Urdu Fake News Detection Using LSTM and Hybrid CNN-LSTM Models

Abstract

Fake news is an increasing point of interest among the research community as it can be transmitted due to numerous media in one of the shortest periods of time. False information, particularly on those languages that lack substantial resources, has turned into a large issue that is increasingly getting worse due to social media and internet crimes. It is difficult to find fake news in Urdu as it is a complex language and there are not many label datasets, which is why it is not a well-researched discipline. The accuracy of the pre-existing machine learning studies in the field of the Urdu fake news detection has been insufficient. In their present work, the authors make use of the deep learning-based methodology of Long Short-Term Memory (LSTM) and a composite method (CNN-LSTM), i.e., both TF-IDF and word embeddings employing LSTMs to address this problem. The LSTM-based word embedding new method was doing significantly better than the best research, having an 95.85% accuracy and a 94.65% F1 score in the D2 dataset. The model was very accurate with 93.87% and F1 score of 94.21% on D1 data. These findings are an important step toward investigating fake news in Urdu and a promising source to reduce the negative impact of information warping in the online society.