SimoGiuffrida/SentimentRL

Emotion Classification with BERT + RL Fine-tuning

This model combines BERT architecture with Reinforcement Learning (RL) for emotion classification. Initially fine-tuned on the dair-ai/emotion dataset (20k English sentences with 6 emotions), we then applied PPO reinforcement learning to optimize prediction behavior.

🔧 Training Approach

1. Supervised Phase:

   • Base BERT model fine-tuned with cross-entropy loss
   • Achieved strong baseline performance

2. RL Phase:

   • Implemented Actor-Critic architecture
   • Policy Gradient optimization with custom rewards
   • PPO clipping (ε=0.2) and entropy regularization
   • Custom reward function: +1.0 for correct, -0.1 for incorrect predictions

📊 Performance Comparison

Metric	Pre-RL	Post-RL	Δ
Accuracy	0.9205	0.931	+1.14%
F1-Score	0.9227	0.9298	+0.77%
Precision	0.9325	0.9305	-0.21%
Recall	0.9205	0.931	+1.14%

Key observation: RL fine-tuning provided modest but consistent improvements across most metrics, particularly in recall.

🚀 Usage

from transformers import pipeline

# Load from your repository
classifier = pipeline("text-classification", 
                     model="SimoGiuffrida/SentimentRL",
                     tokenizer="bert-base-uncased")

results = classifier("I'm thrilled about this new opportunity!")

💡 Key Features

• Hybrid training: Supervised + Reinforcement Learning
• Optimized for nuanced emotion detection
• Handles class imbalance (see confusion matrix in repo)

For full training details and analysis, visit the GitHub repository.