[4L4-GS-1a] Fundamental AI, theory algorithm

2:15 PM - 2:30 PM JST(5:15 AM - 5:30 AM UTC)

[4L4-GS-1a-04] Neural Correlates of Emotion Labels Derived from Natural Language Processing in fMRI Functional Networks

〇Midori Yamamoto¹, Yuko Ishimaru¹, Shunsuke Nakashima¹, Atsushi Kawaguchi¹ (1. SAGA University)

Keywords:

Sentiment Analysis,Neuroimaging,Functional Connectivity,Evaluation of AI models,AI interpretability

This study investigated the neurobiological validity of AI-based sentiment analysis using natural language processing (NLP) by examining the relationship between estimated emotions and functional brain connectivity measured with functional magnetic resonance imaging (fMRI).fMRI data from 26 participants in the ALICE dataset were acquired during listening to an audio narration of Chapter 1 of Alice’s Adventures in Wonderland. Narrative texts were analyzed using three AI sentiment models. A sliding-window approach was applied to incorporate contextual information, generating sentence-level time series of positive, negative, and neutral emotion labels (present/absent).During emotion-present periods, seed-to-voxel analyses were performed using core regions of emotion-related networks as seeds. Functional connectivity was statistically evaluated and compared with established brain networks, and differences among AI models were examined.The results showed increased connectivity in the Default Mode and Frontoparietal Networks for positive emotions, and in the Salience Network for negative emotions. The estimated emotions were consistent with known neural networks, and significant differences were observed among the models.These findings indicate that neuroimaging-based analyses enable validation and comparison of AI sentiment models.