TechSSN at SemEval-2022 Task 5: Multimedia Automatic Misogyny Identification using Deep Learning Models

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 571 - 574 July 14-15, 2022 ©2022 Association for Computational Linguistics TechSSN at SemEval-2022 Task 5: Multimedia Automatic Misogyny Identification using Deep Learning Models Rajalakshmi Sivanaiah, Angel Deborah S, Sakaya Milton R, Mirnalinee T T Department of Computer Science and Engineering Sri Sivasubramaniya Nadar College of Engineering Chennai - 603110, Tamil Nadu, India rajalakshmis@ssn.edu.in, angeldeborahs@ssn.edu.in, miltonrs@ssn.edu.in, mirnalineett@ssn.edu.in Abstract Research is progressing in a fast manner in the field of offensive, hate speech, abusive and sarcastic data. Tackling hate speech against women is urgent and really needed to give re- spect to the lady of our life. This paper de- scribes the system used for identifying misog- ynous content using images and text. The sys- tem developed by the team TECHSSN uses transformer models to detect the misogynous content from text and Convolutional Neural Network model for image data. Various mod- els like BERT, ALBERT, XLNET and CNN are explored and the combination of ALBERT and CNN as an ensemble model provides bet- ter results than the rest. This system was de- veloped for the task 5 of the competition, Se- mEval 2022. 1 Introduction In our society, women are facing lot of challenges in terms of education, employment, career and life. Eventhough women are better off now-a-days, there are not considered as equal to men in many situations. With the invent and rapid usage of so- cial media platforms, offensive images and texts conveying several forms of hate against women are transmitted and spread online in a fast manner. Although opportunities for women have been in- creased on the Internet, systematic inequality and discrimination offline is continued in online in the form of offensive contents through MEMEs. A meme is essentially an image characterized by a pictorial content with an overlaying text introduced by human on it. Most of the memes are created mainly for funni- ness, still it is used for ironic purpose too. The task 5 by Fersini et al. (2022) in SemEval 2022 mainly focused on identifying the misogynous memes us- ing textual and image contents. There are two sub- tasks in this. Subtask A is to categorize the memes as misogynous or not misogynous. Subtask B is to classify the misogynous memes into one of the categories like stereotype, shaming, objectification and violence. 2 Related Work The survey on various techniques used for Auto- matic Misogyny Identification (AMI) tasks hap- pened in EVALITA 2018 and IBERALEVAL 2018 were discussed in detail by Shushkevich and Cardiff (2019). Machine learning (SVM, Naive Bayes, Logistic Regression) and deep learning tech- niques (CNN, GRU, RNN, LSTM) are used and achieved an accuracy of 90% in IBERALEVAL task and 70% in EVALITA task. García-Díaz et al. (2021) uses sentiment analy- sis and social computing technologies with word embeddings and linguistic features for AMI that achieves better results than traditional machine learning algorithms. Hate speech against women is handled by Frenda et al. (2019). Pamungkas et al. (2020) created models using …

Content truncated to 500 words