ABSTRACT
Social media outlets are transforming more and spreading fake news at an incredible scale. It is a tedious process to perform verification of legitimacy news in a manual way since it affects the world on a large scale. There are a lot of existing fake news detection techniques such as content-based and context-based classification, image-based classification, etc. where the text sequence is analyzed in a unidirectional manner. A hybrid approach enables better classification of fake news in real-life scenarios. A hybrid model bidirectional encoder representation from transformers and bidirectional long-short term memory (BERT-BiLSTM) is implemented for the accurate detection of fake news using the FakeNewsNet dataset. The process can be categorized into modules like data preprocessing, and word embedding where feature vectors are generated followed by model training, and testing. BERT, a deep transformer model uses an attention mechanism and can process a very long sequence with two networks, one for encoding the headlines, and another for encoding the body content. BiLSTM will be capable of extracting the maximum number of features in different layers. The popularity prediction for identifying fake news is carried out using a regression model, which classifies the headlines as positive, negative, or neutral and helps people to recognize it.
