text-emotion-distilroberta

DistilRoBERTa for Emotion Classification

Model Description

This model is a fine-tuned version of distilroberta-base for multi-class emotion classification. It can classify text into 7 emotion categories: Neutral, Happiness, Anger, Sadness, Surprise, Disgust, and Fear.

Model Details

• Model Type: Text Classification
• Base Model: distilroberta-base
• Language: English
• Task: Multi-class Emotion Classification
• Classes: ['Anger', 'Disgust', 'Fear', 'Happiness', 'Neutral', 'Sadness', 'Surprise']

Dataset

Overview

The model was trained on a comprehensive emotion recognition dataset created by aggregating and preprocessing three publicly available datasets: DailyDialog, ISEAR, and MELD. The final dataset contains 68,301 samples across 7 emotion classes.

Source Datasets

• DailyDialog: ~13,000 conversational dialogs with emotion labels
• ISEAR: ~7,600 self-reported emotional experiences
• MELD: ~14,000 multimedia dialog utterances with emotion labels

Preprocessing Pipeline

The dataset underwent extensive preprocessing to ensure quality and balance:

1. Data Cleaning & Unification

• Mojibake correction: Fixed encoding artifacts (e.g., ’ → ')
• Text normalization: Standardized quotes, removed escaped characters
• Social media cleanup: Removed mentions, hashtags, and emojis from ISEAR
• Emotion mapping: Unified emotion labels across all three datasets
• Result: 124,723 initial samples

2. Deduplication

• Removed 3 samples with missing text
• Eliminated 18,624 duplicate samples (14.93%)
• Result: 106,096 unique samples

3. Class Balance Correction

Original distribution (severely imbalanced):

• Neutral: 76,762 (72.35%)
• Happiness: 14,337 (13.51%)
• Anger: 4,112 (3.88%)
• Sadness: 3,580 (3.37%)
• Surprise: 2,872 (2.71%)
• Disgust: 2,306 (2.17%)
• Fear: 2,127 (2.00%)

Balancing strategy:

• Undersampled Neutral class to 19,190 samples
• Preserved minority classes for natural distribution
• Result: 37,771 balanced samples

4. Data Augmentation

Applied targeted augmentation techniques to enhance diversity and address minority class underrepresentation:

Techniques used:

• Synonym Replacement: Replaced words with synonyms while preserving emotion keywords
• Character Noise: Added random characters to simulate typos
• Emotion Intensifiers: Added emotion-specific adjectives
• Paraphrasing: Used T5-small model for sentence rephrasing

Augmentation results:

• Fear: 2,127 → 5,000 samples
• Disgust: 2,306 → 5,001 samples
• Surprise: 2,872 → 5,000 samples
• Sadness: 3,580 → 5,002 samples
• Anger: 4,112 → 5,001 samples
• Happiness: 3,584 → 5,001 samples
• Neutral: 19,190 → 38,296 samples

Total augmented samples: 30,530 (44.69% of final dataset)

Final Dataset Characteristics

Size: 68,301 samples total

• Original samples: 37,771 (55.31%)
• Augmented samples: 30,530 (44.69%)

Class Distribution:

Text Statistics:

• Average text length: 74.4 characters
• Average word count: 15.1 words
• Text length range: 1-1,158 characters

Dataset Splits:

• Training: 26,439 samples (70%)
• Validation: 5,666 samples (15%)
• Test: 5,666 samples (15%)

Splits were created using stratified sampling to maintain proportional class representation across all sets.

Performance

Overall Metrics

• Accuracy: 0.8103
• F1-Score (Macro): 0.7286
• F1-Score (Weighted): 0.8096

Per-Emotion Performance

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("jihedjabnoun/text-emotion-distilroberta")
model = AutoModelForSequenceClassification.from_pretrained("jihedjabnoun/text-emotion-distilroberta")

# Prepare text
text = "I'm feeling great today!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)

# Get prediction
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_class = torch.argmax(predictions, dim=-1)

# Map to label
emotions = {0: 'Anger', 1: 'Disgust', 2: 'Fear', 3: 'Happiness', 4: 'Neutral', 5: 'Sadness', 6: 'Surprise'}
predicted_emotion = emotions[predicted_class.item()]
print(f"Predicted emotion: {predicted_emotion}")

Training Details

• Training Framework: Transformers
• Optimizer: AdamW
• Learning Rate: 2e-5
• Batch Size: 16
• Max Sequence Length: 128
• Epochs: 3
• Warmup Steps: 500

Limitations and Considerations

• The dataset maintains some class imbalance with Neutral being dominant (56.06%)
• Augmented samples comprise 44.69% of the dataset, which may introduce synthetic artifacts
• Performance varies across emotion classes, with Fear and Neutral showing highest F1-scores
• The model is optimized for English text and may not generalize to other languages

🤝 Contact & Support

• Model Repository: jihedjabnoun/text-emotion-distilroberta
• Issues: Please report bugs or issues in the model repository
• Improvements: Suggestions for model improvements are welcome
• Mail: [email protected]

Citation

@misc{text-emotion-distilroberta,
    title={DistilRoBERTa for Emotion Classification},
    author={Jihed Jabnoun},
    year={2025},
    url={https://huggingface.co/jihedjabnoun/text-emotion-distilroberta}
}