Comprehension effort as the cost of inference

As you read this text, word by word, you build an understanding of its meaning. What cognitive mechanisms underlie this ability? An influential approach to answering this question comes from viewing comprehension as probabilistic inference over potential interpretations given linguistic input. Motivated within this perspective, a wealth of previous literature in psycholinguistics has focused on an important empirical relationship made precise by surprisal theory (Hale, 2001; Levy, 2008a), the hypothesis that the effort required to process a word scales in its negative log probability, in context. However, the standard derivation of surprisal within the inference framework relies on a crucial assumption: that there is a deterministic relationship between the latent interpretations targeted by inference and the observable input. In this work we propose relaxing this assumption and formalize inference cost directly as the amount of change in probabilistic beliefs. This proposal forms a nontrivial generalization of standard surprisal theory, which provides a more direct connection to algorithmic theories, and naturally explains phenomena where unpredictable input requires little processing effort. To test this framework against surprisal theory, we conduct a self-paced reading time study targeting words with orthographic errors, a specific setting where our approach predicts substantially different patterns. We find that processing effort follows the predictions of belief-update rather than surprisal, in a noisy-channel model of comprehension as inference about intended words. These results demonstrate a clear case where surface surprisal cannot explain human processing cost, and provide further support for models of language comprehension as rational inference.