(Early) context effects on event-related potentials over natural inputs

ABSTRACT

Language understanding requires the integration of the input with preceding context. Event-related potentials (ERPs) have contributed significantly to our understanding of what contextual information is accessed and when. Much of this research has, however, been limited to experimenter-designed stimuli with highly atypical lexical and context statistics. This raises questions about the extent to which previous findings generalise to everyday language processing of natural stimuli with typical linguistic statistics. We ask whether context can affect ERPs over natural stimuli early before the N400 time window. We re-analyse a data set of ERPs over ∼700 visually presented content words in sentences from English novels. To increase power, we employ trial-level ms-by-ms linear mixed-effects regression simultaneously modelling random variance by subject and by item. To reduce concerns about Type I error inflation common to time series analyses, we introduce a simple approach to model and discount auto-correlations at multiple, empirically determined, time lags. We compare this approach to Bonferroni correction. Planned follow-up analyses employ Generalized Additive Mixed Models to assess the linearity of contextual effects, including lexical surprisal, within the N400 time window. We found that contextual information affects ERPs in both early (∼200 ms after word onset) and late (N400) time windows, in line with a cascading, interactive account of lexical access.

KEYWORDS:

• Event-related potentials
• time course analysis
• mixed-effects regression
• generalized additive mixture models
• auto-correlation

Acknowledgments

We are grateful to Dr. Stefan Frank for sharing the data, including EEG data, sentence materials, predictors from language models, and for patiently answering all our questions about his original experiments. We would also like to thank Dr. Gina Kuperberg for providing insightful feedbacks on earlier drafts of the paper. We really appreciate the highly constructive feedback of Dr. Milena Rabovsky, another anonymous reviewer and the editor. We would also like to thank the members of the Human Language Processing Lab – in particular, Zachary Burchill, Wednesday Bushong, Dr. Linda Liu – for proof reading earlier versions of the paper. All remaining mistakes are our own.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 We originally also examined a related alternative measures of semantic distance, based on an interpretation of context vectors as probability distributions over latent semantic spaces. This measure correlates neither with early ERP components nor with N400 amplitude. The results reported here do not change if this predictor is included in the analysis.

2 Additionally, the smoothing technique employed in the n-gram model (Kneser-Key smoothing, Chen & Goodman, 1999) uses back-off to smooth unreliable estimates of n-gram probabilities. While this is an effective smoothing technique and thus desirable, it can further increase the correlation between the estimated n-gram probabilities and word frequency.

3 Typically, the ERP baseline is simply deducted from the ERP wave to perform baseline correction. This assumes that the effect of baselines is constant across ERPs at different times. Here we include the ERP baseline as a regressor to keep our results comparable to Frank and Willems (2017). This decision does not affect our results. As reported in the appendix, the ERP baseline had different effect on different parts of the ERP signal.

Additional information

Funding

This work was funded by the NICHD R01 grant [HD075797] to T. Florian Jaeger; Division of Human Development.