Edited Media Understanding Frames: Reasoning About the Intent and Implications of Visual Misinformation

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2026–2039 August 1–6, 2021. ©2021 Association for Computational Linguistics 2026 Edited Media Understanding Frames: Reasoning About the Intents and Implications of Visual Disinformation Jeff Da♠ Maxwell Forbes♠♥ Rowan Zellers♠♥ Anthony Zheng♣ Jena D. Hwang♠ Antoine Bosselut♠♦ Yejin Choi♠♥ ♠Allen Institute for Artificial Intelligence ♣University of Michigan ♥Paul G. Allen School of Computer Science & Engineering, University of Washington ♦Stanford University jeffda.com/edited-media-understanding Abstract Understanding manipulated media, from auto- matically generated ‘deepfakes’ to manually edited ones, raises novel research challenges. Because the vast majority of edited or manipu- lated images are benign, such as photoshopped images for visual enhancements, the key chal- lenge is to understand the complex layers of underlying intents of media edits and their im- plications with respect to disinformation. In this paper, we study Edited Media Under- standing Frames, a new conceptual formal- ism to understand visual media manipulation as structured annotations with respect to the in- tents, emotional reactions, effects on individu- als, and the overall implications of disinforma- tion. We introduce a dataset for our task, EMU, with 56k question-answer pairs written in rich natural language. We evaluate a wide vari- ety of vision-and-language models for our task, and introduce a new model PELICAN, which builds upon recent progress in pretrained mul- timodal representations. Our model obtains promising results on our dataset, with humans rating its answers as accurate 48.2% of the time. At the same time, there is still much work to be done – and we provide analysis that highlights areas for further progress. 1 Introduction The modern ubiquity of powerful image-editing software has led to a variety of new disinforma- tion threats. From AI-enabled “deepfakes” to low- skilled “cheapfakes,” attackers edit media to en- gage in a variety of harmful behaviors, such as spreading disinformation, creating revenge porn, and committing fraud (Paris and Donovan, 2019; Chesney and Citron, 2019; Kietzmann et al., 2020, c.f.). Accordingly, we argue that it is important to develop systems to help spot harmful manipu- lated media. The rapid growth and virality of social I want to suggest that subject2 has the support of subject1. subject1 subject2 I feel mad, because we’re definitely not friends... How subject1 might feel It seems like subject2 and subject1 are allies! original edited background changed person introduced horizontally flipped Previous work: Understanding the structure of edits (Jhamtani et al., 2018; Tan et al., 2019) This paper: Understanding the intent and implications of edits Why did the editor create this? How the community might respond frame: emotion frame: disinformation frame: intent Figure 1: Edited Media Understanding Frames. Given a manipulated image and its source, a model must generate natural language answers to a set of open-ended questions. Our questions test the under- standing of the what and why behind important changes in the image – like that subject1 appears to be on good terms with subject2. media requires as such, especially …

Content truncated to 500 words