Analysis of contextual features’ granularity for fake news detection

While the world is battling COVID-19 pandemic and its variants; netizens are combating Infodemic – “Proliferation of fake news online”. Spread of fake news during this global pandemic COVID-19 has dangerous consequences. Precise automated fake news detection is the need of the hour. This is the driving force behind this study. Intrinsic quality of news data: precision and objectivity can be studied to detect the credibility. However, to gain the knowledge or inference from it further it is a challenge. To address this challenge, this work proposes derivation of features into two categories: intrinsic word level features (Fine-grained) and sentence level features (Coarse- grained).For experimentation, fine-grained features are learned from word vector representation of the news articles. Psycho-linguistic and sentiment level word level features are derived using Empath library. Coarse-grained features (sentence-level) comprise of vectors generated by Text summarization and DOC2Vec. Taking the advantages of existing approaches, this paper proposes a new framework Granularity based Fake news Detection (GRAFED) that explores fusion of fine and coarse-grained features. The fusion of feature set is more powerful than individual fine or coarse-grained feature sets as they can capture complex interdependence of words in the sentence along with the semantics. Exhaustive experimentation using traditional classifiers with hybrid granular feature vector of GRAFED outperformed the existing approaches for publicly available state-of-art LIAR dataset. Experimental results show that the hybrid feature set is superior to individual feature set and the results are promising when compared to the existing state-of-art approaches as analyzed in the comparative analysis section.