Transformer-based Model Captures Neural Representation Differences between Nouns and Verbs in Spoken Narratives

Area of Science:

• Cognitive Neuroscience
• Computational Linguistics
• Artificial Intelligence

Background:

• Nouns and verbs are core language components with distinct representations in human brains and deep neural networks (DNNs).
• Previous research highlights differences in how biological brains and artificial systems process word classes.

Purpose of the Study:

• To investigate if humans and DNNs represent word class differences comparably within passage contexts.
• To compare neural responses to nouns and verbs in humans with representations in DNNs.

Main Methods:

• Utilized electroencephalography (EEG) to record human neural responses to nouns and verbs during narrative listening.
• Analyzed word embeddings from BERT, a transformer-based DNN, for linguistic representations.
• Correlated DNN representations with human EEG responses.

Main Results:

• Observed significant differences in EEG responses between nouns and verbs in humans.
• Identified comparable word class differences in BERT model representations.
• Demonstrated that DNN representations could predict human EEG response differences.

Conclusions:

• DNNs can develop human-like language representations.
• The hidden layer embeddings of DNNs capture word class differences found in human brain activity.
• Suggests a convergence in how biological and artificial systems process fundamental language structures.