Detecting Harmful Memes and Their Targets

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2783–2796 August 1–6, 2021. ©2021 Association for Computational Linguistics 2783 Detecting Harmful Memes and Their Targets Shraman Pramanik1, Dimiter Dimitrov2, Rituparna Mukherjee1, Shivam Sharma1,3, Md. Shad Akhtar1, Preslav Nakov4, Tanmoy Chakraborty1 1Indraprastha Institute of Information Technology - Delhi, India 2Sofia University, Bulgaria 3Wipro AI Labs, India 4Qatar Computing Research Institute, HBKU, Doha, Qatar {shramanp, shivams, shad.akhtar, tanmoy}@iiitd.ac.in mitko.bg.ss@gmail.com, ritumukherjee23@gmail.com, pnakov@hbku.edu.qa Abstract Among the various modes of communication in social media, the use of Internet memes has emerged as a powerful means to convey politi- cal, psychological, and socio-cultural opinions. Although memes are typically humorous in na- ture, recent days have witnessed a proliferation of harmful memes targeted to abuse various social entities. As most harmful memes are highly satirical and abstruse without appropri- ate contexts, off-the-shelf multimodal models may not be adequate to understand their under- lying semantics. In this work, we propose two novel problem formulations: detecting harm- ful memes and the social entities that these harmful memes target. To this end, we present HarMeme, the first benchmark dataset, con- taining 3, 544 memes related to COVID-19. Each meme went through a rigorous two-stage annotation process. In the first stage, we la- beled a meme as very harmful, partially harm- ful, or harmless; in the second stage, we fur- ther annotated the type of target(s) that each harmful meme points to: individual, orga- nization, community, or society/general pub- lic/other. The evaluation results using ten uni- modal and multimodal models highlight the importance of using multimodal signals for both tasks. We further discuss the limitations of these models and we argue that more re- search is needed to address these problems. 1 Introduction The growing popularity of social media has led to the rise of multimodal content as a way to express ideas and emotions. As a result, a brand new type of message was born: meme. A meme is typically formed by an image and a short piece of text on top of it, embedded as part of the image. Memes are typically innocent and designed to look funny. WARNING: This paper contains meme examples and words that are offensive in nature. Over time, memes started being used for harm- ful purposes in the context of contemporary politi- cal and socio-cultural events, targeting individuals, groups, businesses, and society as a whole. At the same time, their multimodal nature and often camouflaged semantics make their analysis highly challenging (Sabat et al., 2019). Meme analysis. The proliferation of memes online and their increasing importance have led to a growing body of research on meme analy- sis (Sharma et al., 2020a; Reis et al., 2020; Pra- manick et al., 2021). It has also been shown that off-the-shelf multimodal tools may be inadequate to unfold the underlying semantics of a meme as (i) memes are often context-dependent, (ii) the vi- sual and the textual content are often uncorrelated, and (iii) meme images are mostly morphed, and the embedded text is sometimes hard to …

Content truncated to 500 words