Abstract:
Classification of fake news content is one of the challenging problems of Natural
Language Processing (NLP). In classification of fake news detection, the classes of true and
fake news are predefined in which the news are assigned based on model’s judgement. The
increase in social media gave an edge to spreading of fake news easily. it has now become
one of the considerable menaces to journalism, democracy and freedom of expression. In this
era, fake news has emerged as a world topic, and it has become of major concern for the
people to know the authenticity of a news content over the social media. The existing contentbased approaches such as rule based, probabilistic and machine learning are used for
classification. These approaches are far from achieving acceptable accuracy with fake news
detection due to the complex nature of the news content that is generated to mislead the
audience. These models require handcrafted features, which has the possibility of missing
out the important features or considering the unimportant features. Secondly, these traditional
models lack the ability of memory element to keep the track of previous words as well as
current appearing words also known as words dependency, which is one of great importance
in the classification of fake news. In this research, we have proposed classification of fake
news model that comprise of news content representation scheme also known word
embeddings and deep learning model that represent the news articles as latent features of the
text. The proposed model for classification of fake news is a blend of two DL models
consisting of 1D Convolution Neural Network (CNN) as feature extractor and LSTM as
classification model. Our model outperforms the state of art by a well-off accuracy on three
known datasets of fake news detection such as 98.6% on FakeNewsNet datset, 97.30% on
ISOT dataset and 90.08% on FA-KES dataset.
