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

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.

💡 Research Summary

The paper tackles a fundamental bottleneck in applying Discussion Transformers (DTs) to online social media: the heavy reliance on scarce, high‑quality, human‑annotated datasets. While DTs excel at modeling the hierarchical, branching structure of conversations as graphs, existing labeled corpora are limited, costly to produce, and often shrink over time as platforms remove the very content needed for training. Moreover, task‑specific fine‑tuning typically focuses on a single objective (e.g., hate‑speech detection) and fails to capture the broader spectrum of community norms that shape how language is used, who is heard, and what is considered acceptable within a given online group.

To overcome these limitations, the authors propose a paradigm shift from task‑specific fine‑tuning to task‑agnostic, unsupervised pre‑training that is explicitly grounded in the notion of community norms. The central hypothesis is that, by leveraging massive amounts of unlabeled discussion data, a DT can first learn the foundational principles of conversational context—both structural (who replies to whom) and semantic (what content is appropriate)—before being exposed to any downstream, potentially biased, labeled tasks.

Pre‑training Design

The framework distinguishes two families of pre‑training objectives, each adapted to the graph‑structured nature of discussions:

1. Generative Objectives – aimed at learning local conversational rules.

   • Edge‑level “is‑a‑reply” classification: Randomly mask reply‑to edges in the discussion graph and ask the model to predict whether two comments should be linked. Because Graph Transformers rely on learned positional encodings rather than built‑in adjacency, this task forces the model to embed structural conventions (e.g., typical reply depths, turn‑taking patterns) into its representations.
   • Node‑level “semantic norm” reconstruction: Mask an entire comment’s multimodal embedding (text, images, etc.) and require the model to reconstruct it from surrounding context. This mirrors masked language modeling but operates on a dense, high‑information node rather than a token, encouraging the model to internalize what kinds of utterances a community expects in a given conversational slot.

2. Contrastive Objectives – aimed at learning global community‑level signatures.

   • Branch‑sampling contrast: Sample two distinct sub‑trees that emanate from the same root post, treat them as a positive pair, and contrast them against branches from unrelated discussions. This teaches the model to capture the shared “discussion norms” that bind a single thread together (e.g., consistent tone, topic drift).
   • Community‑norm alignment: Sample entire discussion trees from the same subreddit (or a cluster of subreddits with high user‑overlap) as positives, and trees from different communities as negatives. By pulling together embeddings of the same community, the model learns a latent “social fingerprint” that encodes dialect, topical preferences, and implicit etiquette.

Both families are implemented with standard losses (cross‑entropy for the generative tasks, InfoNCE for the contrastive tasks) but the authors stress the need for discussion‑aware loss functions that can handle multi‑answer scenarios (e.g., many replies could be appropriate to a given comment) and the high‑dimensional nature of comment embeddings.

Preliminary Experiments

The authors conduct an initial study using the multidiscussion Transformer (mDT) as a backbone. They apply only the community‑level contrastive pre‑training on a Reddit corpus:

• Data: 8,000 high‑scoring discussions drawn from subreddit clusters representing three social dimensions—politics (left vs. right), age (young vs. old), and gender (male vs. female). The clusters are derived from the subreddit2vec methodology of Waller & Anderson (2021), which groups subreddits by user‑membership similarity.
• Procedure: For each discussion, comment embeddings are averaged to obtain a single discussion vector; positive pairs are formed from discussions within the same cluster, negatives from other clusters. An InfoNCE loss is optimized.
• Results: A UMAP projection of the resulting embeddings shows clear separation along political and age axes, indicating that the contrastive pre‑training successfully captures high‑level community norms. The authors note that adding the generative objectives would likely sharpen intra‑cluster cohesion and improve downstream task performance (e.g., hate‑speech detection).

Critical Discussion

The paper acknowledges several open challenges:

• Negative Transfer: As observed in GNN pre‑training, misaligned objectives can cause the model to learn spurious patterns that hurt downstream performance. The authors argue that this risk is amplified when the pre‑training task does not faithfully reflect the social phenomenon of interest.
• Loss Design for Multi‑Answer Scenarios: The “is‑a‑reply” task must tolerate multiple correct replies; a simple binary cross‑entropy would penalize valid alternatives. The authors suggest soft‑labeling or label‑smoothing strategies but leave concrete implementations for future work.
• Reconstruction Difficulty: Masked comment reconstruction is far more complex than masked token prediction because a comment can convey virtually unlimited semantic content. Designing a loss that balances textual fidelity with computational tractability remains an open research problem.

Interpretability & Social Good

A major contribution is the proposal to pair norm‑centric pre‑training with explainable AI (XAI) techniques. By projecting new discussions into the learned embedding space, one can assign them to the nearest “behavioral prototype” (e.g., a left‑leaning political norm). Mechanistic interpretability methods—such as probing attention heads, analyzing activation patterns of specific MLP neurons, or extracting circuit‑level subgraphs—could reveal which components are responsible for detecting in‑group jokes, sarcasm, or harassment. This transparency is crucial for auditing bias, ensuring equitable treatment across communities, and ultimately building AI systems that promote safe, inclusive online discourse.

Conclusion and Outlook

The paper charts a compelling research agenda: replace scarce supervised fine‑tuning with a two‑pronged unsupervised pre‑training regime that first teaches Discussion Transformers the structural and semantic norms of conversation (generative tasks) and then the global cultural fingerprints of online communities (contrastive tasks). Preliminary results on Reddit demonstrate that even a single contrastive objective can induce meaningful community‑level clustering. Future work must (1) implement and evaluate the generative objectives, (2) refine loss functions to handle multi‑answer and high‑dimensional reconstruction challenges, (3) systematically study negative transfer, and (4) integrate XAI pipelines for norm‑level interpretability. If successful, this line of research could dramatically reduce data bottlenecks, improve model robustness across diverse platforms, and provide the transparency needed for responsible AI deployment in the fight against hate speech, misinformation, and other online harms.