BEAST: Leveraging Contrastive Learning and Unsupervised Sentence Embeddings for Improved Drug Abuse Detection

Prescription drug abuse is a growing public health crisis worldwide. In the digital age, social media platforms offer a unique opportunity to monitor drug abuse trends in real-time. However, traditional machine learning models struggle with the informal language, sarcasm, and figurative speech used on social media. This paper proposes BEAST, a novel approach that leverages contrastive learning to improve the detection of drug abuse references hidden within figurative language. Additionally, the integration of SimCSE and Target-Based Generating Strategy further enhances the model's performance by generating superior representations from both labeled and unlabeled data. We test our model on three datasets, and the experimental results demonstrate the superiority of BEAST over the baseline in accurately identifying drug-related references hidden within figurative language on social media. Our work paves the way for more effective public health interventions in this increasingly digital era.