Lexicon-Stance Based Amharic Fake News Detection

Due to the noisy nature of social media content, and the rapid propagation of fake news, the identification and detection of false information become a challenging problem. In recent years, several studies propose to use content based text representation approaches to automatically detect fake news as early as possible before being propagated into social Medias. However, fake news has a different stylistic nature of writing, and attempting to capture its various unique features may help us improve detection rather than focusing solely on text representation. In this study, we propose to hybrid stance-based features (Page score, headline to article similarity and headline to headline similarities) with the previous text representation lexicon based detection. To build the detection model, we used three machine learning algorithms: Logistic regression, Passive Aggressive and Decision tree. The proposed approach is evaluated using a newly collected Amharic fake news dataset from Facebook. Our experiment results show that the hybrid features (lexicon-stance) are capable of improving the previous lexicon based detection results up to 4.1% accuracy, 3% precision, 4% recall and 4% F1-score using passive aggressive algorithm. In addition, the hybrid feature improves the area under curve from 0.982 to 0.995 by reducing the false positive rate by 4% and improved the true positive rate by 4.4%. Furthermore, we found that Page score, out of the proposed stance features included, has contributed the most to the improvement of lexicon-based fake news detection.