I2C at SemEval-2022 Task 5: Identification of misogyny in internet memes

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 689 - 694 July 14-15, 2022 ©2022 Association for Computational Linguistics Abstract In this paper we present our approach and system description on Task 5 A in MAMI: Multimedia Automatic Misogyny Identification. In our experiments we compared several architectures based on deep learning algorithms with various other approaches to binary classification using Transformers, combined with a nudity image detection algorithm to provide better results. With this approach, we achieved a test accuracy of 0.665. 1 Introduction Misogyny is hatred or contempt for women. It is a form of sexism used to keep women at a lower social status than men, thus maintaining the societal roles of patriarchy. Misogyny has been widely practiced for thousands of years. It is reflected in art, literature, human societal structures, historical events, mythology, philosophy, and religion worldwide (Manne, 2017). The Internet represents for many an extension of our offline interactions, and seemingly mundane everyday practices (e.g. participating in social media) form a significant part of our everyday experiences. Unfortunately, it is too common to see examples of harassment towards women and marginalized groups online within these experiences and practices (Drakett et al., 2018). Women have a solid presence on the web, especially in picture-based web media like Twitter and Instagram: 78% of females utilize online media on numerous occasions each day, in contrast with 65% of men. A popular way of communicating via social media platforms are MEMEs. A meme is an image portrayed through pictorial content with overlaid text which is written a posteriori, with the fundamental objective of being entertaining and/or ironic. Even though most of memes are created with the goal of making amusing jokes, shortly after their standardization individuals began to use them to disseminate hate against women, leading to sexist and aggressive messages in internet environments that allow people to freely express sexism without the fear of retaliation. The detection of this disrespectful content is essential to eliminate it as soon as possible and stop spreading misogyny as a “joke”. In MAMI: Task 5, Track A (meme binary classification) (Fersini et al., 2022), participants must determine whether a meme (text + image) is misogynist or not. Meme sentiment-related tasks analysis is challenging, as memes are used created for various purposes, they are always evolving and often use sarcasm and humour. While misogyny and hate speech detection in text has been widely explored by the NLP community (Badjatiya et al., 2017). Its detection in images and text and how they correlate has not been explored in depth. In this field we used BERT (Devlin et al., 2018), RoBERTa (Liu et al., 2021) for training and classifying text, and nudenet tool for image classification (https://github.com/notAI- tech/NudeNet). Our results show that combining text and image classification results are slightly better than using only one of the two methods. The rest of the paper is organized as follows. Section 2 contains a briefly description of the dataset and its structure, Section 3 features …

Content truncated to 500 words