Bigfoot in Big Tech: Detecting Out of Domain Conspiracy Theories

Proceedings of Recent Advances in Natural Language Processing, pages 353–363 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_040 353 Bigfoot in Big Tech: Detecting Out of Domain Conspiracy Theories Matthew Fort, Zuoyu Tian, Elizabeth Gabel, Nina Georgiades, Noah Sauer, Daniel Dakota, Sandra K¨ubler Indiana University {mattfort,zuoytian,eligabel,ngeorgia,sauerno,ddakota,skuebler} @iu.edu Abstract We investigate approaches to classifying texts into either conspiracy theory or mainstream using the Language Of Conspiracy (LOCO) corpus. Since conspiracy theories are not monolithic constructs, we need to identify approaches that robustly work in an out-of- domain setting (i.e., across conspiracy topics). We investigate whether optimal in-domain set- tings can be transferred to out-of-domain set- tings, and we investigate different methods for bleaching to steer classifiers away from words typical for an individual conspiracy theory. We find that BART works better than an SVM, that we can successfully classify out-of-domain, but there are no clear trends in how to choose the best source training domains. Addition- ally, bleaching only topic words works better than bleaching all content words or completely delexicalizing texts. 1 Introduction With the rise of social media over the last 10 years, there has also been a rise in the uses of the internet to spread different types of information, some of it of a more questionable nature. We are interested in the spread of conspiracy theories, which have mor- phed from a fringe phenomenon to a more widely visible, mainstream phenomenon. Along with the increasing spread of misinformation, conspiracy theories have been shown to polarize opinions to extremes and to incite violence (Douglas and Sut- ton, 2018; Enders et al., 2022). While conspiracy theories are often seen as monolithic belief systems, the truth is more com- plex: People who admit to believing a specific conspiracy theory tend to also believe in other con- spiracy theories, but they may only believe differ- ent subsets of factoids associated with a specific conspiracy theory (Enders et al., 2021). For any computational approach to detecting conspiracy theories, this means that we cannot expect to have access to accurate training data. Instead, we will face novel mixes of factoids and conspiracy theo- ries, which deviate from existing training data. For this reason, we investigate here whether it is possi- ble to find out-of-domain conspiratorial texts. We use the Language Of Conspiracy (LOCO) corpus (Miani et al., 2021) to develop classifiers that label a text as either conspiratorial or mainstream, and we investigate under which conditions such clas- sifiers work robustly out-of-domain. More specif- ically, we investigate bleaching methods to steer the classifiers away from words that are typical for a single conspiracy theory (e.g., ’global warming’ for conspiracy theories revolving around climate change). The remainder of this paper is structured as fol- lows: Section 2 explains our research questions, section 3 describes related work, and section 4 describes our data and methodology. Section 5 describes our results for the in-domain setting (section 5.1), for the out-of-domain setting (sec- tion 5.2), and for the bleaching experiments (sec- tion 5.3). We conclude in section 6. 2 Research …

Content truncated to 500 words