A Coarse-to-fine Cascaded Evidence-Distillation Neural Network for Explainable Fake News Detection

Proceedings of the 29th International Conference on Computational Linguistics, pages 2608–2621 October 12–17, 2022. 2608 A Coarse-to-fine Cascaded Evidence-Distillation Neural Network for Explainable Fake News Detection Zhiwei Yang1,2,3,5, Jing Ma2,∗, Hechang Chen3,5,∗, Hongzhan Lin2, Ziyang Luo2, Yi Chang3,4,5,∗ 1 College of Computer Science and Technology, Jilin University, Changchun, China 2 Department of Computer Science, Hong Kong Baptist University, Hong Kong, China 3 School of Artificial Intelligence, 4 International Center of Future Science, Jilin University, China 5 Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education yangzw18@mails.jlu.edu.cn, chenhc@jlu.edu.cn, {majing, cszyluo, cshzlin}@comp.hkbu.edu.hk, yichang@jlu.edu.cn Abstract Existing fake news detection methods aim to classify a piece of news as true or false and pro- vide veracity explanations, achieving remark- able performances. However, they often tailor automated solutions on manual fact-checked reports, suffering from limited news coverage and debunking delays. When a piece of news has not yet been fact-checked or debunked, cer- tain amounts of relevant raw reports are usually disseminated on various media outlets, con- taining the wisdom of crowds to verify the news claim and explain its verdict. In this pa- per, we propose a novel Coarse-to-fine Cas- caded Evidence-Distillation (CofCED) neural network for explainable fake news detection based on such raw reports, alleviating the de- pendency on fact-checked ones. Specifically, we first utilize a hierarchical encoder for web text representation, and then develop two cas- caded selectors to select the most explainable sentences for verdicts on top of the selected top-𝐴��� reports in a coarse-to-fine manner. Be- sides, we construct two explainable fake news datasets, which is publicly available. Experi- mental results demonstrate that our model sig- nificantly outperforms state-of-the-art detection baselines and generates high-quality explana- tions from diverse evaluation perspectives. 1 Introduction During the COVID-19 pandemic, almost 80% of consumers in the United States received fake news, which has caused confusion and undermined pub- lic health efforts1. The proliferation of fake news has increased the demand for automatic fake news detection (Guo et al., 2022). To further clarify and explain detection results, explainable fake news de- tection has gained more importance recently, aim- ing to classify the truthfulness of a piece of news and generate veracity explanations2 (Kotonya and ∗Corresponding authors. 1https://www.statista.com/topics/3251/fake-news 2Explanations and evidence are used interchangeably Figure 1: An example for veracity explanation genera- tion. The underlined explanations can be semantically inferred from some relevant sentences in the reports 𝐴���1 and 𝐴���𝐴���. “𝐴���” denotes the raw report. Toni, 2020a). However, existing methods have a limitation in detecting fake news timely as they heavily relied on debunked reports of investigated journalism. Thus, it is urgent to develop explain- able yet general methods to mitigate this issue. Many previous approaches detected fake news without any justifications (Wang, 2017; Ma et al., 2018). Recently, some explainable methods high- lighted salient words or phrases in relevant reports as explanations (Popat et al., 2018; Wu et al., 2021), which lack readable complete sentences. To al- leviate these issues, some methods aimed to ex- tract salient sentences from relevant reports via …

Content truncated to 500 words