From Fake to Hyperpartisan News Detection Using Domain Adaptation

Proceedings of Recent Advances in Natural Language Processing, pages 1095–1109 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_117 1095 From Fake to Hyperpartisan News Detection Using Domain Adaptation R˘azvan-Alexandru Sm˘adu1, Sebastian-Vasile Echim1, Dumitru-Clementin Cercel1, Iuliana Marin2, Florin Pop1,3 1University Politehnica of Bucharest, Faculty of Automatic Control and Computers 2University Politehnica of Bucharest, Faculty of Engineering in Foreign Languages 3National Institute for Research & Development in Informatics - ICI Bucharest, Romania razvan.smadu@stud.acs.upb.ro, sebastian.echim@stud.aero.upb.ro {dumitru.cercel,iuliana.marin,florin.pop}@upb.ro Abstract Unsupervised Domain Adaptation (UDA) is a popular technique that aims to reduce the do- main shift between two data distributions. It was successfully applied in computer vision and natural language processing. In the cur- rent work, we explore the effects of various unsupervised domain adaptation techniques be- tween two text classification tasks: fake and hyperpartisan news detection. We investigate the knowledge transfer from fake to hyperpar- tisan news detection without involving target labels during training. Thus, we evaluate UDA, cluster alignment with a teacher, and cross- domain contrastive learning. Extensive experi- ments show that these techniques improve per- formance, while including data augmentation further enhances the results. In addition, we combine clustering and topic modeling algo- rithms with UDA, resulting in improved perfor- mances compared to the initial UDA setup. 1 Introduction Fake news detection is a challenging task in which the goal is to detect whether the news content does not disseminate false information which may harm society. Recently, this problem has broad attention to the research community, especially with the rising interaction with social media plat- forms, which have become one of the primary sources of information for many individuals (Shu et al., 2020). Detecting fake news is challenging for many of us, since some news can be written very convincingly, thus spreading misleading in- formation without control (Ahmed et al., 2017). Therefore, new datasets (such as BuzzFeed-Webis Fake News (BuzzFeed) (Potthast et al., 2018) and ISOT (Ahmed et al., 2017)) and novel detection techniques (Koloski et al., 2022; Mosallanezhad et al., 2022) have emerged in recent years. Especially since the 2016 United States presiden- tial election, a related task, namely hyperpartisan news detection, identifies whether the information spread by the news is in a political extreme (Rae, 2021). Hyperpartisan articles aim to expose infor- mation related to only one perspective, ignoring and, in some cases, even attacking the perspectives from other opposing sides (Kiesel et al., 2019). The consequences of this type of news range from misinformation in the media to an increase in the number of supporters of extreme ideologies (Huang and Lee, 2019). Some works (Potthast et al., 2018; Ross et al., 2021) linked fake news with hyperpartisan news, since their goal is to spread as much as possible and influence people. This phenomenon is related to clickbait (Potthast et al., 2016), as the authors use different techniques to make the content more accessible and viral on the media (Kiesel et al., 2019). Recently, many architectures based on Bidirec- tional Encoder Representations from Transformers (BERT) (Devlin et al., 2019) have been developed and fine-tuned …

Content truncated to 500 words