savalera/isear-emotion-classifier

ISEAR Emotion Classifier

A BERT-based model that classifies text into seven basic emotions, fine-tuned on the ISEAR dataset.

Model Overview

This model predicts one of seven emotions from textual descriptions:

Label	Emotion	Definition
0	Joy	Happiness, pleasure, or satisfaction
1	Fear	Anxiety about threat or danger
2	Anger	Strong displeasure or antagonism
3	Sadness	Feeling of loss or disadvantage
4	Disgust	Aversion or revulsion
5	Shame	Embarrassment or humiliation
6	Guilt	Feeling responsible for wrongdoing

Performance

Validation metrics:

• Accuracy: 70.98%

Top performing emotions: Joy (92% F1), Fear (78% F1)
Most challenging: Shame (59% F1), Anger (60% F1)

Usage

from transformers import pipeline

classifier = pipeline("text-classification", model="savalera/isear-emotion-classifier")
result = classifier("I received a surprise visit from my best friend who I hadn't seen in years.")
print(result) # [{'label': 'joy', 'score': 0.9685400724411011}]

Training

• Base model: bert-base-uncased
• Dataset: ISEAR (International Survey on Emotion Antecedents and Reactions)

Limitations

The model may struggle to distinguish between closely related emotions (especially shame/guilt), which reflects inherent challenges in emotion classification that even humans face.

Ethical Usage Guidelines

This model analyzes textual descriptions to classify emotions. While it has many potential applications, users should consider these ethical guidelines:

Recommended Applications

• Academic research on emotional expression and recognition
• Healthcare applications for mental health support under appropriate supervision
• Enhancing communication effectiveness through better emotional awareness
• Supporting team coaching programs that foster psychological safety, enchance communication effectiventess and collaboration

Usage Considerations

• Some jurisdictions have regulations around emotion recognition technology, particularly in workplace and educational settings
• Privacy and consent are important when analyzing personal emotional expressions
• The model's predictions should be treated as probabilistic rather than definitive assessments of emotional states

Users should be aware of restrictions on emotion recognition systems and ensure any applications developed using this model comply with regulatory requirements. This model should not be used for commercial emotion categorization systems to infer emotions of a natural person.

Citation

If you use this model in your research, please cite it directly:

@misc{savalera2025isear,
  title={ISEAR Emotion Classifier},
  author={Savalera},
  year={2025},
  howpublished={\url{https://huggingface.co/savalera/isear-emotion-classifier}}
}

This model was trained on the ISEAR dataset, which was derived from research by Scherer & Wallbott (1994).

License

Apache License 2.0