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.
翻译:尽管通常使用基于规则的工具来调和在线内容,但人类主持人仍然花很多时间来监测这些工具,以确保它们如预期的那样发挥作用。 根据对使用自动调控器的Reddit主持人的调查和访谈,我们查明了减少自动规则的虚假正面和虚假负面方面的主要挑战:无法事先估计规则的实际效果,也难以确定规则应如何更新。为了解决这些问题,我们建立了ModSandbox,这是一个新颖的虚拟沙箱系统,它能探测出规则中可能存在的虚假正面和虚假反面,并可以想象规则中哪些部分正在引起问题。 我们与在线内容管理员进行了用户研究,发现ModSandbox可以支持快速发现可能存在的虚假正面和虚假反面自动规则,并指导主持人更新规则,以减少未来错误。