BERT-based Sentiment Analysis for COVID-19 Tweets
An end-to-end deep learning system utilizing a fine-tuned BERT model (91.4% accuracy) to classify COVID-19 tweet sentiments, featuring a Flask REST API backend and a cross-platform Flutter client for real-time inference.
Context
Master's thesis project at Constantine 2 University – Abdelhamid Mehri. A related comparative study of machine learning, deep learning and BERT models on COVID-19 tweets is published in the ISPR 2025 proceedings (Springer).
Highlights
- Fine-tuned BERT-base (110M parameters) reaching 91.4% accuracy on three sentiment classes: positive, negative and neutral.
- Trained on 41,157 COVID-19 tweets from March–April 2020, with a full text-preprocessing pipeline and 60-token input sequences.
- Flask REST API for real-time inference, with health-check endpoints and CORS support for cross-platform clients.
- Flutter client for Android, iOS, web and desktop, with Material 3, dark mode and model-performance screens.
