Top-down information flow drives lexical access when listening to continuous speech

Speech is noisy, ambiguous and complex. Here we study how the human brain uses high-order linguistic structure to guide comprehension. Twenty-one participants listened to spoken narratives while magneto-encephalography (MEG) was recorded. Stories were annotated for word class (specifically: noun, verb, adjective) under two hypothesised sources of information: (i) "bottom-up": the most common word class given by the word's phonology; (ii) "top-down": the true word class given the syntactic context. We trained a classifier on trials where the two hypotheses matched (about 90%), and tested the classifier on trials where they mismatched. The classifier predicted only the syntactic word class labels, in line with the top-down hypothesis. These effects peaked around 400ms after word offset over frontal MEG sensors. Our results support that when processing continuous speech, lexical representations are quickly built in a context-sensitive manner. We showcase the utility of multivariate analyses in teasing apart subtle representational distinctions from neural time series. ### Competing Interest Statement The authors have declared no competing interest.