基于加权微调BERT的稀疏RNN虚假新闻检测方法

摘要

虚假新闻指以图像、文章或视频形式传播、伪装成真实新闻以试图操纵公众观点的错误信息或不实报道。然而，由于语言风格、语境与信息来源的多样性，现有检测系统难以捕捉虚假新闻的多样特征，导致识别不准确。为此，本文提出一种用于虚假新闻检测的深度学习（DL）方法——面向Transformer的加权微调双向编码器表示与稀疏循环神经网络相结合的模型（WFT-BERT-SRNN）。首先，从Buzzfeed、PolitiFact、Fakeddit与Weibo数据集获取数据以评估WFT-BERT-SRNN；随后通过停用词去除、分词与词干提取进行预处理，剔除无用短语或词汇；接着用WFT-BERT提取特征；最后用SRNN检测并把新闻分类为真实或虚假。本文把WFT-BERT-SRNN与深度神经网络虚假新闻检测（DeepFake）、联合学习BERT、面向虚假新闻检测的多EDU结构（EDU4FD）、基于图像描述的方法以及细粒度多模态融合网络（FMFN）等现有技术进行了比较。相较上述方法，WFT-BERT-SRNN在Buzzfeed、PolitiFact、Fakeddit与Weibo数据集上分别取得0.9847、0.9724、0.9624与0.9725的更优准确率。

Abstract

Fake news refers to misinformation or false reports shared in the form of images, articles, or videos that are disguised as real news to try to manipulate people's opinions. However, detection systems fail to capture diverse features of fake news due to variability in linguistic styles, contexts, and sources, which lead to inaccurate identification. For this purpose, a weighted fine-tuned-bidirectional encoder representation for transformer-based sparse recurrent neural network (WFT-BERT-SRNN) is proposed for fake news detection using deep learning (DL). Initially, data is acquired from Buzzfeed, PolitiFact, Fakeddit, and Weibo datasets to evaluate WFT-BERT-SRNN. Pre-processing is established using stopword removal, tokenization, and stemming to eliminate unwanted phrases or words. Then, WFT-BERT is employed to extract features. Finally, SRNN is employed to detect and classify fake news as real or fake. Existing techniques like deep neural networks for fake news detection (DeepFake), BERT with joint learning, multi-EDU structure for fake news detection (EDU4FD), image caption-based technique, and fine-grained multimodal fusion network (FMFN) are compared with WFT-BERT-SRNN. The WFT-BERT-SRNN achieves a better accuracy of 0.9847, 0.9724, 0.9624, and 0.9725 for Buzzfeed, PolitiFact, Fakeddit, and Weibo datasets compared to existing techniques like DeepFake, BERT-joint framework, EDU4FD, image caption-based technique, and FMFN.