Abstract

Neural language model gradients predict event-related brain potentials

---

Abstract

Fitz & Chang (2019) argue that event-related brain potentials during sentence comprehension result from the detection and incorporation of word-prediction error. Specifically, the N400 component would correlate with prediction error while the P600 component would be indicative of error backpropagation in the language system. The current work evaluates this hypothesis on a corpus of EEG data recorded during naturalistic sentence reading. Word-prediction error and backpropagated error were estimated by an LSTM language model that processed the same 205 English sentences as the human participants. At each word, the word's surprisal and the total gradient of recurrent-layer connections were collected for comparison to the sizes of the N400 and P600 components. Consistent with the theory, higher surprisal resulted in stronger N400 while higher gradient resulted in stronger P600, and ERPs on content words were more sensitive to surprisal whereas ERPs on function words were more sensitive to gradient. However, a detailed analysis of the neural signal's time course indicates that the apparent P600 effect could be interpreted as a reversed N400 effect.

Keywords: event-related brain potentials, N400, P600, sentence processing, LSTM model, prediction error, error propagation

How to Cite:

Frank, S., (2024) “Neural language model gradients predict event-related brain potentials”, Society for Computation in Linguistics 7(1), 316–323. doi: https://doi.org/10.7275/scil.2215