OCR-Based Multi-class Classification of Hate Speech in Images

The depersonalization and anonymity afforded by ubiquitous social media platforms facilitate open discourse, yet also create a potential avenue for hate speech dissemination. The rising incidence of hate speech on these platforms necessitates vigilant monitoring and intervention. However, the sheer volume of user-generated content renders manual oversight infeasible. Technological solutions must be developed to efficiently identify and mitigate hate speech, striking a balance between maintaining open expression and safeguarding against harmful content. Additionally, when utilizing traditional machine learning methodologies as prediction methods, the language being utilized and the length of the messages provide a barrier. In this study, a Genetic Programming (GP) model for identifying hate speech is presented, where each chromosome acts as a classifier with a universal sentence encoder feature. The performance of the GP model was enhanced by enriching the offspring pool with alternative solutions using a unique mutation strategy that only modifies the feature values in addition to the conventional one-point mutation technique. For the six categories of hate-type hate speech text datasets, the suggested GP model beat all cutting-edge solutions. In contrast to the machine learning models such as Random Forest, Decision Tree, and Naive Bayes which gave the following accuracy of 79.25%, 77.88% and 77.33% whereas GP model outperformed with an accuracy of 97%. The following evaluation matrices are considered as precision, recall training, and testing accuracy.