Yale researchers propose emotion intensity scoring for text analysis
New Yale framework scores text emotions 0-100, outperforming sentiment classification
A team from Yale University (led by Francesco A. Fabozzi, Dasol Kim, and William N. Goetzmann) has published a paper proposing a shift from sentiment classification to emotion intensity evaluation in text analysis. The researchers argue that traditional NLP approaches fail to capture the nuanced degree of emotional content critical for domains like finance, where subtle tonal variations can significantly impact decision-making.
The team constructed a dataset of emotional intensity scores and fine-tuned open-weight generative language models (likely including models like Llama 3 or Mistral) to output continuous values between 0-100. Their framework not only outperformed standard classification baselines but also revealed surprising generalization capabilities, including transfer effects to related constructs like sentiment and arousal. The paper, submitted to arXiv on May 15, 2026, positions emotion intensity evaluation as a more expressive alternative to discrete classification methods.
- Proposes continuous emotion intensity scoring (0-100 scale) instead of binary sentiment classification
- Uses fine-tuned open-weight generative models with new emotional intensity dataset
- Outperforms baselines and shows transfer effects to sentiment/arousal analysis
Why It Matters
Enables more precise financial and psychological text analysis through granular emotion scoring