Improving Fake News Detection of Influential Domain via Domain- and Instance-Level Transfer

Proceedings of the 29th International Conference on Computational Linguistics, pages 2834–2848 October 12–17, 2022. 2834 Improving Fake News Detection of Influential Domain via Domain- and Instance-Level Transfer Qiong Nan1,2, Danding Wang1, Yongchun Zhu1,2, Qiang Sheng1,2 Yuhui Shi1,2, Juan Cao1,2∗, Jintao Li1 1Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences 2University of Chinese Academy of Sciences {nanqiong19z,wangdanding,zhuyongchun18s}@ict.ac.cn {shengqiang18z,caojuan,jtli}@ict.ac.cn shiyuhui221@mails.ucas.ac.cn Abstract Both real and fake news in various domains, such as politics, health, and entertainment are spread via online social media every day, ne- cessitating fake news detection for multiple do- mains. Among them, fake news in specific domains like politics and health has more seri- ous potential negative impacts on the real world (e.g., the infodemic led by COVID-19 misin- formation). Previous studies focus on multi- domain fake news detection, by equally mining and modeling the correlation between domains. However, these multi-domain methods suffer from a seesaw problem: the performance of some domains is often improved at the cost of hurting the performance of other domains, which could lead to an unsatisfying perfor- mance in specific domains. To address this issue, we propose a Domain- and Instance-level Transfer Framework for Fake News Detection (DITFEND), which could improve the perfor- mance of specific target domains. To trans- fer coarse-grained domain-level knowledge, we train a general model with data of all domains from the meta-learning perspective. To trans- fer fine-grained instance-level knowledge and adapt the general model to a target domain, we train a language model on the target domain to evaluate the transferability of each data in- stance in source domains and re-weigh each instance’s contribution. Offline experiments on two datasets demonstrate the effectiveness of DITFEND. Online experiments show that DIT- FEND brings additional improvements over the base models in a real-world scenario. 1 Introduction With the rapid popularization of the Internet, more and more people tend to acquire news through so- cial media platforms, such as Weibo1 and Twit- ter2. Due to the above phenomenon, fake news ∗Corresponding author 1https://weibo.com 2https://twitter.com Politics Politics Entertain ment Politics Entertain ment Politics Entertain ment Politics Politics Training Testing Single-domain Multi-domain Cross-domain … … … Figure 1: Illustration of the differences among single- domain, multi-domain, and cross-domain (ours) fake news detection. has spread widely all over the world. During the 2016 U.S. presidential election campaign, the top 20 frequently discussed fake election stories gener- ated 8,711,000 engagements on Facebook, which is larger than the total of 7,367,000 for the top 20 most-discussed election stories posted by 19 major news websites (Silverman, 2016). The wide spread of fake news may break the authenticity balance of the news ecosystem (Shu et al., 2017), and it not only misleads many people but also leads to the so- cial mobs and social panic (Chen, 2020). Therefore, fake news detection is of significant importance. Real-world news platforms categorize news pieces by topic into various domains, e.g., poli- tics, health, and entertainment. Although it is nec- essary to detect fake news in every domain, the …

Content truncated to 500 words