Two forms of uncertainty hierarchically shape lexical predictions during natural speech processing

Speech comprehension relies on contextual lexical predictions to optimize the processing of incoming words. As in other predictive paradigms, it is commonly assumed that prediction error, or surprisal, is weighted by a single summary statistic: uncertainty. Uncertainty reflects the overall unpredictability of the context and is typically quantified using Shannon entropy. However, for predictions involving many alternatives, as in speech, Shannon entropy is an ambiguous measure. We investigated alternative measures of uncertainty, namely the family of Rényi entropies, of which Shannon entropy is a special case. Using whole-brain intracranial EEG recordings in 33 epilepsy patients listening to natural, continuous speech, we show that, rather than Shannon entropy, contextual lexical uncertainty in speech is best captured by two extremes of the Rényi family. These reflect the probability of the most likely word and the spread across all plausible alternatives, which we term strength and dispersion, respectively. These representations are processed in distinct neural populations in a spatiotemporal hierarchy, with the encoding of dispersion preceding that of strength. Around each word onset, information flows between these clusters in a top-down and subsequently bottom-up sequence. Finally, they interact differentially with word surprisal during word processing. Our findings demonstrate that the brain encodes a multidimensional representation of uncertainty and reveal that multiple summary statistics may coexist in predictive processing. More broadly, this work introduces a flexible framework for characterizing the neural bases of uncertainty. It also highlights the potential mismatch between formal definitions of cognitive variables and their neural implementation.