Demystifying Neural Fake News via Linguistic Feature-Based Interpretation

Proceedings of the 29th International Conference on Computational Linguistics, pages 6586–6599 October 12–17, 2022. 6586 Demystifying Neural Fake News via Linguistic Feature-Based Interpretation Ankit Aich∗and Souvik Bhattacharya∗and Natalie Parde Natural Language Processing Laboratory Department of Computer Science University of Illinois at Chicago {aaich2, sbhatt49, parde}@uic.edu Abstract The spread of fake news can have devastating ramifications, and recent advancements to neu- ral fake news generators have made it challeng- ing to understand how misinformation gener- ated by these models may best be confronted. We conduct a feature-based study to gain an interpretative understanding of the linguistic attributes that neural fake news generators may most successfully exploit. When comparing models trained on subsets of our features and confronting the models with increasingly ad- vanced neural fake news, we find that stylistic features may be the most robust. We discuss our findings, subsequent analyses, and broader implications in the pages within. 1 Introduction The internet is a massive and growing source of information (Lee et al., 2021) of varying veracity. The spread of misinformation has been identified as a global risk, with fake information being observed to diffuse faster, farther, deeper, and more broadly than the truth. Studies have found that falsehood is seventy percent more likely to be shared online than the truth (Vosoughi et al., 2018), and most social media platforms either do not filter fake news or do it poorly (Wardle and Singerman, 2021). Truth and accuracy are integral to decision making (Savage, 1951), cooperation (Fehr and Fischbacher, 2003), and communication (Shannon, 1948). Across numerous modern events (Mendoza et al., 2010; Gupta et al., 2013) as well as historically (Burkhardt, 2017), people have been manipulated by the spread of false news. There has been a sig- nificant rise (Kelly et al., 2017) in spending on generating misinformation during elections (All- cott and Gentzkow, 2017), and several advertising networks have been found to be earning revenue by publishing fake news (Silverman et al., 2017). Health-related misinformation holds an immediate ∗Authors contributed equally. danger to the public (Chou et al., 2018). Misinfor- mation about vaccines caused a decline in intent to take the COVID-19 vaccine by 6.4% in September 2020 (Loomba et al., 2021), and false information by anti-vaxxers on social media fueled a tripling in measles cases in the United Kingdom (Sheridan, 2019). In the Democratic Republic of Congo, it was found that “nearly half of respondents believed that Ebola didn’t exist or was invented to destabi- lize the region or to make money” (York, 2019). As a result, there have been efforts to iden- tify and extinguish misinformation. Manual fact checking is time-consuming and often comes too late—over 50% of viral social media claims happen within the first ten minutes of being posted (Shaar et al., 2020), making automated detection more appealing. Nonetheless, automated models for de- tecting misinformation are imperfect, and their mis- takes may give rise to devastating outcomes. Given the prevalence of deep learning models and the re- cent concerning proliferation of neural fake news generators, it …

Content truncated to 500 words