Constituent-constrained word prediction during language comprehension

Area of Science:

• Neuroscience
• Computational Linguistics
• Psycholinguistics

Background:

• Next-word prediction is a proposed core function of the human language system, similar to large language models.
• Research investigates if the brain prioritizes precise word prediction during speech comprehension.

Purpose of the Study:

• To test the hypothesis that the brain predicts upcoming words with maximum precision during connected speech.
• To examine how constituent boundaries influence neural and behavioral responses to word unpredictability.

Main Methods:

• Magnetoencephalography (MEG) experiments with Mandarin Chinese speakers.
• Behavioral experiments assessing speech perception.
• Analysis of electrocorticography (ECoG) data from English narratives.

Main Results:

• Neural responses to word surprisal (unpredictability) were stronger within constituents than across major boundaries.
• This constituent-boundary effect was also observed behaviorally under normal speech rates.
• The effect was confirmed using ECoG data from natural English speech.

Conclusions:

• The language system does not exclusively optimize for word-prediction precision.
• It balances prediction with structural constraints, managing linguistic context based on constituent boundaries.