Enhancing Vaccine Reaction Detection from Social Media Using Optimized Transformer Fine-Tuning

Abstract

This paper describes the system developed by the UoT team for Task 6 of the 10th Social Media Mining for Health (#SMM4H) Shared Tasks, which focused on detecting personally experienced vaccine reactions in social media posts. We fine-tuned the CardiffNLP Twitter-RoBERTa-Large model using optimized training settings and data preprocessing strategies. Our best submission achieved an F1-score of 0.945 on the test set, outperforming the average system and nearing the benchmark score of 0.946. We describe our training pipeline, evaluation metrics, and results, and compare our system with both large language models and benchmark transformer-based models.