Improving Health Mention Classification of Social Media Content Using Contrastive Adversarial Training

Health mention classification (HMC) involves the of an input text as health or not. Figurative and non-health disease words makes task challenging. Learning context is key to this problem. The idea learn word representation by its surrounding utilize emojis in help improve results. In paper, we using adversarial training that acts a regularizer during fine-tuning model. We generate examples perturbing embeddings model then train on pair clean examples. Additionally, contrastive loss tries similar representations for example perturbed version. evaluate method three public datasets. Experiments show improves performance significantly terms F1-score over baseline methods both BERTLarge RoBERTaLarge all Furthermore, provide brief analysis results utilizing power explainable AI.