Modeling Structure-Building in the Brain With CCG Parsing and Large Language Models

Area of Science:

• Neuroscience
• Computational Linguistics
• Cognitive Science

Background:

• Natural language processing and machine learning tools are increasingly used to model language comprehension.
• Context-free grammars (CFGs) have been predominantly used for syntactic structure modeling, but lack the expressiveness for human languages.
• Combinatory categorial grammars (CCGs) offer greater expressiveness and directly compositional models for incremental interpretation.

Purpose of the Study:

• To evaluate if a more expressive CCG provides a better model of human neural signals than a CFG during naturalistic language comprehension.
• To compare different CCG variants in their ability to model neural data.
• To differentiate the neural correlates of grammatical structure-building from word predictability.

Main Methods:

• Functional magnetic resonance imaging (fMRI) was used to collect neural signals from participants listening to an audiobook.
• The study compared the fit of CCG-derived measures against CFG-derived measures and a baseline of next-word predictability from a transformer model.
• CCG variants differing in adjunct handling were tested.

Main Results:

• CCG-derived measures showed a superior fit to neural signals compared to CFG-derived measures, particularly in the left posterior temporal lobe.
• Neural effects related to CCG structure-building were spatially distinct from bilateral superior temporal effects associated with predictability.
• The findings suggest that grammatical structure-building has separable neural effects from predictability during naturalistic listening.

Conclusions:

• More expressive grammars like CCGs offer a better model for the neural basis of language comprehension than less expressive CFGs.
• Grammatical structure-building and word predictability engage distinct neural mechanisms.
• The study highlights the importance of grammar's expressive power in understanding the neural processing of language.