Continually Detection, Rapidly React: Unseen Rumors Detection Based on Continual Prompt-Tuning

Proceedings of the 29th International Conference on Computational Linguistics, pages 3029–3041 October 12–17, 2022. 3029 Continually Detection, Rapidly React: Unseen Rumors Detection based on Continual Prompt-Tuning Yuhui Zuo1, Wei Zhu2, Guoyong Cai1, 1Guilin University of Electronic Technology, Guilin, China 2East China Normal University, Shanghai, China zzuuoo2019ok@gmail.com,ccgycai@guet.edu.cn,wzhu@stu.ecnu.edu.cn Abstract Since open social platforms allow for a large and continuous flow of unverified informa- tion, rumors can emerge unexpectedly and spread quickly. However, existing rumor de- tection (RD) models often assume the same training and testing distributions and can not cope with the continuously changing social network environment. This paper proposed a Continual Prompt-Tuning RD (CPT-RD) framework, which avoids catastrophic forget- ting (CF) of upstream tasks during sequential task learning and enables bidirectional knowl- edge transfer between domain tasks. Specifi- cally, we propose the following strategies: (a) Our design explicitly decouples shared and domain-specific knowledge, thus reducing the interference among different domains during optimization; (b) Several technologies aim to transfer knowledge of upstream tasks to deal with emergencies; (c) A task-conditioned prompt-wise hypernetwork (TPHNet) is used to consolidate past domains. In addition, CPT- RD avoids CF without the necessity of a re- hearsal buffer. Finally, CPT-RD is evaluated on English and Chinese RD datasets and is ef- fective and efficient compared to prior state-of- the-art methods. 1 1 Introduction Online platforms such as social media are facing new and ever-evolving cyber threats at the infor- mation level — rumor. A rumor is an unconfirmed claim related to an object, event, or issue of pub- lic concern that is spread when its integrity is un- known (Guo et al., 2020). It is very necessary to study automated RD, because rumors are ex- tremely harmful to society and manual detection is time-consuming and labor-intensive (Oshikawa et al., 2018). However, automated RD has significant chal- lenges and still faces the following difficulties: 1Innovation Project of Guangxi Graduate Education Event 1 Event 2 Event 3 (a) MH370 Olympics COVID-19 !! "#$ (b) Event 1 Event 2 Event 3 MH370 Olympics COVID-19 !! "#$ !! "#% !! "#& Time t = 1 t = 2 t = 3 Train Predict ③ Labeled training Performance … … ① Zero-shot Performance ② Few-shot Performance Transfer Figure 1: Illustration of the RD model training proce- dure in the traditional static (a) and continual dynamic event transfer (b) setup. First, rumors are highly event domain-specific (Wang et al., 2018), each event domain may have a different input text distribution and fraudulent in- tent. Second, detecting rumors at their early stage of spreading faces the problem of insufficient la- beled samples (Zhou et al., 2020). Third, rumor detectors operating on online social platforms of- ten encounter continuous event domain changes. This poses a significant challenge to existing RD models. Previous work (Wang et al., 2018; Bian et al., 2020; Zhang et al., 2021; Lin et al., 2021; Ben- David et al., 2021) usually assume the same distri- bution of training and testing data, and have dif- ficulty coping with changing social network en- …

Content truncated to 500 words