From Thought to Action: How a Hierarchy of Neural Dynamics Supports Language Production

Mingfang, Zhang, J P Lévy, Stéphane d’Ascoli, Jérémy Rapin, F.‐Xavier Alario, Pierre Bourdillon, Svetlana PinetarXiv 2025 · DOI · arXiv

Abstract: Humans effortlessly communicate their thoughts through intricate sequences of motor actions. Yet, the neural processes that coordinate language production remain largely unknown, in part because speech artifacts limit the use of neuroimaging. To elucidate the unfolding of language production in the brain, we investigate with magnetoencephalography (MEG) and electroencephalography (EEG) the neurophysiological activity of 35 skilled typists, while they typed sentences on a keyboard. This approach confirms the hierarchical predictions of linguistic theories: the neural activity preceding the production of each word is marked by the sequential rise and fall of context-, word-, syllable-, and letter-level representations. Remarkably, each of these neural representations is maintained over long time periods within each level of the language hierarchy. This phenomenon results in a superposition of successive representations that is supported by a hierarchy of dynamic neural codes. Overall, these findings provide a precise computational breakdown of the neural dynamics that coordinate the production of language in the human brain.

Introduction

The process by which the human brain translates abstract thoughts into the intricate sequences of motor actions required for language is a fundamental mystery in neuroscience. Studying natural language production is notoriously difficult because the act of speaking generates significant auditory feedback and facial muscle artifacts that corrupt neuroimaging signals. Consequently, much of our understanding comes from simplified tasks like single-word reading or covert speech. In this study, we overcome these limitations by investigating the neural activity of skilled typists. Typing provides a unique window into language production, as it removes speech-related artifacts while maintaining the complex, sequential nature of language generation.

Methods and Materials

We recorded magnetoencephalography (MEG) and electroencephalography (EEG) data from 35 healthy volunteers who were highly skilled typists. The experimental design utilized a "read-wait-type" paradigm where participants memorized short Spanish sentences (5-8 words long) and subsequently typed them on a custom, artifact-free keyboard.

Experimental paradigm and computational framework for decoding linguistic hierarchies.

To analyze the neural representations, we mapped the linguistic structure of the sentences onto four hierarchical levels using AI-based embeddings: context (GPT-2 hidden activations), words (SpaCy embeddings), syllables (FastText subword embeddings), and letters (One-Hot Encoding). We then employed linear ridge regression to decode these features from the brain signals and used temporal generalization analysis to characterize the stability and dynamics of the underlying neural codes.

Results

Our analysis revealed a clear top-down cascade in the neural preparation of language. Before a word is produced, the brain sequentially activates representations of context, then the specific word, its syllables, and finally the individual letters. The timing of these activations—specifically the "half-time" when decoding reaches 50% of its peak—perfectly correlated with the linguistic hierarchy.

Neural activity topographies and time series during the read and type phases.

Crucially, we found that neural representations are maintained for extended periods—often ten times longer than the physical motor action itself. For instance, while a keypress takes approximately 100ms, its neural representation is decodable for over a second. This leads to a significant temporal overlap, or superposition, where up to five successive letters can be decoded simultaneously from a single snapshot of brain activity.

To prevent interference between these overlapping representations, the brain employs a hierarchy of dynamic neural codes. Each representation moves through a high-dimensional neural subspace at a speed that correlates with its hierarchical level: high-level context shifts slowly as the sentence unfolds, while low-level motor targets (letters) move rapidly through their neural trajectories.

Discussion

These results provide a precise computational account of how the brain coordinates language production through a top-down hierarchy of representations. The discovery of dynamic neural codes explains how the brain manages the simultaneous preparation of multiple linguistic elements without cross-talk.

While typing provides a clean signal for neuroimaging, several questions remain. It is unclear whether the decoded "letter" representations reflect abstract graphemes or keyboard-specific motor plans. Furthermore, while MEG and EEG offer high temporal resolution, they lack the spatial precision to pinpoint the exact anatomical locations of these hierarchical computations. Future research should explore whether these findings generalize to natural speech and other forms of human communication.

Summary updated

Abstract

Humans effortlessly communicate their thoughts through intricate sequences of motor actions. Yet, the neural processes that coordinate language production remain largely unknown, in part because speech artifacts limit the use of neuroimaging. To elucidate the unfolding of language production in the brain, we investigate with magnetoencephalography (MEG) and electroencephalography (EEG) the neurophysiological activity of 35 skilled typists, while they typed sentences on a keyboard. This approach confirms the hierarchical predictions of linguistic theories: the neural activity preceding the production of each word is marked by the sequential rise and fall of context-, word-, syllable-, and letter-level representations. Remarkably, each of these neural representations is maintained over long time periods within each level of the language hierarchy. This phenomenon results in a superposition of successive representations that is supported by a hierarchy of dynamic neural codes. Overall, these findings provide a precise computational breakdown of the neural dynamics that coordinate the production of language in the human brain.
BibTeX
@misc{mingfang2025from,
  author = {Mingfang and Zhang and J P Lévy and Stéphane d’Ascoli and Jérémy Rapin and F.‐Xavier Alario and Pierre Bourdillon and Svetlana Pinet},
  title = {{From Thought to Action: How a Hierarchy of Neural Dynamics Supports Language Production}},
  year = {2025},
  eprint = {2502.07429},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.07429},
  howpublished = {arXiv},
}
Figure