Encoding EEG Signals to Examine Human-Like Next-Word Prediction Behaviour in Language Models

Short summary

This study examines whether language models that predict next words accurately also capture human-like cognitive signals during reading, using EEG-derived event-related potentials. Results show that only surprisal correlates with language-processing ERPs, particularly for open-class words with high semantic content. The findings challenge the assumption that scaling LMs with more parameters improves convergence with human linguistic processing.

• •Compares LM next-word prediction with human EEG-based event-related potentials during reading
• •Only surprisal correlates with language-processing ERPs, not top-1 prediction accuracy
• •Challenges the assumption that scaling LMs improves alignment with human-like linguistic processing

Generated with AI, which can make mistakes.