Enhancing Social Media Health Prediction Certainty by Integrating Large Language Models with Transformer Classifiers

AbstractThis paper presents our approach for SMM4H 2024 Task 5, focusing on identifying tweets where users discuss their child’s health conditions of ADHD, ASD, delayed speech, or asthma. Our approach uses a pipeline that combines transformer-based classifiers and GPT-4 large language models (LLMs). We first address data imbalance in the training set using topic modelling and under-sampling. Next, we train RoBERTa-based classifiers on the adjusted data. Finally, GPT-4 refines the classifier’s predictions for uncertain cases (confidence below 0.9). This strategy achieved significant improvement over the baseline RoBERTa models. Our work demonstrates the effectiveness of combining transformer classifiers and LLMs for extracting health insights from social media conversations.