A large dataset which consists of linguistic characteristics of fake and real online articles. Specifically, LinCFNA consists of 320,960 total time-stamped article observations, which are labelled as either fake or real, and are characterized by 534 different linguistic features.
In order to analyse how text-linguistic characteristics differ between real and fake sources during time, a plethora of different linguistic features were extracted from online news articles which were cralwed between 2009-2019, by gathering the available knowledge from the text. The extraction of these features is based on a previous recent study (Paschalides, Pallis, and Dikaiakos 2021), where 534 features based on Stylistic, Complexity and Psychological aspects of the articles were utilized.
Due to the large size of the data, the files can be downloaded from the following google-drive repository.
LinCFNA_Features.csv: Contains the 534 different linguistic features for the 320,960 time-stamped article observations, which are labelled as either fake or real
LinCFNA_Text.csv: Contains the text and meta-data for the 320,960 time-stamped article observations, which are labelled as either fake or real
clean_non_missing_features_politics_labels_75_lasso_34_features.csv: Containts the political articles' features that were selected with Lasso Logistic Regression
Regarding the Data Collection process, a list of collected articles was constructed by crawling the WebArchive for news articles published from 2009 through 2019. The articles were divided into untrusted and trusted ones by defining two pairs of domain credibility lists (Hagen 2017). The former contains domain names that usually publish Fake News and are highly scrutinized by fact-checking organizations, including Snopes, PolitiFact, and others. The latter includes high reputation domains, which have rarely or never been criticized by fact-checking sites. To verify the validity of these lists, labels provided by the independent online fact-checking outlet MediaBiasFackCheck-MBFC (Van Zandt 2020) were used. MBFC specifies how often a domain publishes factual news by employing seven labels ranging from VERY LOW to VERY HIGH. MBFC has been widely used in the literature(Chen and Freire 2020). Thus, utilizing the knowledge of these lists, the possibility of false positives was limited.
Please cite our paper if you find the work useful:
https://www.nature.com/articles/s41598-023-32952-3#citeas
@article{petrou2023multiple,
title={A Multiple change-point detection framework on linguistic characteristics of real versus fake news articles},
author={Petrou, Nikolas and Christodoulou, Chrysovalantis and Anastasiou, Andreas and Pallis, George and Dikaiakos, Marios D},
journal={Scientific Reports},
volume={13},
number={1},
pages={6086},
year={2023},
publisher={Nature Publishing Group UK London}
}
Chen, Zhouhan, and Juliana Freire. 2020. “Proactive Discovery of Fake News Domains from Real-Time Social Media Feeds.” In Companion Proceedings of the Web Conference 2020, 584–92.
Hagen, Sam. 2017. “OpenSources: Professionally curated lists of online sources.” https://github.com/BigMcLargeHuge/opensources.
Paschalides, Demetris, George Pallis, and Marios D Dikaiakos. 2021. “POLAR: A Holistic Framework for the Modelling of Polarization and Identification of Polarizing Topics in News Media.” In Proceedings of the 2021 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, 348–55.
Van Zandt, Dave. 2020. “Media Bias/Fact Check News: An American Fact-Checking Website).” https://mediabiasfactcheck.com/.