ARCHIVES

The unregulated and fast propagation of fake news over digital media is today a significant issue in the current information ecosystem which erodes trust in the community, distorts democratic procedures, and forms a specific impact on the ethics of society. Despite the incredible success of transformer-based natural language processor models in detecting instances of misinformation by considering the textual analysis, it ignores two fundamental humanistic indicators: the reputation of the source of the news and the level of emotional emotion or expression (in the text). In order to overcome this weakness, the current paper will present a new fake news detection architecture, which combines three complementary streams of features. Initially, fine-tuned transformer-based semantic representations such as BERT and RoBERTa are used to extract contextual semantic representations. Second, there is explicit modeling of source credibility based on a special trust-scoring mechanism which considers the reliability of the publishers. Third, a sentiment analysis module is used to capture sentiment polarity and emotion cues to recognize language patterns of manipulation or biasness. These heterogeneous features are fused with the help of adaptive weighting mechanism which assigns each stream of features dynamically the importance to particular features according to the specificities of the input data. The resultant fused representation is then subjected to fully connected layers to classify the fused representation into fake and real categories. Experimental analyses of benchmark datasets, such as the WELFake, FakeNewsNet, and PolitiFact, confirm that the suggested methodology always beats the text-only transformers baselines. Findings indicate that credibility of the source and emotional context are crucial to consider because they offer a more holistic and human-centered approach to fake news detection in the field.