Uncertainty Reduction as a Measure of Cognitive Processing Effort

The amount of cognitive effort required to process a word has been argued to depend on the word’s effect on the uncertainty about the incoming sentence, as quantified by the entropy over sentence probabilities. The current paper tests this hypothesis more thoroughly than has been done before by using recurrent neural networks for entropy-reduction estimation. A comparison between these estimates and wordreading times shows that entropy reduction is positively related to processing effort, confirming the entropy-reduction hypothesis. This effect is independent from the effect of surprisal.