Agreement errors are predicted by rational inference in sentence processing
People are able to understand language in challenging settings which often require them to correct for speaker errors, environmental noise, and perceptual unreliability. To account for these abilities, it has recently been proposed that people are adapted to correct for noise during language comprehension, via rational Bayesian inference. In the present research, we demonstrate that a rational noisy-channel framework for sentence comprehension can account for a well-known phenomenon—subject-verb agreement errors (e.g. The key to the cabinets are…). A series of experiments provides evidence that: a) agreement errors are associated with misrepresentations of the sentence consistent with noisy-channel inferences and b) agreement errors are rationally sensitive to environmental statistics and properties of the noise. These findings support the hypothesis that agreement errors in production result in part from a sentence comprehension mechanism that is adapted to understanding language in noisy environments.