Overall, our findings show remarkable longevity and parallel processing across the language hierarchy, which is supported by a dynamic spatiotemporal code. This time-stamps inputs for their relative position, enabling integration across levels, and across ordered elements in the sequence.

This dynamic code serves the computational function of avoiding destructive interference between neighbouring inputs, and provides an implicit "time code" for the relative position of items in a sequence!

8/8

Finally, we find that each level of the hierarchy is encoded in a dynamic ensemble of neural patterns, and neural activity unfolds faster for more sensory features like parts of speech sounds, and unfolds more slowly for more symbolic features like word meaning and sentence structure.

7/8

When we group the features into 6 hierarchical levels, we find that the hierarchy is processed in a remarkably parallel manner, highly overlapping in time, with long-lived neural responses. Also, higher order features are decodable earlier than lower order features - a "reverse hierarchy".

6/8

We confirm that a rich hierarchy of 50 language features can be decoded from neural activity.

5/8

21 participants listened to short stories while we recorded neural activity with MEG. We annotated those stories for a comprehensive set of language features, from the parts of individual speech sounds, to the meaning of words, and the structure of sentences.

4/8

We tested the hypothesis that each level of the hierarchy is encoded dynamically, where different neural patterns are activated in temporal sequence. We call this "Hierarchical Dynamic Coding".

3/8

To process speech, the brain generates a rich hierarchy of representation, from sound to meaning.

2/8

Cool!