Tracking Lexical and Semantic Prediction Error Underlying the N400 Using Artificial Neural Network Models of Sentence...

Area of Science:

• Cognitive Neuroscience
• Computational Linguistics
• Psycholinguistics

Background:

• The N400 event-related potential is sensitive to semantic context during language processing.
• Previous models proposed N400 reflects lexical prediction error (surprisal) or semantic prediction error.
• Artificial neural networks show dynamics similar to N400, suggesting predictive processing.

Purpose of the Study:

• To model N400 amplitudes using a sentence gestalt model's meaning representation update.
• To quantitatively compare the sentence gestalt model to a next-word prediction (surprisal) model.
• To investigate the sub-processes underlying N400 generation during naturalistic sentence comprehension.

Main Methods:

• Utilized a sentence gestalt model trained on a large text corpus to predict N400 amplitudes.
• Employed naturalistic sentence processing to elicit N400 potentials.
• Compared model predictions based on 'gestalt update' with surprisal estimates from a language model.

Main Results:

• Both sentence gestalt update and surprisal significantly predicted N400 amplitudes.
• The findings indicate that N400 reflects aspects of both semantic and lexical prediction.
• The study provides a quantitative link between neural dynamics and computational language models.

Conclusions:

• N400 amplitudes likely arise from at least two distinct, interacting sub-processes.
• These sub-processes may involve semantic prediction error and lexical prediction (surprisal).
• This research offers a more nuanced understanding of the neural basis of sentence comprehension.