ModSandbox: Facilitating Online Community Moderation Through Error Prediction and Improvement of Automated Rules (2210.09569v1)

Abstract: Despite the common use of rule-based tools for online content moderation, human moderators still spend a lot of time monitoring them to ensure that they work as intended. Based on surveys and interviews with Reddit moderators who use AutoModerator, we identified the main challenges in reducing false positives and false negatives of automated rules: not being able to estimate the actual effect of a rule in advance and having difficulty figuring out how the rules should be updated. To address these issues, we built ModSandbox, a novel virtual sandbox system that detects possible false positives and false negatives of a rule to be improved and visualizes which part of the rule is causing issues. We conducted a user study with online content moderators, finding that ModSandbox can support quickly finding possible false positives and false negatives of automated rules and guide moderators to update those to reduce future errors.