Prompt-Based Contrastive Learning to Combat the COVID-19 Infodemic

Zifan Peng, Mingchen Li, Yue Wang, Daniel Y. Mo

Abstract: The COVID-19 pandemic has triggered a surge in misinformation and disinformation online, particularly on social media platforms. The World Health Organization has highlighted the urgent need to combat this infodemic, as false information can lead to the spread of conspiracy theories, false remedies, and xenophobia. Efforts to combat the COVID-19 infodemic have faced challenges, with existing research often oversimplifying the issue by focusing solely on verifying the reliability of information. To effectively address this complex problem, a multifaceted approach is necessary. This study aims to broaden the perspective by analyzing texts from diverse angles, considering societal implications and the necessity of government intervention. We employ a prompt-based contrastive learning framework to meet the challenges. A prompt is a passage of text or query fed into a pretrained language model (PLM) to elicit a response. Prompts provide clear guidance and precise context for language models, which has been shown to make them more effective with limited training samples. This approach can help mitigate class imbalance and data sparsity issues in combating COVID-19 infodemics. We also incorporate prompts into a contrastive learning framework to better understand complex utterances within short, informal, unstructured social media texts. Contrastive learning is effective at distinguishing between useful and irrelevant input samples and focusing on the most discriminative features. Our research demonstrates prompt-based contrastive learning can reinforce each other and provide more accurate assessments of input text reliability than current baseline techniques. The framework not only addresses the challenges of data scarcity and class imbalance but also shows potential for application in other text classification tasks, particularly in low-resource languages with extreme class imbalance issues. In conclusion, the prompt-based contrastive learning approach presented in this study offers a promising strategy to combat the infectious diseases infodemic on social media platforms. By considering the multifaceted nature of misinformation and incorporating prompts and contrastive learning, this method provides a more accurate assessment of text reliability, contributing to the broader efforts to mitigate the harmful effects of misinformation during the pandemic.