Multimodal cyberbullying detection in Hinglish memes using a classroom framework based on large language models

The widespread use of social media has led to the emergence of harmful memes, which can perpetuate hate speech and cyberbullying. Detecting such harmful content is crucial for maintaining a safe online environment. Cyberbullying detection in code-mixed languages, particularly Hinglish (Hindi-English), poses significant challenges due to the complexity of multimodal content. This paper presents a novel approach for multimodal cyberbullying detection in Hinglish memes, leveraging the power of Large Language Models (LLMs) and a teacher-student learning based classroom framework. The proposed method involves generating explanations from both harmless and harmful perspectives using LLMs, followed by a final decision made by a smaller language model, mT5. This gives our model the ability to use multimodal explanations derived from both harmful and harmless arguments to execute dialectical reasoning over complex and implicit harm-indicative patterns. The model is trained on the MultiBully dataset, a benchmark dataset of Hinglish memes annotated with bully and non-bully labels. Experimental results demonstrate the effectiveness of the proposed approach, achieving an accuracy of 73.31% and an F1 score of 73.23%, outperforming several baselines. The paper concludes by discussing the implications of the findings and potential future directions for research in this area.