Preventing rumor spread with deep learning

Abstract

The spreading of false rumors through online media is a serious social problem. The current process of fact-checking mostly relies on the responses of crowds or journalists to perform investigations, which can be performed after a rumor has widely spread. This study proposes a bi-directional encoder representations from transformers (BERT)-based model that only takes a claim sentence of a rumor, which can be used to identify false rumors before it goes viral. By evaluating the proposed model with the rumor dataset that is collected from Snopes and Politifact, this study demonstrates the effectiveness of the model that can accurately identify false rumors with only a given claim text. We also reveal that the performance of the models that are trained from the rumors in specific categories (e.g., business, politics) can be improved by transfer learning. Transfer learning uses the model parameters that are trained from a category as an initial state of the model for another category. Our analysis shows that a pre-trained model from a category that deals with a broad range of topics (e.g., fauxtography) is a useful source that can be transferred to other categories (e.g., entertainment). We believe that the proposed model can help mitigate potential social risks such as social turmoil or monetary chaos that are caused by false rumors, as the rumors can be detected before they go viral.