Robust assessment of the cortical encoding of word-level expectations using the temporal response function

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Linguistics

Background:

• Speech comprehension relies on predicting upcoming words based on semantic context.
• Existing neural evaluation metrics for speech processing overlook word-level features, limiting insights into predictive mechanisms.
• Traditional methods struggle with the discrete nature of lexical information in neural signals.

Purpose of the Study:

• To address limitations in evaluating lexical prediction in neural data.
• To introduce novel metrics for assessing the neural encoding of lexical surprise.
• To improve the understanding of predictive coding during naturalistic speech comprehension.

Main Methods:

• Analysis of simulated and actual electroencephalography (EEG) data during speech comprehension.
• Development and application of two new assessment metrics for neural lexical surprise.
• Comparison of novel metrics against the conventional temporal response function (TRF) approach.

Main Results:

• The novel metrics yielded effect sizes over 140% larger than the TRF method.
• These enhanced effect sizes were consistently observed in both simulated and real EEG datasets.
• The new metrics provide more precise measurements of neural lexical prediction.

Conclusions:

• The developed metrics significantly advance the evaluation of lexical prediction in neural signals.
• These findings offer deeper insights into the brain's predictive representations during language processing.
• The study opens new research pathways for predictive coding in naturalistic language.