You are viewing the SNL 2017 Archive Website. For the latest information, see the Current Website.

Poster E44, Friday, November 10, 10:00 – 11:15 am, Harborview and Loch Raven Ballrooms

Comprehenders Rationally Adapt Semantic Predictions to the Statistics of the Local Environment: A Bayesian model of trial-by-trial modulation on the N400

Gina Kuperberg1,2, Nathaniel Delaney-Busch1, Emily Morgan1, Ellen Lau3;1Department of Psychology, Tufts University, 2Department of Psychiatry and the Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, 3Department of Linguistics, University of Maryland

Introduction: When semantic information has been pre-activated by a context prior to new bottom-up input becoming available, semantic processing of the predicted incoming word is typically facilitated, attenuating the amplitude of the N400 event related potential (ERP). This N400 modulation is observed even when the context is a single semantically related “prime” word. The magnitude of the N400 semantic priming effect is larger in experimental environments that contain a larger proportion of semantically related prime-target pairs (e.g. (1), suggesting that participants adapt the strength of their predictions to the predictive validity of the wider experimental environment. Such adaptation makes rational sense: probabilistic prediction is only beneficial if these predictions actually approximate the statistical structure of the input. In the present study, we asked whether Bayesian principles of rational adaptation can explain how lexico-semantic processing in the brain, as indexed by the N400, adapts over time to changes in the broader experimental environment. Methods: We built a formal computational model of rational adaptation of lexico-semantic processing in a semantic priming paradigm. This model combined three basic assumptions: that contexts probabilistically inform lexico-semantic expectations for upcoming target words, that these expectations adapt rationally (in an optimal Bayesian manner) over time, and that the N400 component is sensitive to the amount of information (surprisal) conveyed by target words. Additional inputs to the model included distributional knowledge that participants could use to generate probabilistic predictions about the target, given the prime (specifically, frequency and forward association strength). We then asked whether this Rational Adapter model could explain how the amplitude of the N400 to target words changes after participants switch from an experimental environment with 10% semantically related prime-target pairs (a lower proportion block) to an environment with 50% of semantically related prime-target pairs (a higher proportion block). Results: The Rational Adapter model was able to explain the trial-by-trial pattern of N400 modulation across the higher-proportion block (β = -2.21, t = -2.76, p = 0.006). We further showed that the explanatory power of this model was not simply due to the inclusion of items-level information like frequency and forward association strength (β = -2.30, t = -2.11, p = 0.036). Finally, we confirmed that, given both probabilistic prediction and rational adaptation, word surprisal was a significantly better predictor of N400 amplitude than raw estimates of word probability. Conclusions: These findings provide evidence that the brain probabilistically predicts upcoming words, and that it updates these probabilistic predictions in a rational trial-by-trial fashion in responses to changes in the broader statistics of its environment. They also provide strong evidence that the N400 component indexes new (unpredicted) information: surprisal. These findings hold implications for theories of language processing, the functional significance of the N400 component, and the design of psycholinguistic experiments (in which participants are likely to adapt to trial probability). (1) Lau, Holcomb & Kuperberg, JCN, 2013

Topic Area: Computational Approaches

Back to Poster Schedule