Prompt Learning for Low-Resource Multi-Domain Fake News Detection

Abstract

The spread of fake news has caused severe damage to people's lives and society nowadays. Social media is inundated with fake news from multiple domains. Previously, the methods used to detect fake news have tended to be limited to single domains and have performed inadequacy in other or multiple domains. Therefore, the detection of multi-domain fake news has garnered significant interest. However, multi-domain fake news detection approaches rely more on sufficient training samples. In the real world, the low-resource problem has become a significant challenge that restricts the detection of fake news in multi-domain. In that case, prompt learning approaches have significant advantages in low-resource scenarios, but the existing fake news detection approaches based on prompt learning differ greatly in performance across different domains. In addition, the verbalizer in the prompt learning framework is the key module for mapping label words to classification labels, and the performance of existing methods is also limited by simply designed verbalizer modules, which makes the label words coverage small and the label words prediction inaccurate, especially in zero-shot scenarios. In this paper, we propose prompt learning for low-resource multi-domain fake news detection(PLDFEND). We incorporate domain-aware and relational learning into the prompt learning framework to improve the prompting effect. In addition, the verbalizer is optimized to adapt to different scenarios and map the label words to classification labels, thus achieving multi-domain news classification detection. After conducting comprehensive experiments in settings with limited resources and abundant data, we have confirmed the effectiveness of PLDFEND.