A Novel Approach for Tweet Similarity in a Context-Aware Fake News Detection Model

In today’s information-driven world, identifying fake news is more critical than ever. The spread of false information threatens societal stability, eroding trust in institutions and fostering community polarization. While numerous algorithms and methods have been developed to detect fake news, their effectiveness varies depending on the context and data used. Recent advancements, particularly those leveraging transformer-based models like BERT, have significantly improved detection performance, especially in multimodal environments. Social media platforms are the primary focus of fake news research, as they serve as fertile grounds for the unchecked spread of misinformation. Their rapid dissemination capabilities and global reach make them ideal channels for propagating disinformation. Addressing these challenges requires innovative approaches and robust solutions. In response to these essential issues, this study presents a formal framework for detecting fake news using a multilayered approach. The proposed three-tiered model includes the topic, social, and context layers, providing a comprehensive approach to identifying fake news. One of the critical tasks within the topic layer involves assessing the similarity between two messages. To achieve this, we developed a novel method for evaluating the similarity of two tweets by extending a state-of-the-art model. Our approach leverages FastText vectors, cosine similarity, and word positions within sentences to improve accuracy. For verification, we utilized the STSBenchmark dataset and employed correlation values to measure our model’s quality. We conducted a rigorous 50-fold evaluation of the validation data. Ultimately, our model achieved a superior correlation value (median 0.673528) compared to another leading model, which had previously delivered the best results in the field. This paper presents one of the important elements of a comprehensive tool for detecting fake news, with some components already published in other works by the authors or planned for future development (e.g. multimodal content processing or multi-lingual similarity).