Prediction, syntax and semantic grounding in the brain and large language models.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Linguistics

Background:

• Language comprehension relies on predicting future linguistic input.
• Integrating syntax and semantics rapidly is crucial for understanding.
• Spatiotemporal dynamics of anticipation in naturalistic speech are complex.

Purpose of the Study:

• Investigate word-class-specific neural responses during continuous speech perception.
• Examine anticipatory processes using combined EEG-MEG.
• Relate neural findings to word class predictability in a language model.

Main Methods:

• Simultaneous electroencephalography (EEG) and magnetoencephalography (MEG) recordings.
• Analysis of neural responses to word classes in German audio book.
• Source-space analysis of event-related fields and potentials.
• Computational analysis of word class predictability in the Llama language model.

Main Results:

• Reproducible spatiotemporal signatures for different word classes were identified.
• Significant pre-onset activity for nouns indicated enhanced anticipatory processing.
• Neural activity extended to sensorimotor cortices, suggesting grounded semantics for nouns.
• Language model analysis provided a computational reference for neural findings.

Conclusions:

• Combined EEG-MEG is powerful for studying predictive language processing.
• Nouns show enhanced anticipatory processing and deeper semantic grounding.
• Computational models complement neural data in understanding language comprehension mechanisms.