Reproduce “CHEER-Ekman: Fine-grained Embodied Emotion Classification” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Phan Anh Duong, Cat Luong, Divyesh Bommana, Tianyu Jiang

Researchers at the University of Cincinnati extended the CHEER dataset into CHEER-Ekman, enabling fine-grained classification of embodied emotions into Ekman's six basic categories. Using Large Language Models with Best-Worst Scaling, the work achieved an F1-score of 50.6 for classification, outperforming a fine-tuned BERT model, while also demonstrating significant improvements in binary embodied emotion detection via prompt simplification and Chain-of-Thought techniques.

Key insights

  • Prompt engineering, particularly simplification and CoT, dramatically improves LLM performance in embodied emotion detection, enabling smaller models to achieve competitive results.
  • Automatic Best-Worst Scaling allows LLMs to perform complex, training-free fine-grained emotion classification, surpassing traditional supervised approaches like fine-tuned BERT.
  • LLMs face challenges in distinguishing literal, functional, and metaphorical uses of body parts in emotional contexts, often prioritizing surface-level textual cues over nuanced embodied signals.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.