Poster E70, Monday, March 27, 2:30 – 4:30 pm, Pacific Concourse
Classification of neural responses to contextually constrained sentence endings using single trial EEG data
James J. S. Norton1, Ryan J. Hubbard1, Cybelle Smith1, Timothy Bretl1; 1University of Illinois
Decades of cognitive neuroscience research have demonstrated that the brain rapidly integrates contextual and semantic information during sentence reading. Investigations into the neural basis of these processes -- predominantly using event-related potentials (ERPs) -- have revealed several neural signatures of information integration, including the N400, frontal positivity, and frontal theta responses. ERP investigations, however, are limited in that they assume the average response of many subjects over many trials is representative of single-trial activities in the brain. Several aspects of language comprehension, such as whether or not users make specific word predictions, cannot be explored after averaging. To test whether single trial data can be used to assess these aspects of semantic processing, we performed a classification analysis on EEG data from three separate datasets. The data were collected during three experiments in which subjects read highly constraining sentences with either expected or unexpected but plausible endings. For each of these datasets, we extracted features from the EEG data relating to spatial (using common spatial patterns), temporal, and time-frequency changes in brain activity following the presentation of the final word across two conditions (expected or unexpected). Using linear classifiers (LDA, naive Bayes, SVM), we successfully classified expectancy at accuracies above chance. We also explored classification performance on words from more weakly constraining sentences, as well as on semantically anomalous words. These results provide a first step towards the development of new single trial methodologies investigating language comprehension with EEG, and potentially have additional application to brain-computer interfaces.
Topic Area: LANGUAGE: Other