Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
MEG/EEG responses to spoken ambiguous words reveal neural representations of selected meanings but not prediction error during sentence comprehension
Poster B28 in Poster Session B, Tuesday, October 24, 3:30 - 5:15 pm CEST, Espace Vieux-Port
Victoria Poulton1, Mate Aller1, Lucy MacGregor1, Matt Davis1; 1University of Cambridge
This preregistered MEG/EEG investigates the neural responses to ambiguous words during spoken sentence comprehension. Participants (N=34) listened to semantically constraining sentences such as “While sailing down the river, she noticed the trees along the…”. Sentence-final words were predictable ambiguous homophones in British English (AMB; e.g., “bank”); selected-meaning synonyms (SMS; e.g., “shore”); or alternative words corresponding to a contextually-inconsistent interpretation of the homophone (ALT; e.g., “cashpoint”). To understand ambiguity, listeners must select the correct meaning and suppress the incorrect alternative. This design contrasts two neural computations for combining contexts and words: “sharpening”, in which representations of predictable meanings are enhanced; and “prediction error”, in which the brain represents the difference between the predictable meaning and heard words. While both computations can underlie Bayesian inference, reductions in ERP magnitude for predicted meanings cannot distinguish between the accounts as both explain changes in overall neural activity (Aitchison & Lengyel, 2017). We therefore conducted multivariate representational similarity analyses (RSA; Guggenmos et al., 2018) on pairs of trials since these accounts make contrasting predictions for the similarity of neural responses: whereas sharpening predicts similarity for AMB and SMS trials (bank-vs-shore), prediction error predicts similarity for AMB and ALT trials (bank-vs-cashpoint). Multivariate dissimilarity is measured via correlation distance (0 indicates low dissimilarity), as this metric is less susceptible to differences in response magnitude between conditions. For each subject, we calculated a timeseries of distances for each contrast. Non-parametric cluster-based permutation tests (Maris & Oostenveld, 2007) were used for group-level statistical comparisons of the dissimilarity timeseries. For evoked responses (spatial patterns per time sample), we found evidence for sharpening (i.e., low dissimilarity for AMB-SMS words compared to a baseline contrast of predictable but semantically unrelated word). Significant differences in dissimilarity were observed at both MEG magnetometers and EEG sensors between approximately 200-400ms following the onset of the sentence-final word. However, no significant differences were observed for the AMB-ALT timeseries compared to its analogous baseline. Since correlation distance limits the influence of magnitude on our comparisons, we are confident that shared semantic information is a key contributor to the low dissimilarity for the AMB-SMS timeseries. We also observed a significant difference between our baseline conditions such that the dissimilarity is lower for predictable words, in the absence of shared semantics. This result highlights a need for further investigation into the contributions of predictability and shared semantics on observed dissimilarities. We will discuss these results in comparison with other conflicting findings in the literature concerning the representation of sharpening and prediction error computations for phonological and semantic information in speech, as well as explore the insights of multivariate and univariate approaches in the interpretation of N400 effects seen for anomalous words. We will also present the results of ongoing RSA of time-frequency features (power and phase), since previous research shows that neural responses in different frequency bands (theta, beta) is associated with predictability and violation detection (Heilbron, 2022). Overall, these results support a multivariate approach to studying ambiguity resolution and the neural representation of semantic computation during sentence comprehension.
Topic Areas: Meaning: Lexical Semantics, Speech Perception