A Compositional Neural Architecture for Language.

Area of Science:

• Neuroscience
• Computational Linguistics
• Cognitive Science

Background:

• Human language's hierarchy and compositionality are key features, yet current models lack neurobiological explanations for their origin.
• Existing formal and neurobiological models do not fully explain how the brain generates complex linguistic structures.

Purpose of the Study:

• To propose a neurobiological theory reconciling hierarchy and compositionality with cell assembly computation.
• To explain how the brain converts perceptual data into hierarchical linguistic structures with incremental meaning.
• To bridge insights from linguistics, psycholinguistics, and neuroscience.

Main Methods:

• Developed a model based on a multidimensional coordinate system using neurophysiological principles of sensory processing.
• Incorporated gain modulation and inhibition to tune neural trajectories representing linguistic states.
• Synthesized concepts from neural entrainment, systems neuroscience, category theory, and a symbolic-connectionist model using time and rhythm.

Main Results:

• Proposed a model where neural trajectories in a manifold encode sensory, motor, and linguistic states.
• Demonstrated how gain modulation infers latent structure and enables predictive information processing.
• Showcased a formal and mechanistic alignment between neural oscillations and linguistic structure building.

Conclusions:

• The proposed model offers a potential mechanism for the brain's hierarchical language processing.
• This framework integrates symbolic and connectionist approaches, utilizing time and rhythm for information structuring.
• The theory provides a unified perspective linking linguistic principles with neural computation mechanisms.