Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Distinct mechanisms of inhibition in language production
Poster Session D, Saturday, September 13, 5:00 - 6:30 pm, Field House
Nazbanou Nozari1, Tara Pirnia2, Svetlana Pinet3, Leila Wehbe4; 1Indiana University, 2University of Pittsburgh, 3Basque Center on Cognition, Brain and Language, 4Carnegie Mellon University
Introduction. It is widely assumed that behavioral interference signals representational conflict, tapping inhibitory control. Some assume this inhibitory control to be a unitary function, regardless of task. In contrast, theoretical models distinguish between tasks in which stimulus-driven information is sufficient to arrive at the correct response vs. those in which such information leads to the incorrect response. An example of the former is contextual similarity. A picture of a “cat” is harder to name in the context of semantically/phonologically (“dog”/”mat”) similar pictures, but the target picture ultimately drives processing more strongly towards the correct answer. In contrast, tasks like picture-word-interference contain stimuli that naturally elicit a prepotent incorrect response unless overridden by task goals. The current project uses robust machine-learning (ML) techniques for analyzing EEG data to examine the generalizability of the neural states representing conflict across these two task types. Methods. Twenty-eight native speakers of English completed a picture-naming paradigm. In each block, two pictures were introduced and then appeared randomly for 8 trials each. Two manipulations were embedded: (a) A contextual-similarity manipulation: 5 blocks containing semantically-similar pairs, 5 phonologically-similar pairs, and 5 unrelated pairs. (b) A name-reversal manipulation: in the first half of each block, participants called pictures by their canonical names. In the second half, they called each picture by the name of the other picture in that block, creating a Stroop-like effect. Behavioral data were collected along with EEG data using a 128-channel system. We conducted behavioral analyses, traditional stimulus- and response-locked EEG analyses, a representational similarity analysis (RSA), and a set of ML-based analyses. The ML techniques have the advantage of using the entire data structure and timeline, without cherry-picking electrodes and temporal windows. Three classifiers (linear ridge, support vector machine, and decision tree) were trained to decode neural signatures of conflict within tasks. The best-performing classifier, evaluated using nested k-fold cross-validation, was selected for subsequent analyses. Additionally, we performed temporal generalization analyses to map the stability and generalization of neural patterns over time, as well as a leave-one-out cross-validation analysis to investigate across-participant generalizability. Results. Behaviorally, both manipulations created interference. Traditional EEG analyses and RSA showed stark differences between the neural states induced by contextual-similarity and name-reversal. Five sets of analyses were conducted using the ML technique. The first two showed robust classification abilities within each task. The third set showed successful cross-classification of semantic and phonological similarity within the contextual-similarity manipulation. The fourth set showed successful classification of new participants’ data within each task. Finally, the fifth set showed no evidence of successful classification across the two manipulations. Conclusions. We found that conflict states are generalizable within tasks, and even to new participants in those tasks, showing the ML method’s power in detecting generalization when present in the neural data. However, the neural states induced by conflict in the two tasks were shown to be unequivocally distinct. This finding is compatible with theoretical models of inhibitory control that distinguish between conflict that can or cannot be resolved solely by stimulus-driven processing.
Topic Areas: Control, Selection, and Executive Processes, Language Production