Are word predictability effects really linear? A critical reanalysis of key evidence

Find paper here

One prominent divergence between procedural and inferential theories of sentence comprehension concerns whether processing cost is linear or logarithmic in word predictability. We revisit the evidence for a linear effect reported by Brothers & Kuperberg (2021), reanalysing their self-paced reading and cross-modal picture naming experiments using predictability estimates from GPT-2 in addition to the original cloze and trigram estimates. We find that, while the cloze effect is linear on a raw-probability scale, the GPT-2 effect is superlinear—favouring a logarithmic relationship—and argue that the best available synthesis of the evidence is that predictability effects primarily reflect the costs of probabilistic inference, rather than predictive preactivation.

@inproceedings{shain-etal-2024-linear,
    author = {
        Cory Shain and
        Clara Meister and
        Tiago Pimentel and
        Ryan Cotterell and
        Roger Levy
    },
    booktitle = {Human Sentence Processing Conference (HSP)},
    title = {Are word predictability effects really linear? A critical reanalysis of key evidence},
    year = {2024},
    url = {https://hsp2024.github.io/abstracts/submission_250.pdf},
    pages = {},
}