Cyberbullying Detection: An Innovative MultiTask Learning Framework for Analyzing Harmful Memes
Fowzia Binta Faruk Ananna, Subarno Ranjan Barua, Mohammed Nazim Uddin · 2024
Cyberbullying encompasses bullying behaviors on digital platforms, including spreading false information, sharing embarrassing content, sending threats, or impersonating others to send harmful messages. This study delves into the growing problem of cyberbullying, focusing on the spread of damaging memes on social media. The research explores the potential of hybrid models for detecting cyberbullying memes. This hybrid model combines elements from some existing multitask learning frameworks: MultiAchiever and CentralNet, DMTL and MTEIR, MTEIR and MTLAttention, DMTL and MTLAttention, as well as DMTL and MoE. By leveraging these frameworks, the study investigates the effectiveness of these hybrid approaches in analyzing memes for characteristics associated with cyberbullying, including harmfulness score, sentiment analysis, sarcasm detection, and target identification. The research explores the use of advanced pre-trained models for both image and text feature extraction. A comprehensive dataset consisting of 6,006 memes with textual annotations is used to train and evaluate the models. This nuanced analysis aims to contribute to a more sophisticated understanding of online memes and enhance content analysis capabilities, ultimately fostering safer digital spaces. Furthermore, among all the hybrid models, The MultiAchiever-CentralNet algorithm, in conjunction with ResNet50 and DistilBERT for image and text feature extraction, outperforms existing models with a notable accuracy of 68.64% with f1_score of 65.54%.