Predictability effects in reading: linear or logarithmic?

What is it about?

When words are more “predictable” given the previous sentential context, we read them ever-so-slightly faster than we do less predictable words. Researchers have debated the cognitive mechanism(s) behind these predictability effects, leveraging the form of the relationship—whether processing time decreases linearly or logarithmically as a function of a word’s predictability—as one way to understand how the brain processes language. One notable study used a paradigm called self-paced reading (SPR) to investigate this issue, finding a linear relationship. Here, we used the materials from this earlier study in a highly powered eye tracking paradigm, which is more naturalistic than SPR. Contrary to the previous results, we found evidence that predictability effects in reading are logarithmic.

Featured Image

Photo by Thought Catalog on Unsplash

Photo by Thought Catalog on Unsplash

Why is it important?

The question of the form of the relationship between a word’s predictability and its associated reading time is critical to our understanding of the “why” behind predictability effects. One of the prominent explanations, Surprisal theory (Hale, 2001; Levy, 2008), predicts this relationship to be logarithmic. There hasn’t been much experimental work testing this hypothesis directly, and the existing evidence was somewhat ambiguous, with the answer potentially dependent on how researchers estimated predictability (i.e., with cloze norms or LLM-generated values). Our study leveraged the same materials as earlier experimental work, but here we used a more naturalistic paradigm and found compelling evidence for a logarithmic relationship. We also found that while this relationship was clearest when using highly granular, LLM-generated predictability estimates, it still held even with human-derived cloze norms. In sum, we interpret these results as support for Surprisal.

Perspectives