gabe-zhang/twitter-emotion-bert

Twitter Emotion BERT

A fine-tuned bert-base-cased model for classifying emotions in text into six categories: sadness, joy, love, anger, fear, and surprise.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="gabe-zhang/twitter-emotion-bert",
)

classifier("I am so happy today!")
# [{'label': 'joy', 'score': 0.95}]

classifier("This makes me really angry")
# [{'label': 'anger', 'score': 0.89}]

Labels

ID	Emotion
0	sadness
1	joy
2	love
3	anger
4	fear
5	surprise

Performance

Overall: Accuracy 0.9385 | Weighted F1 0.9399

Emotion	Precision	Recall	F1-Score
Sadness	0.94	0.93	0.93
Joy	0.95	0.96	0.95
Love	0.86	0.89	0.87
Anger	0.93	0.90	0.91
Fear	0.91	0.89	0.90
Surprise	0.85	0.82	0.83

Training

• Base model: bert-base-cased
• Dataset: dair-ai/emotion (~20,000 English Twitter messages)
• Optimizer: AdamW, learning rate 1e-5
• Dropout: 0.5
• Max sequence length: 512
• Early stopping: patience 5
• Class balancing: Random oversampling for minority classes
• Framework: PyTorch Lightning

Limitations

• Trained on English Twitter text only; may not generalize well to other languages or formal writing styles.
• "Love" and "surprise" have lower F1 scores due to smaller representation in the training data and overlap with other emotions.
• Short, informal text with slang or emojis may affect predictions.

Source Code

Full training code, notebooks, and baseline comparisons: github.com/gabe-zhang/twitter-emotion-classification