An improved multi-modal framework for fake news detection using NLP and Bi-LSTM

Fake news and misinformation, particularly through social media, are some of the biggest threats to society and individual well-being. Given that the volume of misleading content is on the rise, it becomes crucial to identify efficient ways of detecting such content. Currently, there is a great improvement in using natural language processing (NLP) and machine learning in fake news detection, but the existing methods are often based on a limited set of features. This work aims at fulfilling this gap through the development of a framework that integrates multiple data forms to support detection. In this work, we propose a new approach to fake news detection called Multi-Modal Fake News Detection (MM-FND). This research employs three datasets: The ISOT Fake News Dataset, the LIAR Dataset, and the COVID-19 Fake News Dataset. Global features are obtained using Word2Vec and Term Frequency-Inverse Document Frequency (TF-IDF) and temporal features are obtained using Bi-directional Long Short-Term Memory (Bi-LSTM) networks. Further, spatial features are extracted through Named Entity Recognition (NER) along with Global Vectors for Word Representation (GloVe) embeddings. The last level of classification of news articles is done by the Random Forest classifier, since the various feature extraction methods are expected to complement each other. The proposed framework shows a better detection accuracy than the conventional methods as depicted in the proposed framework. The proposed MM-FMD model has an accuracy of 96.3% with a precision of 95.8%, recall of 97.1% and an F1 score of 96.4% when tested on the ISOT Fake News Dataset. On the LIAR Dataset, the model achieves an accuracy of 95.6% which is slightly lower than the previous dataset, with precision of 93.5%, recall of 95.0% and F1-score of 94.2%. For the COVID-19 fake news dataset, the model performs well, with an accuracy of 97.1%, precision of 98.5%, recall of 96.7% and F1 score of 97.9%.