Improving the Robustness of Summarization Systems with Dual Augmentation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 627–638 July 9-14, 2023 ©2023 Association for Computational Linguistics Causal Intervention and Counterfactual Reasoning for Multi-modal Fake News Detection Ziwei Chen1 Linmei Hu2∗ Weixin Li3 Yingxia Shao1 Liqiang Nie4 1Beijing University of Posts and Telecommunications 2 Beijing Institute of Technology 3Beihang University 4 Harbin Institute of Technology (Shenzhen) {chen_zw,shaoyx}@bupt.edu.cn hulinmei@bit.edu.cn weixinli@buaa.edu.cn nieliqiang@gmail.com Abstract Due to the rapid upgrade of social platforms, most of today’s fake news is published and spread in a multi-modal form. Most existing multi-modal fake news detection methods ne- glect the fact that some label-specific features learned from the training set cannot generalize well to the testing set, thus inevitably suffering from the harm caused by the latent data bias. In this paper, we analyze and identify the psy- cholinguistic bias in the text and the bias of inferring news label based on only image fea- tures. We mitigate these biases from a causality perspective and propose a Causal intervention and Counterfactual reasoning based Debiasing framework (CCD) for multi-modal fake news detection. To achieve our goal, we first utilize causal intervention to remove the psycholin- guistic bias which introduces the spurious cor- relations between text features and news label. And then, we apply counterfactual reasoning by imagining a counterfactual world where each news has only image features for estimating the direct effect of the image. Therefore we can eliminate the image-only bias by deducting the direct effect of the image from the total ef- fect on labels. Extensive experiments on two real-world benchmark datasets demonstrate the effectiveness of our framework for improving multi-modal fake news detection. 1 Introduction Fake news quietly sneaks into people’s daily life, mixed with massive information, causing serious impact and harm to society. Fake news often uti- lizes multimedia information such as text and im- ages to mislead readers, spreading and expanding its influence. Thus, it is crucial and urgent to find a way to discern multi-modal fake news. Today, most existing methods train on known fake news instances expecting to capture the label- specific features for judging the authenticity of un- seen news (Singhal et al., 2020; Wu et al., 2021; ∗Corresponding author Anx. Anger Neg.emo Pos.emo Sad Insight Cause. Discrep. Tentat. Certain Differ. Feel Hear See Swear Netspeak Assent Nonflu. Category 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 Frequency Fake Real Figure 1: The word frequency distributions of psycho- logical categories on different labels (Twitter dataset). Qian et al., 2021b; Qi et al., 2021). However, such label-specific features may expose the models to hidden data bias when confronted with unseen fake news samples (Wang et al., 2018; Cheng et al., 2021; Zhu et al., 2022). To address the problem, we investigate the biases underlying the multi-modal fake news detection data and identify the psycholin- guistic bias in the text and the bias of inferring news label based on image features only (i.e. image-only bias). These biases could lead to spurious correla- tions between …

Content truncated to 500 words