Hierarchical Predictive Processing during Natural Reading

Abstract

Key neural mechanisms underlying reading, especially in the context of sentences and texts, remain elusive. An influential theory states that the brain rapidly generates a hierarchy of probabilistic predictions about upcoming words at multiple levels of linguistic representation. When encountering a word, a reader integrates sensory input with the linguistic predictions, and rely on the resulting prediction error signals to achieve efficient comprehension. Here, we tested this hierarchical predictive account with a natural reading task while EEG and eye movements were simultaneously recorded. We find distinct prediction error signals attributable to word contextual, lexical, and orthographic levels of predictions. These prediction error signals exhibit a temporal order. Our results also indicate that context-based predictions rapidly constrain orthographic-level predictions. We thus provide evidence for a hierarchical predictive process central to efficient natural reading.