Constituent-constrained word prediction during language comprehension

Abstract

Next-word prediction has been hypothesized as the central computational objective of the human language system, akin to that of current large language models. Here we put this conjecture to the test, investigating whether the brain predicts each upcoming word as precisely as possible when listening to connected speech. In three magnetoencephalography experiments with Mandarin Chinese speakers, we demonstrate that the response related to word unpredictability, that is, word surprisal calculated using large language models, is significantly stronger for words within an ongoing constituent than words across a major constituent boundary, and this effect is further modulated by the certainty of a constituent boundary. This constituent-boundary effect is also observed behaviorally unless speech is very slowly presented, and it is confirmed by analyzing a dataset of electrocorticography responses to natural English narratives. The constituent-boundary effect demonstrates that the language system does not solely optimize word-prediction precision; rather, it balances word-prediction contributions by constituent-constrained management of linguistic contextual representations. Zou et al. reveal a key difference between human brains and large language models (LLMs). While LLMs are optimized to predict the next word, the human brain modulates prediction efficiency by strategically grouping words into phrases.