Decoding the Rule Book: Extracting Hidden Moderation Criteria from Reddit Communities
This work addresses the challenge of making opaque moderation systems transparent for platform administrators and researchers, though it is incremental in applying interpretable methods to a specific domain.
The paper tackles the problem of extracting implicit content moderation criteria from online communities like Reddit by developing an interpretable architecture that represents criteria as lexical score tables. The approach successfully replicates neural model performance while revealing significant variations in how shared norms are enforced, uncovering undocumented patterns such as community-specific language tolerances and subcategories of toxic speech.
Effective content moderation systems require explicit classification criteria, yet online communities like subreddits often operate with diverse, implicit standards. This work introduces a novel approach to identify and extract these implicit criteria from historical moderation data using an interpretable architecture. We represent moderation criteria as score tables of lexical expressions associated with content removal, enabling systematic comparison across different communities. Our experiments demonstrate that these extracted lexical patterns effectively replicate the performance of neural moderation models while providing transparent insights into decision-making processes. The resulting criteria matrix reveals significant variations in how seemingly shared norms are actually enforced, uncovering previously undocumented moderation patterns including community-specific tolerances for language, features for topical restrictions, and underlying subcategories of the toxic speech classification.