Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Late Frontal Positivities during self-paced reading are sensitive to lexical predictability, not misprediction: Evidence from EEG and Predictive Coding simulations
Poster Session E, Sunday, September 14, 11:00 am - 12:30 pm, Field House
Anthony Yacovone1,3,4, Samer Nour Eddine2,3, Thomas Hansen3,4, Gina Kuperberg3,4; 1Boston University, 2University of Maryland, College Park, 3Tufts University, 4Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital
INTRODUCTION: It is well established that predictable words evoke smaller N400s relative to unpredictable words. Within a predictive coding (PC) framework, the N400 can be simulated as bottom-up residual error, i.e., information within the input that has not been reconstructed (predicted) by higher levels. In studies with unexpected (but plausible) words that disconfirm strong predictions, the N400 is often followed by a late frontal positivity (LFP, 600–1000ms). In PC, this response may reflect top-down residual error, which contains “mispredicted” information at higher levels that does not appear in the input. A challenge, however, is that LFPs can also be observed in less constraining contexts, where mispredictions are unlikely to occur. While the conditions that produce LFPs remain unclear, they are often seen during self-paced reading—possibly because self-pacing encourages active engagement with each word. When present, these LFPs are sensitive to lexical predictability (smaller, less positive-going LFPs to more predictable words). Here, we ask (1) during self-paced reading of short stories, are LFPs routinely evoked by all content words, and (2) can we reconcile both LFP patterns within a single PC framework? METHOD AND RESULTS: (ERPs) We recorded EEG while participants (N=22) self-paced through short, engaging vignettes. We measured the N400 (300–500ms, centroparietal sites) and LFP (500–1000ms, frontal sites) evoked by each content word. Predictability was operationalized using an online cloze task (N=361). Linear mixed effects models revealed an effect of Predictability (target cloze probability) but not Misprediction (the probability of the most common “mispredicted” word) on the N400, replicating its sensitivity to cloze and insensitivity to contextual constraint. We found similar patterns on the LFP with smaller (less positive-going) responses to more predictable target words (regardless of the degree of mispredicting). (Simulations) We implemented a PC model of lexico-semantic processing. To simulate lexical predictability, top-down activation was fed into the highest conceptual level, while bottom-up input was presented to the lowest orthographic level. N400s and LFPs were operationalized as the magnitude of bottom-up and top-down residual error (RE), respectively, as the PC algorithm converged. In all simulations, the bottom-up RE (simulated N400) was sensitive to predictability and not misprediction. In contrast, the effects of predictability and misprediction on top-down RE (simulated LFP) varied depending on the number of iterations over which top-down and bottom-up inputs were presented, as well as the extent of overlap between these phases. For longer durations of top-down pre-activation and bottom-up presentation, the pattern of top-down RE mirrored the LFP effects from our self-paced ERP experiment (larger LFPs to less predictable inputs). This top-down RE effect arose because newly activated, unpredicted higher-level semantic information was retroactively propagated to lower-level lexical states that had not yet encoded it. In contrast, when the model was given limited time to process the input, the pattern of top-down RE more closely resembled patterns of LFPs observed in previous experimental studies using faster RSVP, i.e., the simulated LFP was primarily sensitive to misprediction. In our model, this was due to continued propagation of mispredicted higher-level representations to the lower lexical level.
Topic Areas: Meaning: Lexical Semantics, Computational Approaches