Bridging social structure and discourse

This work addresses the challenge of Reddit communities exhibiting escalating polarization by developing a hybrid RAG pipeline that combines user-language features, fine-tuned language models, and continuous-time graph neural networks.

Abstract: Reddit communities were exhibiting escalating polarization, making proactive moderation impossible and increasing user churn. This work builds a hybrid RAG pipeline that retrieves user-language features via embeddings, fine-tunes LLAMA and BERT models for sentiment and persuasion-signal prediction, and trains a continuous-time GNN to model interaction dynamics.

Key Contributions:

• Hybrid RAG pipeline for social media analysis
• Fine-tuning of LLAMA and BERT models using Hugging Face Transformers
• Continuous-time GNN for modeling interaction dynamics
• 70% accuracy in forecasting community fragmentation
• Analysis of linguistic and interaction factors driving polarization

Results: Successfully forecast community fragmentation weeks ahead with 70% accuracy, enabling moderators to implement targeted interventions.

Recommended citation: Z. Guo, Z. Li, B. Tyler, X. Dong, and J. Pierrehumbert, “Bridging social structure and discourse,” Preparing to submit to the ACL Rolling Review (ARR), 2025.