalistic speech.
• To bridge computational language models with neurophysiological data for language comprehension research.

Main Methods:

• Simultaneous electroencephalography (EEG) and magnetoencephalography (MEG) recordings from 29 participants.
• Participants listened to an audiobook and rated noun predictability using BERT and LLaMA language models.
• Analysis focused on neural responses (N400 amplitude) and pre-word onset activity.

Main Results:

• Higher word predictability correlated with reduced N400 amplitudes during recognition.
• Increased pre-stimulus activity was observed before word onset, especially for low-predictability words.
• EEG indicated left fronto-temporal pre-activation; MEG suggested sensorimotor engagement for low-predictability words.

Conclusions:

• The brain dynamically integrates top-down predictions with bottom-up sensory input for language comprehension.
• Findings provide novel insights into predictive processing in naturalistic speech, linking AI models with human brain function.
• This research advances understanding in cognitive computational neuroscience and inspires neuroscience-inspired AI development.