Sentiment Analysis Using Fine-Tuned BERT

By Ansh Sharma

---

Deployed Link--

• Google Colab -- https://colab.research.google.com/drive/1xtIr_MBp--2J8Qhasv7I1Bv6KYLlSHJv?usp=sharing

• Wandb AI -- https://wandb.ai/anshsharma210504

• 

Welcome to my sentiment analysis project! This repository showcases the fine-tuning of a BERT model using Hugging Face, deployment on WandB AI and Heroku via FastAPI, and the complete workflow of testing, evaluating, and productizing the model for real-world applications.

---

Project Overview

The project aims to classify sentiment in student reviews about campus events and amenities into Positive, Negative, or Neutral categories. This end-to-end implementation covers:

• Data Preparation: Creating a custom dataset with ReviewText and SentimentLabels.
• Fine-Tuning: Leveraging the pre-trained BERT model to adapt to the dataset.
• Evaluation: Measuring performance with accuracy, F1-score, and confusion matrix.
• Visualization: Showcasing sentiment distributions and insights via Matplotlib.
• Productization: Deploying the model for public use via FastAPI on both local and cloud servers.

---

Key Features

1. Fine-Tuned BERT Model

   • Adapted pre-trained BERT using Hugging Face for sentiment classification.
   • Optimized with AdamW optimizer and evaluated on metrics like accuracy and F1-score.

2. Deployment

   • FastAPI: Designed an intuitive REST API for real-time sentiment predictions.
   • Hosted on Heroku and tested locally for seamless integration.

3. Monitoring & Logging

   • Tracked model performance and fine-tuning experiments on WandB AI.

4. Evaluation Metrics

   • Performance visualized with confusion matrices and sentiment distribution graphs.

---

Technologies Used

• Programming: Python
• Frameworks: Hugging Face Transformers, FastAPI
• Deployment: Heroku, Local Server
• Monitoring: WandB AI
• Visualization: Matplotlib, Seaborn

---

Project Workflow

1. Dataset Creation

• Custom dataset with ReviewText and SentimentLabels (Positive, Negative, Neutral).
• Preprocessing: Text cleaning, tokenization, and encoding of labels.

2. Model Fine-Tuning

• Base Model: BERT (bert-base-uncased).

• Added a classification head for sentiment analysis.

• Optimizer: AdamW with learning rate scheduling.

• Metrics: Accuracy, F1-score, and confusion matrix.

  

3. Evaluation

• Tested on unseen data with detailed analysis of predictions.
• Plotted sentiment distributions for better insights.

4. Productization

• Built a REST API with FastAPI:
  • Endpoint: /predict
  • Input: ReviewText
  • Output: Predicted sentiment and confidence score.
• Deployed on Heroku and tested via a local server.

---

Getting Started

Structure of the Repository

├── app.py # FastAPI application code
├── requirements.txt # Python dependencies
├── data/ # Dataset folder
├── models/ # Saved fine-tuned BERT model
├── visualizations/ # Evaluation metrics and graphs
└── README.md # Project documentation