Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Understanding the Influence of Surprisal in Joke Comprehension through N400, P600, and Late Frontal Positivity
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.
Poster C14 in Poster Session C, Wednesday, October 25, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Haoyin Xu1, Masaki Nakanishi2, Sean Trott1, Seana Coulson1; 1University of California, San Diego, 2Swartz Center of Computational Neuroscience, University of California, San Diego
To understand a joke, comprehenders need to incorporate unexpected information into their discourse models. For example, in the joke “My doctor has never violated his oath: the oath he took years ago to become a millionaire,” the word ‘millionaire’ is unexpected. However, it also prompts us to activate different aspects of background knowledge regarding doctors. Here we use regression models to examine whether single trial EEG amplitude associated with language ERP components, namely N400, P600, and late frontal positivity (LFP), are sensitive to the predictability of sentence-final words in jokes and in non-funny controls, utilizing a predictability measure derived from the large language model GPT3. EEG was recorded from the scalp as healthy adults read 80 Jokes, 80 Straights, viz., non-funny controls in which the critical word in the joke was replaced by a word with a similar cloze probability, and 80 Expecteds, filler sentences with predictable sentence final words. After preprocessing, single trial EEG amplitude was measured by averaging across relevant time windows (N400: 300 - 500ms, LFP and P600: 700 - 900ms) at six electrodes for each component based on the ERP language literature. The surprisal of each critical word was computed by taking the negative log of word probability provided by GPT3. A series of mixed effect models were constructed to predict the amplitude of single trial N400, P600, and LFP. All models included random intercept terms for each critical word, subject, and channel; fixed effects included experimental condition (i.e. joke, straight, and expected), and surprisal. Akaike information criterion (AIC) was used for statistical model comparisons. Single factor regressions revealed N400 amplitude was significantly predicted by surprisal (p < 0.001) and experimental condition (condition[Joke] p < 0.001; condition [Straight] p < 0.001). In the additive model including surprisal and experimental conditions, surprisal remained significant (p = 0.02), as did the condition effect for jokes (p = 0.004); the condition effect for straights was no longer significant (p = 0.09). These analyses suggest that while N400 elicited by unexpected straights was captured by predictability estimates from the language model, N400 elicited by jokes was not. In addition, single factor regressions also indicated that P600 amplitude was significantly predicted by experimental condition in jokes (condition[Joke]: p = 0.012) but not straights (condition[Straight]: p = 0.17); surprisal was not a significant predictor (p = 0.16). These analyses suggest the enhanced P600 elicited by jokes indexes processing demands not directly related to the predictability of the word in context. Because the LFP is typically elicited only in high constraint contexts, joke/straight sentences were divided into high and low constraint items and modeled separately. For high constraint materials, LFP amplitude was significantly predicted by surprisal (p < 0.001), and by experimental condition for jokes (condition[Joke]: p <0.001) and but not straights (condition[Straight]: p = 0.355). In models of low constraint materials, none of the predictors were significant. This analysis suggests that in high constraint sentences, jokes elicited enhanced LFP that may reflect the detection of situation-model level violations.
Topic Areas: Meaning: Discourse and Pragmatics, Syntax and Combinatorial Semantics