Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Effects of target entropy and word surprisal on fixation duration, fixation location, and brain potentials
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Lucas Y. H. Chan1, Urs Maurer1,2; 1Chinese University of Hong Kong, 2Brain and Mind Institute, Chinese University of Hong Kong
Reading engages an action-perception cycle in the brain. It requires constant coordination between eye-movement control and the visual perception system. A theory has suggested that action and perception can be simulated by minimizing "expected free energy" and "variational free energy" respectively. In reading, "expected free energy" can be minimized through programming saccades towards a potentially more informative target (i.e., higher information gain), while "variational free energy" can be minimized through prediction error signals generated after perceiving the target word. We tested this notion by investigating the effects of LLM-derived target entropy and word surprisal on eye movement and EEG during a natural reading task. Here, higher target entropy represents higher contextual information gain, while higher word surprisal represents stronger contextual prediction error. We expected that target entropy would influence saccade target selection and pre-saccadic brain activity, and word surprisal would influence fixation durations and post-fixation brain activity. For fixation duration, we found that words with higher surprisal received longer fixation durations in all duration measures (i.e., first fixation, gaze, single fixation, and total fixation duration). Fixation towards higher entropy targets only resulted in shorter gaze duration. For fixation location, we found that higher entropy targets and higher surprisal words were more likely to be fixated. We also found that first fixations towards higher entropy target were more likely to land on the center of a word. This effect was not observed for word surprisal. For brain potentials, a cluster-analysis from 300 ms to 0 ms before saccade onset revealed a significant cluster elicited by target entropy that started at about 50 ms prior to saccade onset. Cluster analysis on the word surprisal effect from 0 ms to 500 ms after fixation onset showed a significant cluster that resembles the traditional N400 effect elicited by cloze probability. These results provide preliminary evidence that target entropy and word surprisal may indeed capture the underlying parameters governing the action-perception cycle during reading comprehension.
Topic Areas: Reading,