Automatic detection of fake news using recurrent neural network—Long short-term memory

Abstract

The propagation of deliberate misinformation is gaining significant momentum, especially across different social media platforms. Despite the fact that there are several fact-checking blogs and websites that distinguish between news that is genuine or fake, and regardless of the reality that there is ongoing research being conducted to restrict the propagation of fake news, the issue is still one that needs to be addressed. The most major barrier is a failure to spot monitors and to disclose false news in a reasonable timeframe, both of which are critical components. In this research, a system is proposed that utilizes deep learning model of Recurrent Neural Network—Long Short-Term Memory (RNN-LSTM) in order to put an end to the circulation of misleading information and put a stop to, and expose instances of fake news that are spread through reliable channels. The process of extracting hybrid features from text data, such as Lemmas, Bi-Gram, Tri-Gram, N-gram, Term Frequency Inverse Document Frequency (TF-IDF), part-of-speech, and dependency-based natural language processing features, is developed as a strategy. When compared to other traditional approaches to machine learning classifier, the RNN-LSTM method that was proposed obtains a greater level of accuracy than those other approaches. It achieves an accuracy rate of 99.10% both during training and testing, which is better to the accuracy achieved by common machine learning approaches such as Support Vector Machine (SVM), Artificial Neural Network (ANN), Random Forest (RF), Naïve Bayes (NB), and Decision Tree (DT). In the proposed approach for detecting false news using RNN-LSTM, the three experiments are conducted to acquire, accuracy, precision, recall, and F-score with varying forms of cross validation (5-Fold, 10-Fold and 15-Fold). Based on the findings of the empirical research, the conclusion can be drawn that the RNN-LSTM with ReLu function provides more accurate detection than both the RNN-LSTM (Tan h) function and the RNN-LSTM (sigmoid) function with 15 fold cross validation.