FinD: Fine-grained discrepancy-based fake news detection enhanced by event abstract generation

In recent years, a plethora of studies have used fact-checking technology to detect fake news due to their ability of modeling the content discrepancy between the evidence from the fact corpus and the news. These models usually extract the evidence based on text similarity, which may not be relevant to the event’s facts. Moreover, they model coarse-grained content discrepancy that merely reflects the difference from a limited aspect and may be irrelevant to detection. Therefore, we propose a fine-grained discrepancy-based approach for fake news detection, named FinD. Specifically, we propose a dual-similarity calculation method to generate concise event abstract from semantic and literal aspects. In addition, we propose fine-grained discrepancy calculation methods to measure the content discrepancy. After that, we utilize sentences weights formed from the discrepancy to obtain effective news representation for the detection. We conduct extensive experiments on three datasets, validate the superiority of our framework FinD on fake news detection, and show the rationality of event abstract and fine-grained discrepancy of our framework.