Sentiment Analysis of YouTube Comments with BERT

Overview

This project fine-tunes a BERT model to classify YouTube comments into sentiment categories (positive, negative, neutral). The model achieves a macro F1-score of 0.82 on an imbalanced dataset of 10,000+ comments. The trained model is deployed as a containerized API for easy integration into applications.

• Dataset: Kaggle YouTube Comments Dataset
• Hugging Face Model: bert-finetuned-youtube_sentiment_analysis
• Docker Image: rahulk98/youtubecomment-sentiment-predictor

Table of Contents

• Project Structure
• Model Performance
• API Usage
• Docker Setup
• Local Development
• Example Response
• How It Works
• Training Process

Project Structure

├── Dockerfile                # Docker configuration for containerizing the API
├── README.md                 # Project documentation
├── images/                   # Screenshots and images for documentation
│   └── postman_response.jpg  # Example API response from Postman
├── notebooks/                # Jupyter notebooks for model training
│   ├── main.ipynb            # Training notebook for BERT fine-tuning
│   └── requirements.txt      # Python dependencies for training
└── src/                      # Source code for the API
    ├── app.py                # Flask API entry point
    ├── predict.py            # Prediction module using the trained model
    └── requirements.txt      # Python dependencies for the API

Model Performance

The model achieves excellent performance on the YouTube comments dataset:

F1 score (macro): 0.8182257811696632

Confusion matrix:

[[ 161   59   14]
 [  16  375   71]
 [   8   85 1047]]

The model can classify comments into three sentiment categories:

• Neutral
• Positive
• Negative

API Usage

The API is accessible via HTTP POST requests to the /predict endpoint.

Request Format

• URL: http://<host>:5000/predict
• Method: POST
• Headers: Content-Type: application/json
• Body:

  {
    "text": "Your YouTube comment text here"
  }

Response Format

{
  "sentiment": "positive"
}

The response will contain the predicted sentiment, which can be one of:

• "positive"
• "negative"
• "neutral"

Docker Setup

The API is containerized and available as a Docker image for easy deployment.

Pull the Docker Image

docker pull rahulk98/youtubecomment-sentiment-predictor

Run the Docker Container

docker run -p 5000:5000 rahulk98/youtubecomment-sentiment-predictor

This will start the API service on port 5000.

Local Development

If you want to run the API locally without Docker:

1. Clone the repository:

   git clone https://github.com/rahulk98/Sentiment-Analysis-with-BERT-Model.git
   cd Sentiment-Analysis-with-BERT-Model

2. Install dependencies:

   pip install -r src/requirements.txt

3. Run the API:

   cd src
   python app.py

Example Response

Below is an example of the API response from Postman:

How It Works

The API uses a fine-tuned BERT model to predict the sentiment of YouTube comments. The model was trained on a dataset of over 10,000 YouTube comments with manually labeled sentiments.

The prediction process:

1. The input text is tokenized using BERT tokenizer
2. The tokenized text is passed through the fine-tuned BERT model
3. The model outputs probabilities for each sentiment class
4. The class with the highest probability is returned as the prediction

Training Process

The model was fine-tuned using:

• Base model: bert-base-uncased
• Custom weighted loss function to handle class imbalance
• Learning rate: 5e-5
• Training epochs: 5
• Batch size: 16

For more details on the training process, check the notebooks/main.ipynb file.