Hidden Biases in Unreliable News Detection Datasets

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 2482–2492 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 2482 Hidden Biases in Unreliable News Detection Datasets Xiang Zhou1, Heba Elfardy2, Christos Christodoulopoulos2 Thomas Butler2, Mohit Bansal1 1UNC Chapel Hill 2Amazon {xzh, mbansal}@cs.unc.edu {helfardy, tombutl}@amazon.com, chrchrs@amazon.co.uk Abstract Automatic unreliable news detection is a re- search problem with great potential impact. Recently, several papers have shown promis- ing results on large-scale news datasets with models that only use the article itself without resorting to any fact-checking mechanism or retrieving any supporting evidence. In this work, we take a closer look at these datasets. While they all provide valuable resources for future research, we observe a number of prob- lems that may lead to results that do not gen- eralize in more realistic settings. Specifically, we show that selection bias during data collec- tion leads to undesired artifacts in the datasets. In addition, while most systems train and pre- dict at the level of individual articles, overlap- ping article sources in the training and evalua- tion data can provide a strong confounding fac- tor that models can exploit. In the presence of this confounding factor, the models can achieve good performance by directly memo- rizing the site-label mapping instead of mod- eling the real task of unreliable news detec- tion. We observed a significant drop (>10%) in accuracy for all models tested in a clean split with no train/test source overlap. Us- ing the observations and experimental results, we provide practical suggestions on how to create more reliable datasets for the unreli- able news detection task. We suggest future dataset creation include a simple model as a difficulty/bias probe and future model develop- ment use a clean non-overlapping site and date split.1 1 Introduction The proliferation of unreliable news is widely ac- knowledged (Del Vicario et al., 2016; Lazer et al., 2018; Vosoughi et al., 2018), and its identification 1Our code is publicly available at https://owenzx. github.io/unreliable_news is a socially important problem. In this work we use the label unreliable news as a broad term for all un- verifiable and misleading news content, regardless of whether the content is malicious (targeted mis- information) or not. Accordingly, while specific definitions vary in different datasets used in this work, we refrain from using the term “fake” since identifying the intent of the author(s) is beyond the scope of this work. To mitigate the problem of surfacing unreliable news content, various websites (e.g., PolitiFact2, Media Bias/Fact Check (MBFC)3, GossipCop4, etc.) determine the reliability of news by manually fact-checking the important claims in given news articles. Beyond requiring investigative expertise, manual fact-checking is time-consuming and is thus limited to only a small set of selected news articles. Recent research has explored automating this process using machine learning methods to auto- matically determine news veracity (P´erez-Rosas et al., 2018; Baly et al., 2018; Nie et al., 2019; Wright and Augenstein, 2020). These efforts were made possible due to the …

Content truncated to 500 words