Slide Slam M7
N400 amplitudes reflect change in a probabilistic representation of meaning: Evidence from large scale modelling
Alessandro Lopopolo1, Milena Rabovsky1; 1University of Potsdam
The N400 component of the event-related brain potential (ERP) is widely used in research on language and meaning processing, but its functional basis remains actively debated (Kutas & Federmeier 2011). Recent work showed that the update of the predictive representation of sentence meaning (semantic update, or SU) generated by the Sentence Gestalt (SG) model (McClelland et al. 1989), a neural network model of sentence comprehension, consistently displayed a similar pattern to the N400 amplitude in a series of conditions known to modulate this event-related potential, suggesting that the N400 might reflect the change in a probabilistic representation of meaning corresponding to an implicit semantic prediction error (Rabovsky et al. 2018). In previous work, the model was trained on a small artificial training corpus and thus could not be presented with the same naturalistic stimuli presented in empirical experiments, making the testing of the hypothesis implemented in the model also somewhat indirect and based on the assumption that the small synthetic environment adequately captures the relevant statistical properties of human environments. In the present study, we attempt to directly predict the amplitude of the N400 generated during sentence processing by using as predictor the update of the inner representation of a SG model trained on a large corpus of naturalistic texts (part of the Rollenwechsel-English corpus (Sayeed et al. 2018). We used EEG data collected while subjects were asked to read sentences (Frank et al. 2015). We fit a linear mixed effect model predicting the N400 amplitude as a function of the SU over the stimulus words. The results indicate that SU significantly predicts the amplitude of the N400 (β = 0.07, z = 9.52, p < 0.001). Larger word-wise updates of the SG layer representation correspond with a stronger negative deviation of the ERP signal in the N400 time segment. Moreover, to assess the contribution of the SU on the amplitude of the N400 beyond the effect of surprisal, previously shown to predict N400 amplitudes (Frank et al. 2015), we fitted two nested linear mixed-effects models, one containing as predictors only surprisal, the other containing also the SU. The log-likelihood test between the two models showed that the fit of the model including SU was significantly better (χ2 = 79.03, p < .0001). Even with the presence of surprisal (β = −0.06, z = −8.08, p < 0.001), SU makes a significant contribution to the amplitude of the N400 (β = 0.06, z = 8.90, p < 0.001). The present analyses showed a significant relation between the amplitude of the N400 and the update of the probabilistic semantic representation (SU) generated by the corpus-trained SG model. Further analyses indicate that word position, word frequency, and surprisal have similar effects on the SU as they have on N400 amplitudes. These results suggest that the SU is a valid approximate of the N400, in line with the hypothesis that its amplitudes reflect the change in a probabilistic representation of sentence meaning corresponding to an internal prediction error at the level of meaning.