Eye movements reveal a dissociation between prediction and structural processing difficulty in language comprehension

Abstract

Significance When we read sentences, our eyes move back and forth in complex ways. One factor that explains this pattern is prediction: Our experience allows us to anticipate upcoming words, making it is easier to recognize predictable ones. What are the limits on what can be explained by prediction? Here, we use contemporary language models and recordings of eye movements from hundreds of participants to answer this question. We find that language models provide only a partial explanation: Their next-word probability estimates explain the early stages of reading, but not the difficulty of integrating a word into the context, reflected in backward eye movements and rereading. This highlights the challenge of creating cognitive models that can explain all stages of reading.