Narimatsu, Hiromi; Nomura, Keishi; Wei, Xijia; Berthouze, Nadia; Kumano, Shiro; (2025) Emotion in Art-Elicited Texts, Interpreted by GPT-4o: Self vs. Third-Person Perspectives. In: Proceedings of 13th International Conference on Affective Computing & Intelligent Interaction (ACII 2025). IEEE (In press).

Preview

Text ACII2025_Narimatsu_-3.pdf - Accepted Version Download (200kB) | Preview

Abstract

This study investigates whether GPT-4o can serve as a proxy in emotion research by estimating human emotional states from text. Unlike prior resources such as EmoBank, which include inferred writer labels and emotionally neutral texts, we collected a dataset in which participants viewed visual artworks, reported their own emotions using the valence–arousal–dominance (VAD) model, alongside natural language descriptions. GPT-4o was prompted to estimate emotional intensity from both writer (self) and reader (third-party) perspectives, and its predictions were compared to human self-reports and reader judgments. Using pairwise comparison tasks, we evaluated agreement based on forced-choice judgments. Results show that GPT-4o aligns closely with mean human readers for valence, moderately for arousal, and poorly for dominance. Notably, for arousal, the alignment between GPT-4o’s writer and reader perspectives varied depending on whether the texts shared the same emotional context—that is, whether they pertained to the same image. We also examined prompt design effects but found no benefit from intermediate reasoning for the pairwise comparison of the art-elicited texts. Our findings suggest that with appropriate task framing, LLMs like GPT-4o may support pilot studies in cognitive and affective research.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item