On the Mathematical Relationship Between Contextual Probability and N400 Amplitude

Area of Science:

• Cognitive Science
• Neuroscience
• Computational Linguistics

Background:

• The relationship between word probability and processing difficulty in language comprehension is debated.
• Existing models propose linear, logarithmic, or super-logarithmic relationships, but empirical evidence is mixed.
• Current theories do not fully explain how statistical language properties impact comprehension.

Purpose of the Study:

• To investigate the mathematical relationship between corpus-derived word probabilities and the N400 brain response.
• To determine which transformation of word probability best predicts N400 amplitude.
• To test predictions from contemporary transformer language models against N400 data.

Main Methods:

• Utilized 37 transformer language models to compute contextual word probabilities.
• Analyzed data from 6 experimental studies measuring the N400 response.
• Tested linear, logarithmic, super-logarithmic, and sub-logarithmic probability transformations, including combinations.

Main Results:

• Replicated findings that combined linear and logarithmic probability transformations predict N400 amplitude better in some datasets.
• Identified sub-logarithmically transformed probability as the superior single predictor across most models and datasets.
• Demonstrated that sub-logarithmic transformation explains variance previously attributed to linear and logarithmic transformations.

Conclusions:

• Sub-logarithmic probability is a novel and highly effective predictor of N400 amplitude in language comprehension.
• This finding challenges existing theoretical accounts of language processing.
• Highlights the importance of statistical language regularities in understanding neural responses during comprehension.