The project explores the application of continual learning techniques to a T5 NLP model for news summarization, addressing the challenge of catastrophic forgetting, where a model loses accuracy on previously learned tasks when trained on new data. This is accomplished by incrementally (LoRA) fine-tuning the model first on the CNN/DailyMail dataset and subsequently on daily news articles sourced from NewsAPI. An interactive web interface developed using Flask enables real-time user interactions for custom news retrieval and summarization tasks. The project evaluates the model’s performance using ROUGE scores and employs advanced training methodologies such as Elastic Weight Consolidations and Deep Generative Replay to maintain effectiveness across datasets. Automation and cloud hosting on Google Cloud ensure efficient and scalable data processing, model training, and deployment, making the system robust and user-friendly for real-world applications.
