Strong Prediction: Language Model Surprisal Explains Multiple N400 Effects

Area of Science:

• Cognitive Neuroscience
• Computational Linguistics
• Psycholinguistics

Background:

• The N400 brain response is a key indicator of language processing.
• Existing theories debate whether N400 amplitude reflects prediction, semantic similarity, or both.

Purpose of the Study:

• To investigate which theoretical account of the N400 is best supported by empirical evidence.
• To differentiate the roles of predictability and semantic similarity in N400 generation.

Main Methods:

• Utilized GPT-3, a large language model, to quantify word predictability (surprisal).
• Employed GloVe and fastText vector representations to measure contextual semantic similarity.
• Constructed regression models predicting single-trial N400 amplitudes from these linguistic variables.

Main Results:

• GPT-3 surprisal emerged as the strongest predictor of N400 amplitude.
• Statistical model comparison favored predictability over semantic similarity alone.
• N400 effects attributed to expectancy, plausibility, and semantic similarity were unified under predictability.

Conclusions:

• Word predictability is the primary driver of N400 amplitude.
• Findings support predictive coding theories within the human language network.
• This research offers a unified account of factors influencing the N400.