Mapping 99 emotion terms with GPT4 prompting reveals nuanced semantic conceptual structure

Area of Science:

• Cognitive Psychology
• Computational Linguistics
• Affective Science

Background:

• Dimensional theories organize emotions along Pleasure, Arousal, and Dominance.
• Current computational methods like word embeddings require significant technical expertise.
• Accessible methods are needed to analyze complex emotion structures.

Purpose of the Study:

• To evaluate GPT-4 prompting as an accessible alternative to word embeddings for capturing emotion's dimensional structure.
• To compare GPT-4 prompting and word embeddings against human spatial reasoning.
• To analyze extensive emotion vocabularies beyond human cognitive limits.

Main Methods:

• Study 1: Validated GPT-4 prompting and word embeddings using six basic emotions against human judgments.
• Study 2: Extended analysis to 99 emotion terms, comparing both computational methods.
• Utilized human spatial reasoning as a benchmark for semantic structure analysis.

Main Results:

• GPT-4 prompting showed strong convergence with human judgments, unlike word embeddings.
• Both methods identified two clusters based on Pleasure and Dominance, with clearer separation by GPT-4.
• Arousal emerged as a secondary dimension at larger vocabulary scales, not a primary cluster-defining one.

Conclusions:

• GPT-4 prompting is a promising, accessible tool for investigating the semantic structure of emotion.
• This method effectively uncovers patterns in broad emotion lexicons, surpassing limitations of traditional approaches.
• Findings advance understanding of emotion representation and computational analysis in psychology.