"You" Is Not Toxic: How LLM Labelers Rewrite the Web's Conversation Laws

📄 arXiv: 2610.23446v1 📥 PDF

作者: Liang Yue, Eric A. Vance

分类: cs.SI, cs.CL

发布日期: 2026-10-16

备注: 43 pages, 4 figures, 44 tables. Labels, prompts and code: https://github.com/liyu2195/you-is-not-toxic

💡 一句话要点

提出新方法以解决LLM标签者对网络对话的误判问题

🎯 匹配领域: 支柱九：具身大模型 (Embodied Foundation Models)

关键词: 大型语言模型 毒性标记 网络对话 人类评审 数据分析 内容审核 偏差纠正

📋 核心要点

  1. 现有的毒性标记方法主要依赖于Google的Perspective API，但其即将关闭，导致对网络对话的毒性测量面临挑战。
  2. 论文提出了一种新的方法Threadsafe，通过利用LLM标签者的评分来优化人类评审的选择和预测，提高标记的准确性。
  3. 实验结果显示，LLM标签者对包含“你”的评论存在显著的过度标记现象，而重写为第三人称后可减少约三分之二的误判。

📝 摘要（中文）

本研究探讨了大型语言模型（LLMs）在标记网络对话中的毒性时所存在的偏差，尤其是将第二人称的使用误判为毒性。通过分析125万条新闻评论，发现LLM标签者在标记时对包含“你”的评论存在过度标记现象。研究表明，重写为第三人称可以显著减少这种误判。此外，提出了一种新的方法Threadsafe，利用标签者的评分来选择人类评审的评论，从而提高标记的准确性。该方法在模拟审计中表现出优越的覆盖率，超越了现有的均匀抽样方法。

🔬 方法详解

问题定义：本研究旨在解决LLM标签者在标记网络对话毒性时的偏差，尤其是将第二人称“你”误判为毒性的问题。现有方法在处理长对话时未能有效识别这种偏差，导致标记不准确。

核心思路：论文提出通过重写评论中的第二人称为第三人称来减少误判，并引入Threadsafe方法，利用标签者的评分来优化人类评审的选择，从而提高标记的准确性。

技术框架：整体架构包括数据收集、LLM标记、重写机制和人类评审选择模块。首先收集评论数据，然后使用LLM进行初步标记，接着通过重写机制调整标记，最后利用Threadsafe方法选择人类评审的评论。

关键创新：最重要的技术创新在于提出了Threadsafe方法，通过结合LLM标签者的评分与人类评审，显著提高了标记的准确性，克服了传统方法的局限性。

关键设计：在实验中，重写机制的参数设置和人类评审的选择策略是关键设计，确保了在不同的评论长度和内容下，能够有效减少误判并提高标记的可靠性。实验中还使用了线性法则来估计标记的偏差。

🖼️ 关键图片

img_0
img_1
img_2

📊 实验亮点

实验结果显示，Claude Haiku对不含“你”的评论标记率为12%，而含有四个或更多“你”的评论标记率高达51%。通过重写为第三人称，误判率减少约66%。Threadsafe方法在模拟审计中表现出优越的覆盖率，标记准确性显著提高。

🎯 应用场景

该研究的潜在应用领域包括社交媒体平台、在线评论系统及任何需要自动化内容审核的场景。通过提高毒性标记的准确性，可以有效改善网络环境，减少有害内容的传播，促进更健康的在线交流氛围。未来，该方法可扩展至其他语言和文化背景下的对话分析。

📄 摘要（原文）

What the Web community believes about toxic conversations, such as that toxicity rises in longer and deeper threads, was measured mostly with Google's Perspective API, which is shutting down; large language models (LLMs) will make the next measurements. We find that LLM labelers mistake second-person address for toxicity, and that this one bias manufactures laws that human labels do not show. On 1.25 million news comments with crowd labels, Claude Haiku flags 12% of the non-toxic comments that contain no "you" and 51% of those with four or more; a single crowd rater flags 6% and 10%. Deeper replies address someone more often, so under Haiku toxicity rises with reply depth (+0.93 percentage points per level), while under the crowd's labels it slightly declines (-0.12). Rewriting "you" in the third person removes two thirds of the false rise; flagging only as many comments as the crowd does not remove it, and showing Haiku the parent comment enlarges it. All seven LLMs we test over-flag "you", and at the crowd's flag rate, without the parent, any false rise of the open ones vanishes at a fixed number of second-person words; Haiku also over-flags "you" on Reddit, Gab and Wikipedia talk pages. Agreement with human labels does not reveal the bias: we prove that a fixed Cohen's kappa leaves a law nearly unconstrained. We therefore certify the law instead of the labeler. Threadsafe uses the labeler's scores to choose which comments human raters label and to predict the others, and corrects the predictions with the human labels; linear laws are estimated without bias whatever the labeler does. In simulated audits its intervals for the laws keep nominal coverage from 250 labels on, where uniform-sample methods such as PPI++ fall short, and at 1,000 to 8,000 labels are 0.75 times as wide as those of design-based supervised learning (DSL); a thousand human labels rule out Haiku's rising law.