Fake News Detection For Amharic Language Using Deep Learning

Abstract

The new media age we are living in has enabled many of us to be connected through the internet, which has been changing the way updates arrive at hand and from whom we get those pieces of information. Due to the usage of such convenient technology for ill intent, fake news dissemination has become a great problem to societies all over the globe. Especially the phenomenon is more common in the time of election and chaos like we have been experiencing in the COVID-19 pandemic. Despite the issue that fake news and misinformation are not constricted to language and culture, most of the attempts to automate the detection of fake news are concerned with a specific group of languages, especially the English language. In this study, we used a newly collected and annotated dataset of 12000 news to build an automated fake news detection system for one of a low-resourced language; Amharic, using deep learning algorithms. The research employed several experiments to determine the best performing deep learning architecture among the ones used in the field of Natural Language Processing and got the Bidirectional Gated Recurrent Unit(Bi-GRU) and Convolutional Neural Network(CNN) to outperform over the other recurrent and attention-based models. The CNN model surpasses all other models with an accuracy of 93.92% and an f1-score of 94%. The effect of Morphological normalization on the Amharic fake news detection was also accessed over the best two performing models and the experiment revealed that applying normalization has an unfavorable effect on the classification performance that reduces the f1- measure of both models from 94% to 92%. In addition to that, different combinations of CNN hyperparameters are tested. Even though significant improvement in terms of performance was not seen from the tuning process, an important correlation between the performance of the model and hyperparameters was seen. Besides our contribution in the evaluation of these deep learning models to one of the morphologically rich languages, we expect the newly proposed dataset will be a reason for more research and findings regards to the detection and prevention of Amharic fake news soon