Sentiment Analysis Model for Product Reviews

Project Overview

This project is a production-ready DistilBERT Sentiment Analysis model for product reviews desgined as a low cost market research tool with the nuiance of an actual market researcher, it was trained with a dataset of over 4 million Amazon product reviews to work for a wide variety of products in a highly scalable manner. This project comes with designs for a serverless live demo in AWS.

Key Features

• Pre-trained DistilBERT model for sentiment analysis on product reviews
• Designed for low compute costs to maximize scalability
• Conversion from PyTorch to TensorFlow
• Data processing and tokenization pipeline
• Comprehensive testing of both original and converted models
• Designs of an AWS serverless architecture for a live demo with future deployment(TBD)

Project Structure

1. 1-Data processing: Script for data preparation
2. 2-CSV files(Production test): CSV file used for production testing(Full dataset https://www.kaggle.com/datasets/bittlingmayer/amazonreviews)
3. 3-Tokenization and Training: Tokenization process and initial model training
4. 4-distilbert_sentiment_model: Original PyTorch DistilBERT model
5. 5-Testing model: Scripts for testing the PyTorch model, both a write in and a CSV input script
6. 6-PyTorch2TensorFlow: Conversion script from PyTorch to TensorFlow
7. 7-tensorflow_model: Converted TensorFlow model
8. 8-Testing converted model: Scripts for testing the TensorFlow model, includes a write in one, and 2 CSV input ones, one for testing the model, one for production usage

Documentation

• MODEL.md: https://github.com/DimitriVavoulisPortfolio/aws-serverless-nlp-sentiment-4M-product-reviews/blob/main/MODEL.md
• PROCESS.md: https://github.com/DimitriVavoulisPortfolio/aws-serverless-nlp-sentiment-4M-product-reviews/blob/main/PROCESS.md
• AWS-IMPLEMENTATION-AND-COST-REPORT.md: https://github.com/DimitriVavoulisPortfolio/aws-serverless-nlp-sentiment-4M-product-reviews/blob/main/AWS-IMPLEMENTATION-WITH-COST-REPORT.md

Model Performance

• 4M-dataset-sentiment_distribution.png: Visualization of sentiment distribution in the dataset
• 400K-dataset-testing.png: Results of model testing on a 400K sample dataset
• Accuracy: 0.9973
• Precision: 0.9968
• Recall: 0.9979
• F1: 0.9973

Quick Start Guide

1. Clone the repository:

   git clone https://github.com/DimitriVavoulisPortfolio/aws-serverless-nlp-sentiment-4M-product-reviews.git
   cd aws-serverless-nlp-sentiment-4M-product-reviews
   

2. Install dependencies:

   pip install tensorflow transformers torch pandas numpy scikit-learn onnx onnx-tf 
   

3. To test the PyTorch model:

   python 5-Testing model\sentiment-analysis-testing-script-write-in.py
   

4. To test the TensorFlow model:

   python 8-Testing converted model\tensorflow-model-test-script.py
   

DISCLAIMER: The paths for the models have to be put in by the user

Future Work

• Implement AWS deployment
• Create API for real-time sentiment prediction
• Optimize model performance and size

License

This project is licensed under the Apache-2.0 license - see the LICENSE file for details.

Contact

For any questions or feedback, please open an issue in this repository or contact Dimitri Vavoulis.