Multi-Class Sentiment Analysis with BERT Fine-tuning

Portfolio Note: Portfolio recreation of NLP pipeline built at Omfys Technologies.

🎯 Overview

Fine-tuned BERT on 100K+ reviews achieving 93% F1-score with PyTorch, ONNX optimization, and Streamlit dashboard achieving 87% satisfaction improvement and 22% reduction in negative reviews.

📊 Key Metrics

• Dataset: 100K+ customer reviews
• F1-Score: 93%
• Classes: 5 (Very Negative to Very Positive)
• Inference: <100ms latency
• Business Impact: 87% satisfaction improvement

🛠️ Tech Stack

• DL Framework: PyTorch, Transformers (Hugging Face)
• Model: BERT-base-uncased
• Optimization: ONNX Runtime (3x speedup)
• Dashboard: Streamlit
• API: FastAPI
• Topic Modeling: Gensim (LDA)
• Cloud: OCI Data Science (A10 GPUs)

⚡ Key Features

1. BERT Fine-Tuning

• Transfer learning from BERT-base-uncased
• Mixed-precision training (FP16) on A10 GPUs
• Gradient accumulation
• Weighted loss for class imbalance
• Training time reduced by 50%

2. Data Augmentation

• Back-translation
• Synonym replacement
• Random insertion/deletion
• 40% increase in training data

3. ONNX Optimization

• Model quantization
• 3x inference speedup
• Reduced model size by 4x
• Deployment-ready format

4. Interactive Dashboard

• Real-time prediction (<100ms)
• Confidence scores
• Attention visualization with BertViz
• Batch processing (10K reviews/hour)
• LDA topic modeling

📁 Project Structure

sentiment-analysis-bert/
├── src/
│   ├── training/            # Fine-tuning scripts
│   ├── inference/           # ONNX inference
│   ├── dashboard/           # Streamlit app
│   └── preprocessing/       # Text processing
├── notebooks/               # Jupyter notebooks
├── Dockerfile
└── README.md

🚀 Getting Started

git clone https://github.com/Amanroy666/sentiment-analysis-bert.git
cd sentiment-analysis-bert
pip install -r requirements.txt

# Run training
python src/training/train_bert.py

# Start dashboard
streamlit run src/dashboard/streamlit_app.py

📈 Model Performance

Class	Precision	Recall	F1-Score
Very Negative	91%	89%	90%
Negative	92%	91%	91.5%
Neutral	94%	95%	94.5%
Positive	93%	94%	93.5%
Very Positive	95%	96%	95.5%
Weighted Avg	93%	93%	93%

👤 Author

Aman Roy - Data Engineer at Omfys Technologies
📧 contactaman000@gmail.com | 💼 LinkedIn