Localizing syntactic predictions using recurrent neural network grammars.

Area of Science:

• Neuroscience
• Computational Linguistics
• Cognitive Science

Background:

• Brain activity in perisylvian regions is influenced by the predictability of language.
• Computational parsing models offer new ways to understand linguistic representations.
• Neural network language models can quantify word unexpectedness.

Purpose of the Study:

• To investigate which linguistic representations guide brain activity during language comprehension.
• To evaluate computational parsing model outputs against fMRI data.
• To identify neural correlates of predictive processing in the brain's language network.

Main Methods:

• Utilized Recurrent Neural Network Grammars (RNNGs) to model linguistic derivations and extract complexity metrics.
• Analyzed functional magnetic resonance imaging (fMRI) data from participants listening to an audiobook.
• Correlated word surprisal (from LSTM models) and RNNG-derived metrics with brain activity patterns.

Main Results:

• Word surprisal, derived from word-sequence models, correlated with activity in temporal and frontal brain regions.
• RNNG-based hierarchy encoding was associated with activity in left posterior temporal areas.
• Derivational step count correlated with activity in the left temporal lobe and inferior frontal gyrus.

Conclusions:

• Linguistic predictability and hierarchical structure significantly modulate brain activity in the language network.
• Computational models like RNNGs provide valuable insights into the neural basis of language processing.
• These findings refine our understanding of the representations supporting predictive language comprehension.