AMUSED: A Multilingual Urdu/English Stance and Fake News Detection Corpus

Online fake news is a significant concern today, as many people rely on social media for news consumption. Although there are various datasets for detecting fake news, most are in English and primarily focus on news content. Only a few datasets include social context and user profile information. Furthermore, no multilingual (English/Urdu) event-based dataset exists specifically for fake news detection. This highlights the need for a comprehensive dataset with features such as user profiles, social interactions on posts, and news content on various events. To address this gap, this study collected a multilingual dataset (AMUSED) focused on recent events in Pakistan. This dataset contains information on 12 recent multi-domain events, 66 claims, 468 source posts, and 36,940 comments. The data set is labeled for stance detection and fake news detection. This data set includes all relevant social context, user profiling, and news content characteristics associated with the source posts. In addition, a benchmark study is presented to evaluate the performance of state-of-the-art models on the proposed dataset. BERT-Base-Multilingual-Cased model is trained for the stance detection task, achieving an 71% F1 score, whereas RoBERTa is used for fake news detection task achieving an F1 score of 79%.