What's Surprising About Surprisal.

Abstract

In the computational and experimental psycholinguistic literature, the mechanisms behind syntactic structure building (e.g., combining words into phrases and sentences) are the subject of considerable debate. Much experimental work has shown that surprisal is a good predictor of human behavioral and neural data. These findings have led some authors to model language comprehension in a purely probabilistic way. In this paper, we use simulation to exemplify why surprisal works so well to model human data and to illustrate why exclusive reliance on it can be problematic for the development of mechanistic theories of language comprehension, particularly those with emphasis on meaning composition. Rather than arguing for the importance of structural or probabilistic information to the exclusion or exhaustion of the other, we argue more emphasis should be placed on understanding how the brain leverages both types of information (viz., statistical and structured). We propose that probabilistic information is an important cue to the structure in the message, but is not a substitute for the structure itself-neither computationally, formally, nor conceptually. Surprisal and other probabilistic metrics must play a key role as theoretical objects in any explanatory mechanistic theory of language processing, but that role remains in the service of the brain's goal of constructing structured meaning from sensory input.

Supplementary information: The online version contains supplementary material available at 10.1007/s42113-025-00237-9.