Fake News Detection using CosSim Rule and Semantic Embedding

Abstract

The proliferation of fake news on digital media platforms and online news portals is a majorproblem with information and public trust. With the widespread use of social media, misinformationhas spread at a rapid pace and it is becoming increasingly hard to tell the truth fromthe fiction. Most methods used for traditional fake news detection are based on large annotateddatasets and computationally expensive machine learning models, which lack transparencyand interpretability. To overcome such restrictions, in this research, an explainable fakenews detection framework is proposed, which uses semantic similarity, association rule interestingnessand cross-source verification. The proposed approach involves web scrapping, naturallanguage processing, semantic embedding, cosine similarity, and rule-based classification.Multiple trusted, untrusted, national, international and regional news sources are automaticallycollected using the news articles. The comparison is made between a reference articlefrom an authenticated source and articles about the same event in various media outlets.The Sentence-BERT model is used to generate semantic representations of articles, and cosinesimilarity is used to calculate the contextual similarity between the articles. The articles areclassified either as Real or Fake based on pre-defined similarity thresholds and an interpretablerule-based decision mechanism. The reliability of the generated rules is assessed by usingassociation rule interestingness measures such as Support, Confidence and Lift. Experimentswere performed using the dataset of 51 cross-source news articles, which were automaticallycollected, and manually organized. The results show that the semantic similarity scores andclassification results are significantly better with text preprocessing. Moreover, the rules createdhave confidence levels between 0.76 and 0.91, and lift greater than 1, both of which arestatistically significant, and both associations and classification decisions are reliable. The resultsdemonstrate the efficiency of using semantic similarity approaches for cross-source verificationand the need for explainable rule-based detection in misinformation. The proposedframework offers an alternative of transparency and low overhead to black-box machine learningmodels and serves as a basis for future hybrid fake news detection systems based on thecombination of semantic reasoning and data-driven approaches.