Certain investigation for text classification on fake news data using optimized deep learning techniques

Abstract

In social media platforms, numerous data are generated every second newlineand the users of these platforms share the contents instantly when they like to newlineshare. The contents posted in social media are mostly unverified and there is a newlinehigh possibility of the presence of inaccurate information. Spread of fake newlineinformation on social media misleads the users about some situations or newlinetopics in political and social sectors. Rumour data classification in social newlinemedia seems to be an area of recent research due to its dependency on digital newlinecommunications, and this rumour data make social media unstable. newlineIn the field of Natural Language Processing (NLP), research on rumour newlinedetection has its challenges with word embedding, feature extraction and newlineclassification tasks. Previously, many researches have been conducted, but the newlineaccuracy of rumour detection is a big issue for identifying infrequent/rare newlinewords. The traditional word embedding models like Word2Vec, FastText and newlineso on., perform word level analysis for finding word vectors. In recent days, newlinethese models are replaced with the novel transformer-based BERT model newlinewhich considers the combination of word pieces for producing efficient newlinefeature vectors. Still, it is difficult to identify which word piece of each token newlineis to be used for vector formation. In the first part of the research, an automated fake news detection system is introduced with three different word embedding techniques such as Word2Vec, GloVe and FastText. Further, it is classified by using two newlineproposed algorithms namely Optimized Cost Sensitive Long Short-Term newlineMemory (OC-LSTM) and Optimized Cost Sensitive Gated Recurrent Unit newline(OC-GRU). For the LIAR and Fake and Real News (ISOT) datasets, the newlineFastText embedding with OC-GRU model yields improved accuracy. newlineTo improve the quality and accuracy of features, in the second part of newlinethe research, two phases of feature extraction techniques are involved for newlineperforming word embedding and deep feature extraction. In the first phase, newlineBidirectional Encoder Representations from Tran