Séminaire de Recherche en Linguistique

Modelling the Human Language Model

Surprisal theory posits that the cognitive effort required to comprehend a linguistic unit depends on the unit’s surprisal—its negative log probability given prior context. While this relationship is well-supported empirically, reducing explanations of language comprehension to a single measure is more often than not an oversimplification. In this talk, I will present a family of sampling-based uncertainty measures that generalise surprisal theory and allow expressing a wider range of incremental processing hypotheses about the workings of language comprehension. I will show how these uncertainty measures, estimated via Monte Carlo sampling with neural network language models, exhibit predictive power comparable or superior to surprisal for a range of human neural and behavioural processing signatures, including eye-tracked and self-paced reading times, event-related brain potentials, cloze responses, and acceptability judgments.