| Resumen |
The fact that it might be difficult to distinguish between real news and fake news is of the utmost significance. It is difficult, time-consuming, and expensive to manually review enormous volumes of digital data in order to detect false news, which is information that is not real in line with the facts. There is a new piece of fake news published every second; hence, the creation of automated systems is absolutely necessary in order to detect instances of fake news. This study provides a comprehensive analysis of various machine and deep learning approaches, as well as how the systems perform in identifying fake Urdu news, which are submitted in the shared tasks named UrduFake@FIRE2020 and UrduFake@FIRE2021. It was hoped that the shared tasks would attract new researchers to come up with algorithms that could detect Urdu fake news articles available in the digital media. More than 50 teams from ten different nations participated to find a solution to this binary classification issue. The objective of the shared tasks was to classify a given Urdu news instance as either real or fake, and the teams proposed numerous algorithms to achieve this goal. Among the several methods of text representation that have been investigated in this study are count-based BoW features and word vector embeddings. Traditional neural and non-neural network approaches, including BERT, XLNet and RoBERTa are also examined, and their analysis is presented. Moreover, precision, recall, F1Real, F1Fake, and F1Macro are used for evaluation. The winning system in 2021 employed the linear classifier function and achieved 0.679 F1Macro, while the highest performing system in 2020 was based on BERT and obtained 0.907 F1Macro. Thus, the results of this research show that machine learning classifiers underperformed in detecting Urdu fake news than deep learning techniques. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG. |
