Transformer-based Model Captures Neural Representation Differences between Nouns and Verbs in Spoken Narratives
Yaoyao Wang, Jiaying Zhang, Siyi Tu, Cheng Luo · 2024
Nouns and verbs constitute the fundamental elements of human language systems. Abundant studies have demonstrated that nouns and verbs exhibit different representations in both biological brains and various advanced deep neural networks (DNNs). Here, we investigate whether humans and DNNs represent the word class difference in a comparable way under the context of passages. For humans, we used electroencephalography (EEG) recordings to analyze neural responses to nouns and verbs when participants naturally listened to spoken narratives. For DNNs, we analyzed word embeddings from a transformer-based model, i.e., BERT, that reach human-level performance when trained to perform natural language processing (NLP) tasks. A significant word class difference was observed in both the EEG responses and model representations. Further analyses revealed that the model representations could predict the EEG response difference between nouns and verbs. All these results suggest that DNNs can naturally evolve human-like language representations, and their hidden layers' embeddings capture the word class difference in neural representations of human brains.