Finding structure during incremental speech comprehension.

Area of Science:

• Neuroscience
• Computational Linguistics
• Cognitive Science

Background:

• Human speech comprehension relies on incrementally integrating words into structured interpretations.
• This process is governed by complex linguistic and contextual probabilistic constraints.
• Understanding the neural basis of this dynamic integration is crucial.

Purpose of the Study:

• To investigate the neural substrates of incremental speech comprehension.
• To map the spatiotemporal dynamics of integrating linguistic and contextual information in the brain.
• To validate artificial neural networks as models for cognitive processes.

Main Methods:

• Extracted word-by-word sentential structure measures from BERT (a deep language model).
• Employed representational similarity analysis comparing BERT measures with brain activity.
• Recorded spatiotemporally resolved brain activity using electro-/magnetoencephalography (EEG/MEG).

Main Results:

• Identified specific brain regions and timing for the emergence of coherent interpretations.
• Demonstrated that coherent interpretation involves evaluating and integrating multifaceted constraints.
• Revealed engagement of bilateral brain regions beyond the classical fronto-temporal language system.

Conclusions:

• Neural processes for speech comprehension dynamically integrate multiple constraints.
• Artificial neural networks like BERT can model the neural dynamics of cognitive processes.
• Speech comprehension involves a widespread neural network, not limited to traditional language areas.