IITR CodeBusters at SemEval-2022 Task 5: Misogyny Identification using Transformers

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 728 - 732 July 14-15, 2022 ©2022 Association for Computational Linguistics IITR CodeBusters at SemEval-2022 Task 5: Misogyny Identification using Transformers Gagan Sharma Indian Insitute of Technology Roorkee gagan_s@cs.iitr.ac.in Gajanan Sunil Gitte∗ Indian Insitute of Technology Roorkee gajanan_s@cs.iitr.ac.in Shlok Goyal∗ Indian Insitute of Technology Roorkee shlok_g@cs.iitr.ac.in Raksha Sharma Indian Insitute of Technology Roorkee raksha.sharma@cs.iitr.ac.in Abstract This paper presents our submission to task 5 ( Multimedia Automatic Misogyny Identifica- tion) of the SemEval 2022 competition. The purpose of the task is to identify given memes as misogynistic or not and further label the type of misogyny involved. In this paper, we present our approach based on language pro- cessing tools. We embed meme texts using GloVe embeddings and classify misogyny us- ing BERT model. Our model obtains an F1- score of 66.24% and 63.5% in misogyny classi- fication and misogyny labels, respectively. 1 Introduction SemEval 2022 task 5 (Fersini et al., 2022) is a sen- timent analysis task aimed at memes1, divided into two subtasks of increasing complexity. Subtask A is misogynous meme identification, i.e., whether a meme should be categorized either as misogy- nous or not misogynous; subtask B, on the other hand, is a multi-label classification problem, i.e., classifying what kind of misogyny is involved in the meme. A meme is usually intended to convey a sarcastic message, but in a short time, people have started to use them to deliver sexist messages in an online environment. The anonymity of the in- ternet tends to make them more aggressive. Such an environment amplifies the offline world’s sexual stereotyping and gender inequality. Meme sentiment analysis is a challenging task as often it relies on implicit themes or knowl- edge and trending news at the time of the cre- ation of the meme. Meme analysis has been of growing interest for the NLP community. Our approach also relies on various NLP-based tools and the code to implement the paper is available at https://github.com/gagansh7171/IITR- CodeBusters . ∗These two authors contributed equally. 1The term meme used in this task refers to an idea or a message conveyed via an image and embedded text. The rest of the paper is organized as follows. In Section 2 we briefly describe the details regard- ing the datasets used. In Section 3 we describe some recent related work conducted in the field of meme classification. In Section 4 we describe and define the models and baselines for the task. In sec- tion 5 we describe the low-level details including scoring parameters, pre-processing, libraries and hyper-parameters used for the experimental setup. In Section 6 we describe the results obtained and some insight of why the models are performing the way they performed. In section 7 we conclude the paper. 2 Background The dataset used for this task consists of a set of 10000 memes images whose text has already been extracted. Each meme has already been labeled as misogynous for subtask A and as shaming, stereo- type, …

Content truncated to 500 words