AN INTELLIGENT FRAMEWORK FOR AUTOMATED MODERATION: CLASSIFYING MALICIOUS SOCIAL MEDIA MESSAGES USING CAUSAL GRAPH TRANSFORMERS WITH CONTRASTIVE LEARNING

Abstract

Social media sites such as Facebook, Twitter and Instagram have seen dramatic increases in user-generated content over the last decade, which has led to a significant rise in the prevalence of malicious content including hate speech, cyberbullying, offensive language and deliberate misinformation. State-of-the-art automated moderation systems, largely based on traditional machine learning models and transformer-based architectures find it difficult to characterize implicit contextual information or causal associations in textual data.

In a bid to overcome these limitations, we propose a new hybrid framework CGT-CL-BERT that closes the gap between causal graph representations and contextual embeddings based on transformer architecture by combining information obtained from them using contrastive learning. The method uses a pre-trained BERT encoder to get deep semantic features and relies on a graph-based complementary module to learn relational dependencies among textual entities. In addition, it also includes a contrastive learning objective to increase the robustness of representation by maximizing the inter-class distance and minimizing the intra-class distance.

The model is tested on benchmark datasets, specifically toxic comment classification, showing TQ-ETR outperforming baseline models using LSTM and a regular transformer architecture. We demonstrate improvements in accuracy, precision, recall andF1-score over 40 percent with respect to our earlier work while also being more effective in identifying implicit toxicity and toxicity requiring context to determine.

This framework demonstrates a system for smart social media moderation that can effectively and efficiently enable safe digital environments.

Future work will involve multilingual extensions and running in real-time large-scale content moderation systems.