Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Identifying EEG activity related to the linguistic processing of natural language.
Poster Session A - Sandbox Series, Thursday, October 24, 10:00 - 11:30 am, Great Hall 3 and 4
This poster is part of the Sandbox Series.
Jin Dou1, Andrew J. Anderson2, Aaron R. Nidiffer1, Aisling E. O’Sullivan3, Jens Hjortkjær4,5, Alain De Cheveigné6,7, Edmund C. Lalor1; 1University of Rochester, Rochester, United States of America, 2Medical College of Wisconsin, Milwaukee, United States of America, 3Trinity College, Dublin, Ireland, 4Technical University of Denmark, Kgs. Lyngby, Denmark, 5Copenhagen University Hospital - Amager and Hvidovre, Copenhagen, Denmark, 6Laboratoire des Systèmes Perceptifs, Paris, France, 7Ecole Normale Supérieure, Paris, France
Speech is central to human life. An enormous amount of research has been conducted on the neurobiology of spoken language. In recent years, this has included significant effort aimed at modeling brain responses to natural speech based on different acoustic and linguistic features of that speech. In the case of EEG research, one major unresolved issue centers on how to determine the success of these models. More specifically, it is unknown how much variance in EEG responses to speech derives from the linguistic processing of that speech rather than the obligatory sensory processing of the speech acoustics. Indeed, given the low spatial resolution of EEG, activity related to acoustic and linguistic processing is typically mixed into the same EEG recordings channels. The goal of the present project was to distinguish EEG activity related to linguistic processing from EEG activity related to low-level stimulus feature processing. We did this based on the hypothesis that activity in language processing regions will be similar for reading and listening, while activity in sensory (i.e., auditory and visual) regions will be different across those two conditions. Specifically, we recorded EEG from 10 participants who were presented with segments of a book in both audio and text format with their orders balanced across participants. Crucially, the text presentation was controlled to have the same timing as the audiobook. As a result, we could directly relate EEG data from the reading and listening conditions in an effort to identify any common activity that – by hypothesis – should necessarily be related to language processing. To identify common signals across all 10 participants and 2 modalities, we used multi-way canonical correlation analysis (MCCA). To avoid learning correlations between low-level activity in auditory and visual cortices across channels (based on common timing), we used MCCA to extract a single canonical component (CC) shared across all subjects and modalities for each EEG channel separately. However, including a range of time lags in the analysis allowed us to do this while accounting for any differences in the temporal profile of EEG responses between participants or modalities. We then conducted a series of additional analyses to validate if the CCs extracted for individual channels reflected linguistic processing. We found that: 1) comparing the spatial distribution of pairs of CCs from the two modalities revealed common activity around centro-parietal scalp; 2) the average of all CCs for each channel, predicted EEG most accurately also around the centro-parietal scalp; and 3) the strong activation around centro-parietal scalp remained even after regressing out EEG activity related to lexical surprisal. The fact that distribution of common activity across reading and listening is strongest over scalp regions that have been implicated in linguistic processing in previous research, leads us to conclude that we have successfully isolated linguistic processing from lower-level sensory processing in reading and listening. Notably, the variance of this EEG activity is not fully explained by lexical surprisal – which is a linguistic feature that has been strongly linked with EEG responses to natural language.
Topic Areas: Methods, Computational Approaches