Transformers at SemEval-2022 Task 5: A Feature Extraction based Approach for Misogynous Meme Detection

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 550 - 554 July 14-15, 2022 ©2022 Association for Computational Linguistics Transformers at SemEval-2022 Task 5: A Feature Extraction based Approach for Misogynous Meme Detection Shankar Mahadevan1, Sean Benhur2, Roshan Nayak3, Malliga Subramanian4, Kogilavani Shanmugavadivel4, Kanchana Sivanraju2, Bharathi Raja Chakravarthi5 1 Thiagarajar College of Engineering, India 2PSG College of Arts and Science, India 3B.M.S College of Engineering, India 4Kongu Engineering College, India 5Insight SFI Research Centre for Data Analytics,National University of Ireland, Galway shankarmahadevan12901@gmail.com1, seanbenhur@gmail.com2, roshannayak610@gmail.com3 mallinishanth72@gmail.com4, kogilanvani.sv@gmail.com4, kanachana@psgcas.ac.in2, bharathi.raja@insight-centre.org5 Abstract Social media is an idea created to make the world smaller and more connected. Recently, it has become a hub of fake news and sexist memes that target women. Social Media should ensure proper women’s safety and equality. Fil- tering such information from social media is of paramount importance to achieving this goal. In this paper, we describe the system developed by our team for SemEval-2022 Task 5: Multi- media Automatic Misogyny Identification. We propose a multimodal training methodology that achieves good performance on both the sub- tasks, ranking 4th for Subtask A (0.718 macro F1-score) and 9th for Subtask B (0.695 macro F1-score) while exceeding the baseline results by good margins. The code will be available here1 1 Introduction With the rising usage of social media and the Inter- net, it is tougher to establish an inclusive and wel- coming community among users. Offensive speech, hate speech, and targeted insult have been increas- ing among users, disturbing everyone. With the rising utilization of the Internet in a pandemic, hate speech prevalence on the Internet is also increased. Online hate speech with targeted discrimination also creates threats and crimes offline. Among these, online misogyny or sexist com- ments have been increasing among women (Salter et al., 2018), which includes name-calling, sexual threats, shaming. This emphasizes the need for 1https://github.com/shankrmahadevan/ semevalmami2022 specialized automatic misogyny detection in on- line platforms. Platforms such as Twitter 2 and Facebook already have policies for banning hate- ful content. However, these systems are primarily through manual methods and might not scale well for large users and multimodal content. Moreover, hate speech is also prevalent in multimodal form since most social media platforms support Images, text, audio and video content. These memes have been popular among users to express their opinions since people express information through memes, GIFs, and videos. But, unfortunately, this also causes the rise of multimodal hate, and offensive content, which is disturbing to users (Suryawan- shi et al., 2020). This includes misogynistic posts, which are targeted towards women. Previous works on Misogyny detection have been primarily focusing on one modality, which is text (Pamungkas et al., 2020). Misogyny detec- tion in text falls under an area of text classification. Text classification methods such as BERT, LSTM, Naive Bayes have been used to detect misogynistic comments. In this work, we have used the provided data, which contains both images of the memes and the extracted OCR text from the memes. This paper …

Content truncated to 500 words