LiteDetective🕵️

Abstract

Harmful content detection is a critical task for any social media platform, as the presence of misinformation, age, gender, and racial discrim- ination can lead to a reduction in active users. This paper introduces a novel approach that leverages large language models (LLMs) to an- alyze specific social media data and generate training data, combined with a BERT-based Dy- namic TextCNN architecture. We first crawl potential harmful comments from targeted com- munities (e.g., "ShunBa"). These comments are then subjected to random filtering and clus- tering using a smaller LLM to generate policy- guided seed examples. Next, we employ a large LLM (Qwen-3) for context-aware and context- free data augmentation. Finally, we integrate BERT embeddings with a Dynamic TextCNN classifier on our custom dataset.