Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Community Norms in the Spotlight: Enabling Task-Agnostic Unsupervised Pre-Training to Benefit Online Social Media

Liam Hebert, Lucas Kopp, Robin Cohen

TL;DR

This work addresses the data scarcity and interpretability challenges in analyzing online social dynamics by proposing norm-centric unsupervised pre-training for Discussion Transformers. The authors combine generative tasks that teach local structural and semantic norms with contrastive tasks that capture global community norms, enabling task-agnostic pre-training from unlabeled discussions. Preliminary experiments on Reddit-derived clusters show that norm representations separate by community attributes, supporting potential benefits for robust, interpretable moderation and sociotechnical governance. The approach offers a path toward data-efficient, explainable analysis of social behavior, aligning with AI for Social Good objectives by focusing on social context understanding rather than surface signals alone.

Abstract

Modelling the complex dynamics of online social platforms is critical for addressing challenges such as hate speech and misinformation. While Discussion Transformers, which model conversations as graph structures, have emerged as a promising architecture, their potential is severely constrained by reliance on high-quality, human-labelled datasets. In this paper, we advocate a paradigm shift from task-specific fine-tuning to unsupervised pretraining, grounded in an entirely novel consideration of community norms. We posit that this framework not only mitigates data scarcity but also enables interpretation of the social norms underlying the decisions made by such an AI system. Ultimately, we believe that this direction offers many opportunities for AI for Social Good.

Community Norms in the Spotlight: Enabling Task-Agnostic Unsupervised Pre-Training to Benefit Online Social Media

TL;DR

This work addresses the data scarcity and interpretability challenges in analyzing online social dynamics by proposing norm-centric unsupervised pre-training for Discussion Transformers. The authors combine generative tasks that teach local structural and semantic norms with contrastive tasks that capture global community norms, enabling task-agnostic pre-training from unlabeled discussions. Preliminary experiments on Reddit-derived clusters show that norm representations separate by community attributes, supporting potential benefits for robust, interpretable moderation and sociotechnical governance. The approach offers a path toward data-efficient, explainable analysis of social behavior, aligning with AI for Social Good objectives by focusing on social context understanding rather than surface signals alone.

Abstract

Modelling the complex dynamics of online social platforms is critical for addressing challenges such as hate speech and misinformation. While Discussion Transformers, which model conversations as graph structures, have emerged as a promising architecture, their potential is severely constrained by reliance on high-quality, human-labelled datasets. In this paper, we advocate a paradigm shift from task-specific fine-tuning to unsupervised pretraining, grounded in an entirely novel consideration of community norms. We posit that this framework not only mitigates data scarcity but also enables interpretation of the social norms underlying the decisions made by such an AI system. Ultimately, we believe that this direction offers many opportunities for AI for Social Good.
Paper Structure (8 sections, 1 figure)

This paper contains 8 sections, 1 figure.

Table of Contents

  1. Introduction
  2. A Framework for Norm-Centric Pre-Training
  3. Generative Pre-Training for Structural and Semantic Norms
  4. Contrastive Pre-Training for Global Community Norms
  5. Preliminary Experiments
  6. Discussion and Conclusion
  7. Acknowledgements
  8. Paper Checklist

Figures (1)

  • Figure 1: UMAP projection of the embedding space after applying community-centric contrastive pre-training on mDT hebert2024multi, where each dot is a discussion. We observe a clear separation between age-oriented discussions and political similarities, suggesting the potential to capture distinct behaviours.