HYBRID NEURAL NETWORK FOR FAKE NEWS STANCE DETECTION

Abstract

Fake news is becoming more readily available as technology advances, which sometime mislead the readers and leads to inaccurate social opinions. Fake news may be found on the Internet, news sources and social media platforms. The spread of low-quality news has harmed both individuals and society. In this thesis work, we analyze three hybrid models, CNN+simple RNN, CNN+GRU and CNN+BiLSTM in encoder decoder architecture to predict the stance between headline and article of the news. Pre-trained GloVe word embedding is used for word to vector representation as it can capture the inter-word semantic information. The CNN-RNN combination had been shown efficient in deep learning applications because they can capture sequential and local features of input data. The models were successfully trained and tested on both binary (ISOT) and multiclass (FNC-1) fake news datasets. It is found that the CNN+ BiLSTM model had better results than other two hybrid models in both binary and multiclass classification task for the fake news stance detection system.