Multi-task detection of harmful content in code-mixed meme captions using large language models with zero-shot, few-shot, and fine-tuning approaches

Abstract

In today’s digital world, memes have become a common form of communication, shaping online conversations and reflecting social events. However, some memes can negatively impact people’s emotions, especially when they involve sensitive topics or mock certain groups or individuals. To address this issue, it is important to create a system that can identify and remove harmful memes before they cause further harm. Using Large Language Models for text classification in this system offers a promising approach, as these models are skilled at understanding complex language structures and recognizing patterns, including those in code-mixed language. This research focuses on evaluating how well different Large Language Models perform in identifying memes that promote cyberbullying. It covers tasks like cyberbullying detection, sentiment analysis, emotion recognition, sarcasm detection, and harmfulness evaluation. The results show significant improvements, with a 7.94% increase in accuracy for cyberbullying detection, a 2.68% improvement in harmfulness evaluation, and a 1.7% boost in sarcasm detection compared to previous top models. There is also a 1.07% improvement in emotion detection. These findings highlight the ability of Large Language Models to help tackle cyberbullying and create safer online spaces.