What is the shape of semantic predictions?

It is common for language researchers to claim that humans consistently make predictions about upcoming linguistic input, such as the meaning of an upcoming word.  While many researchers agree that semantic predictions are graded in nature, rather than all-or-none, there is little direct work exploring the contours of these predictions. In this talk, I attempt to answer several questions about semantic predictions, including: How do raters and readers respond as presented words become more and more semantically distant from an expected word? How semantically distant must a presented word be before it becomes anomalous and triggers a ‘prediction error’? How do semantic predictions change as contextual constraint increases?  Do modality (offline ratings vs. online reading) and time (early processing vs. late) influence these patterns? Is prediction an obligatory or optional process? To answer these questions, I rely on a variety of data, including corpus analyses, semantic acceptability ratings, and eye tracking data.