Recent studies link surprisal —a measure of conditional probability of words in context—to the N400 component size in event-related potentials (ERP), supporting a role for predictive coding in language comprehension. An alternative account argues that N400 variations are better explained by a retrieval mechanism sensitive to the semantic similarity between a word and its preceding context. Because jokes often rely on the presence of unexpected words that relate to the prior context multiple ways, they afford observation of the relative importance of contextual predictability and contextual similarity. We employed state-of-the-art machine learning to assess the surprisal and contextual semantic similarity of critical words in jokes and control stimuli. Using regression models to predict ERP, we found contextual similarity best explains N400 and P600 responses, supporting the semantic similarity account. Additionally, jokes elicit enhanced N400 and P600 responses that go beyond that attributable to their surprisal and contextual semantic similarity.

Authors:

Haoyin Xu: University of California, San Diego; Masaki Nakanishi: 1986; Seana Coulson: UC San Diego