Human-in-the-loop Evaluation for Early Misinformation Detection: A Case Study of COVID-19 Treatments

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 15817–15835 July 9-14, 2023 ©2023 Association for Computational Linguistics Human-in-the-loop Evaluation for Early Misinformation Detection: A Case Study of COVID-19 Treatments Ethan Mendes, Yang Chen, Wei Xu, Alan Ritter Georgia Institute of Technology {emendes3, yangc}@gatech.edu {wei.xu, alan.ritter}@cc.gatech.edu Abstract We present a human-in-the-loop evaluation framework for fact-checking novel misinforma- tion claims and identifying social media mes- sages that support them. Our approach extracts check-worthy claims, which are aggregated and ranked for review. Stance classifiers are then used to identify tweets supporting novel misin- formation claims, which are further reviewed to determine whether they violate relevant poli- cies. To demonstrate the feasibility of our ap- proach, we develop a baseline system based on modern NLP methods for human-in-the-loop fact-checking in the domain of COVID-19 treat- ments. We make our data1 and detailed annota- tion guidelines available to support the evalua- tion of human-in-the-loop systems that identify novel misinformation directly from raw user- generated content. 1 Introduction As many people now get information from social networking websites such as Facebook and Twit- ter, misinformation has become a serious societal problem. To address this, social media companies have spent billions on content moderation.2 Prior work on developing natural language processing systems to combat misinformation has mainly fo- cused on various sub-tasks (Lee et al., 2021; Guo et al., 2022), including claim detection (Eger et al., 2017; Li et al., 2022), evidence retrieval (Jiang et al., 2020; Samarinas et al., 2021; Wan et al., 2021; Aly and Vlachos, 2022), fact verification (Aly et al., 2021; Wu et al., 2022; Chen et al., 2022; Gu et al., 2022), stance classification (Thorne et al., 2017; Conforti et al., 2018; Li et al., 2019), and fal- lacy recognition (Alhindi et al., 2022). Researchers have also attempted to perform early detection of novel misinformation claims (Yue et al., 2022), as it 1https://github.com/ethanm88/hitl-eva luation-early-misinformation-detection 2https://www.cnbc.com/2021/02/27/cont ent-moderation-on-social-media.html is crucial for supporting early interventions such as pre-bunking (Lewandowsky and Van Der Linden, 2021). However, evaluations are often set up auto- matically using datasets that were retrospectively constructed based on a predefined set of debunked claims. Recent work by Glockner et al. (2022) pre- sented convincing evidence that existing NLP fact- checking pipelines are unsuitable for detecting novel real-world misinformation. They show these systems rely on leaked counter-evidence from news sources that have already fact-checked the claim. In general, it is unrealistic to assume this type of evidence will be available for new claims that have not yet been widely spread. In this paper, we address this challenge by pre- senting a more realistic human-in-the-loop detec- tion and evaluation framework that can measure a system’s capabilities for detecting novel check- worthy claims in the wild (see Figure 1). We fo- cus on discovering new, domain-specific claims from raw tweets which are then verified by humans, rather than relying on a pre-defined list of claims that have already been fact-checked for evaluation. More importantly, we consider not …

Content truncated to 500 words