A BERT–RESNET MULTIMODAL FUSION FRAMEWORK FOR FAKE NEWS CLASSIFICATION

Source abstract

The amount of data, especially text data, expands very fast over time. Recently, fake news has become a big topic worldwide. People use social media for news because it's affordable, easy to use, and spreads information quickly. It can cause problems in public, politics, the economy, and organizations. People and society may tolerate greatly as a result of the increase of false information. Finding fake news during recent events is a major challenge in social media detection. A multi-modal feature extractor is used to get textual and visual information. It works with fake news detectors to learn features that help in identifying fake news. This identifier is used to remove unique features of an event while keeping features that are common across all events. It is advised to use them for tokenization and feature extraction from text data since this process results in feature extraction and vectorization. The best features for the highest precision will then be tested and selected using features selection techniques based on the confusion matrix result, which goal is to help users determine if the news is true or not and check if the website is trustworthy. In this project BERT is used, which can learn bidirectional representations from unlabeled text because it is trained on both left and right context in every layer. Therefore, the pre-trained BERT model can be improved to produce high-performing models for a variety of tasks, with just an additional output layer. To deal with these issues, using a ResNet50 hybrid deep neural network(DNN) model for identifying fake news using both text and image components. This study approves a multimodal approach to identify and detecting fake news as real or fake.