Prediction as a basis for skilled reading: insights from modern language models

Abstract

Reading is not an inborn human capability, and yet, English-speaking adults read with impressive speed. This study considered how predictions of upcoming words impact on this skilled behaviour. We used a powerful language model (GPT-2) to derive predictions of upcoming words in text passages. These predictions were highly accurate and showed a tight relationship to fine-grained aspects of eye-movement behaviour when adults read those same passages, including whether to skip the next word and how long to spend on it. Strong predictions that were incorrect resulted in a prediction error cost on fixation durations. Our findings suggest that predictions for upcoming words can be made based on the analysis of text statistics and that these predictions guide how our eyes interrogate text at very short timescales. These findings open new perspectives on reading and language comprehension and illustrate the capability of modern language models to inform understanding of human language processing.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 36 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.