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

Abstract

In this article, we address two problems: the detection of fake news using a multimodal, content, and context-driven approach and the evaluation of short-text messages (in our case, tweets) to compare how similar or different they are. For the first problem, we developed a framework for detecting fake news using a multilayered, multimodal approach consisting of a three-tiered model: the topic layer, the social layer, and the context layer, helping to establish a methodology for detecting fake news. Within the topic layer, one of its tasks is the calculation of how similar two messages are to each other. We developed an improved version of an existing model, adapted it to our framework, and included calculating certain words and their positions as features and a novel embedding method using Cosine Similarity and POS (Parts of Speech) tags. We used the dataset from STSBenchmark and the correlation value to measure the quality of our model and performed a 50-fold evaluation over the validation data. Ultimately, our model has a better correlation value (median 0.673528) than the benchmarked model. Our main contribution is the delivery of a reliable, formal, and adaptable framework for identifying fake news; we also present a means of comparing tweets by their content, utilizing FastText models, cosine similarity, and other measures, making our contribution practical and effective.