README.md

metadata

license: apache-2.0
title: CustomerFeedbackClassification
sdk: gradio
sdk_version: 5.49.1

BERT Sentiment Classification

A fine-tuned BERT model for customer feedback sentiment analysis, deployed as a Gradio web application.

🚀 Features

• Real-time sentiment analysis using fine-tuned BERT model
• Interactive web interface built with Gradio
• Confidence score visualization with bar charts
• Support for 3 sentiment classes: Positive 😊, Negative 😞, Neutral 😐
• Professional UI with examples and detailed results
• Model flexibility - works with fine-tuned or base BERT models

🧠 Model Details

• Base Model: bert-base-uncased (Google's BERT)
• Task: Multi-class sentiment classification
• Classes: 3 (positive, negative, neutral)
• Training: Fine-tuned on customer feedback dataset
• Architecture: BERT encoder + classification head
• Performance: ~85-90% accuracy on validation data

🔧 Technical Specifications

• Framework: PyTorch + Transformers
• Interface: Gradio
• Model Size: ~109M parameters
• Max Sequence Length: 128 tokens
• Batch Processing: Optimized for real-time inference

📦 Dependencies

The application requires the following Python packages:

torch>=1.9.0
transformers>=4.20.0
gradio>=3.40.0
pandas>=1.3.0
numpy>=1.21.0
scikit-learn>=1.0.0

🚀 Usage

1. Enter text in the input box
2. Click "Analyze Sentiment" to get predictions
3. View results including:
   • Predicted sentiment with emoji
   • Confidence percentage
   • Detailed probability breakdown
   • Visual confidence chart

💡 Example Inputs

Try these sample texts to see the model in action:

• "This product exceeded all my expectations! Outstanding quality."
• "I'm completely disappointed with this purchase."
• "The product is decent. It works as described but nothing extraordinary."
• "Best purchase I've made this year! Highly recommend."
• "The product I received was damaged. Unacceptable."

🔍 How It Works

1. Text Processing: Input text is tokenized using BERT tokenizer
2. Encoding: BERT encoder processes tokens with self-attention mechanisms
3. Classification: A classification head outputs probability scores for each sentiment class
4. Prediction: The class with the highest probability is selected as the final prediction

🏗️ Architecture

Input Text → BERT Tokenizer → BERT Encoder → Classification Head → Softmax → Prediction

📊 Model Performance

• Accuracy: ~85-90% on validation dataset
• Response Time: <2 seconds per prediction
• Confidence Scores: Clear differentiation between sentiment classes
• Robustness: Handles various text lengths and styles

🌐 Deployment

This application is designed for deployment on:

• Hugging Face Spaces (recommended - free & permanent)
• Google Colab (for development and testing)
• Local environments (with proper dependencies)
• Cloud platforms (AWS, GCP, Azure)

🔧 Model Files

The application supports multiple model formats:

• sentiment_pipeline.pkl - Complete pipeline with model and tokenizer
• bert_sentiment_model/ - HuggingFace format directory
• Fallback to base BERT model if no fine-tuned model is available

📝 License

This project is open source and available under the Apache 2.0 License.

🤝 Contributing

Contributions, issues, and feature requests are welcome!

📧 Contact

For questions or support, please open an issue in the repository.

---

Built with ❤️ using BERT, PyTorch, and Gradio