My Account

Poster C23, Wednesday, August 21, 2019, 10:45 am – 12:30 pm, Restaurant Hall

Early Neural Signals of Prediction During Language Comprehension Are Modulated by Constraint and Expectancy

Ryan Hubbard1, Kara Federmeier1;1University of Illinois, Urbana-Champaign

Numerous studies of language comprehension have identified that individuals can utilize contextual information of a sentence in order to predict or pre-activate upcoming features of words, both with behavioral and electrophysiological measurements (Federmeier et al., 2007). While this research has been fruitful and has led to greater understanding of prediction, the measures utilized are relatively coarse – behavioral responses give little insight into the specific neural processes occurring, and event-related potentials (ERPs) may primarily index more global low-frequency neural changes while missing subtle changes in patterns of activity. Neural representations of linguistic and semantic information are likely distributed across networks in the brain (Maess et al., 2006), and thus multivariate analysis techniques that can detect subtle changes in patterns of activity may be effective for understanding the specifics of what features are pre-activated during comprehension, and when they are activated. To this end, we implemented Representational Similarity Analysis (RSA), a multivariate analysis technique that calculates similarity between neural representations over time or space. We used RSA to examine the similarity of neural representations of expected or unexpected sentence final words and immediately preceding words in high and low constraint sentences (data from Federmeier et al., 2007). If the preceding word cues pre-activation of features of the final word, then representational similarity may be increased in more predictive contexts, but decreased when the final word is unexpected. Time-based RSA revealed a peak of similarity between pre-final and final words 100-250 ms post-onset. The magnitude of this peak was graded based on constraint and expectancy, with strong constraint expected endings having the greatest similarity. Space-based RSA of this time-window showed that representational similarity was greatest over occipital sites, suggesting the pre-final word cued pre-activation of visual features of the sentence final word. We extended the time-based RSA analysis with a generalization approach to examine similarity between early activity following the pre-final word and later activity (in the N400 time window) of the final word. This analysis revealed a striking pattern – namely, similarity between early pre-final word activity and late final word activity was lowest for strongly expected endings, and was in fact negative, suggesting dissimilarity or repulsion (Chanales et al., 2017). These findings suggest that prediction during language comprehension is multi-dimensional, resulting in pre-activation of not just semantic but also visual features, and can occur rapidly, allowing for fast matching of incoming information to pre-activated representations. However, once representations are verified, they may be inhibited to focus on other processing, leading to downstream consequences of impoverished representations for predicted information (Rommers & Federmeier, 2018).

Themes: Meaning: Combinatorial Semantics, Meaning: Lexical Semantics
Method: Electrophysiology (MEG/EEG/ECOG)

Back