Predicting Fake News using GloVe and BERT Embeddings

Abstract

The growth of fake news in multiple fields such as in the political or health sector has become a great concern as it possess huge impact on the reader's mind. Identifying the fake news or differentiating between fake and authentic news is quite challenging. The focus of this research is to identify fake news by applying different artificial intelligence techniques along with different embeddings and to assess the performance of all the applied models. The performance of these models and the embeddings is compared based on precision, accuracy, Fl-score and recall. For machine learning techniques SVM, KNN, Naive Bayes, Logistic Regression and Decision Trees are used, while for deep learning techniques CNN and LSTM are used with GloVe and BERT embeddings. Multiple experiments using these techniques are performed on the LIAR and Fake-or-Real dataset. Naïve Bayes has shown the best results from machine learning techniques on both datasets. While in deep learning techniques, LSTM with GloVe has shown the best results on the LIAR dataset and CNN with BERT has shown the best performance on the Fake-or-Real dataset. Overall GloVe word embeddings performed well on the LIAR dataset while BERT sentence embeddings have shown good performance on the Fake-or-Real dataset.