|
--- |
|
language: en |
|
tags: |
|
- emotion-classification |
|
- text-classification |
|
- distilbert |
|
datasets: |
|
- dair-ai/emotion |
|
metrics: |
|
- accuracy |
|
--- |
|
|
|
# Emotion Classification Model |
|
## Model Description |
|
This model fine-tunes DistilBERT for a multi-class emotion classification task. The dataset that is used is dair-ai/emotion containing six emotion classes: sadness, joy, love, anger, fear and suprise |
|
## Training and Evaluation |
|
- Training Dataset: dair-ai/emotion (16,000 examples) |
|
- Validation Dataset: dair-ai/emotion (2,000 examples) |
|
- Test Dataset: dair-ai/emotion (2,000 examples) |
|
- Validation Accuracy: 96.6 % |
|
- Test Accuracy: 93.2 % |
|
- Training Time: 18,078.18s (5h 1m 18s) |
|
|
|
## Hyperparameters |
|
- Learning Rate: 5e-5 |
|
- Batch Size: 16 |
|
- Epochs: 4 |
|
- Weight Decay: 0.01 |
|
|
|
## Usage |
|
```python |
|
from transformers import pipeline |
|
classifier = pipeline("text-classification", |
|
model="your-username/emotion-classification-model") |
|
text = "I’m so happy today!" |
|
result = classifier(text) |
|
print(result) |
|
# Output: [{'label': 'LABEL_1', 'score': 0.9996680021286011}] ; which corresponds to 'joy' with a 99.9% accuracy |
|
``` |
|
|
|
## Limitations |
|
|
|
- Performance on Non-English Texts: The model works best with English texts. It may not handle other languages or regional dialects. |
|
- Bias From Training Data: Reflects patterns from the training data, which may not cover all cases it could see. May lead to unfair predictns |
|
- Issues with Sarcasm and Complex Language: Model might misinterpret sarcasm or subtle language, ex, "Great product... not!" |
|
- No Confidence Scores: Model does not judge it's own output, so need some form of double check system, ie. yourself |
|
|
|
## License |
|
The MIT License (MIT) |
|
|
