Strong Prediction: Language Model Surprisal Explains Multiple N400 Effects

Abstract Theoretical accounts of the N400 are divided as to whether amplitude response a stimulus reflects extent which was predicted, is semantically similar its preceding context, or both. We use state-of-the-art machine learning tools investigate these three best supported by evidence. GPT-3, neural language model trained compute conditional probability any word based on words that precede it, used operationalize contextual predictability. In particular, we an information-theoretic construct known surprisal (the negative logarithm probability). Contextual semantic similarity operationalized using two high-quality co-occurrence-derived vector-based meaning representations for words: GloVe and fastText. The cosine between vector representation sentence frame final derive estimates. A series regression models were constructed, where variables, along with cloze plausibility ratings, predict single trial amplitudes recorded from healthy adults they read sentences whose varied in predictability, plausibility, relationship likeliest completion. Statistical comparison indicated GPT-3 provided account suggested apparently disparate effects expectancy, can be reduced variation predictability words. results argued support predictive coding human network.