内容审核术语的通俗解释

简短回答。 内容审核是决定什么内容可以在共享空间发布，什么内容不能发布的工作。它的词汇部分来自大型平台的信任与安全团队，部分来自机器学习。以下是志愿者管理员实际会遇到的术语，每个术语都附有关于其在实践中重要性的说明。

Moderation has acquired a vocabulary of its own, partly from the trust-and-safety teams of large platforms and partly from machine learning. Much of it is useful once translated. This glossary explains the terms a volunteer admin is likely to meet, in the order of the alphabet, with a note on why each matters in practice.

A to C

Appeal. A request to have a moderation decision looked at again. Even a small page benefits from having a way to ask, if only an email address, because some decisions will be wrong.

Astroturfing. Coordinated activity designed to look like spontaneous public feeling. On a confession page it takes the form of many submissions pushing the same line, written to appear independent. See brigading.

Automated moderation. The use of software to screen content with no human involved, or before a human sees it. It is useful for sorting and for catching the obvious, and unreliable as a final judge. Explained in AI 内容审核如何运作.

Blocklist. A list of words or phrases that cause a post or comment to be hidden or held. Quick to set up and blunt in effect: it catches innocent uses and misses creative spellings.

Brigading. A group organising in one place in order to descend on another — submitting, commenting or reporting en masse. The volume is meant to look like consensus.

CAPTCHA. A test intended to tell people from programs. Effective against bots; adds friction and can raise accessibility and privacy concerns.

Classifier. A model trained to give a score for how strongly content matches a category such as "harassment" or "sexual". It produces a number, not a decision.

Community guidelines. The published rules of a space. The term is used both for a platform's rules, which you must follow, and for your own page's rules, which you write.

Content warning. A short notice ahead of material some readers may find distressing, so that they can choose whether to read on. Helpful when used sparingly; ignored when placed on everything.

Context collapse. What happens when words written for one audience are read by another. A joke between friends, published to five thousand strangers, is no longer the same statement.

CSAM. Child sexual abuse material. Illegal to possess or distribute. It must be reported and never forwarded, even for the purpose of reporting it.

D to H

Decision log. A record of non-obvious moderation decisions and the reasons for them. The cheapest tool a team has for staying consistent.

Defamation. A false statement, communicated to others, that damages a person's reputation. In writing or other permanent form it is libel. See 诽谤与告白页面.

Doxxing. Publishing someone's identifying or private information — address, workplace, phone number, identity documents — in order to expose them to harassment. A criminal offence in Malaysia since 2025 when done with that intent or effect.

Escalation. Passing a decision to someone with more authority or experience, or requiring a second person to agree. A team needs a clear list of what triggers it.

Evasion. Deliberate attempts to get round a filter: swapping letters for numbers, inserting spaces, using lookalike characters, or placing text inside an image.

False negative. Harmful content that a check failed to flag. The mistake you learn about from a complaint.

False positive. Harmless content that a check flagged wrongly. The mistake you learn about from an annoyed submitter.

Flag. A marker that content needs attention. It may be raised by software or by a user's report. A flag is a question, not an answer.

Hash matching. Comparing a file's digital fingerprint with a database of known harmful files. Large platforms use it to detect known abuse imagery without anyone having to view it.

Hate speech. Attacks on people on the basis of characteristics such as race, religion, nationality, gender, sexuality or disability. Definitions differ between platforms and between countries.

Human in the loop. A system design in which software makes recommendations and a person makes the final decision. The sensible default for anything with consequences.

I to N

Identifiability. Whether a person could be recognised from a piece of content by people who know them. It does not require a name.

Impersonation. Pretending to be a specific person or organisation. On a confession page this includes submissions written as though by someone else in order to embarrass them.

Jigsaw identification. Identifying a person by putting together several details, each harmless on its own. Course, year and society role, taken together, may point to exactly one person.

Large language model (LLM). The technology behind chat assistants. It can be given instructions in ordinary language and asked to judge content against them. Flexible, and capable of being confidently wrong.

Metadata. Information about a file as distinct from its content: when a photo was taken, on what device, and sometimes where. It can identify a person even when the image itself does not.

Moderation queue. The list of items waiting for a decision. The age of the oldest item in it is the best single measure of a team's health.

Named-person detection. Automated flagging of submissions that appear to contain a personal name, so that a human can consider whether someone is being identified. It detects that a name is present. It cannot judge whether publishing it is harmful.

Non-consensual intimate imagery. Sexual images shared without the consent of the person shown, including images that have been fabricated. Sometimes called "revenge porn", a term many people avoid because it implies that the victim did something to provoke it.

NSFW. "Not safe for work" — shorthand for sexually explicit or graphic content, and the usual label for classifiers that detect it.

P to S

Pile-on. A large number of people directing criticism or abuse at one person in a short time. Each participant sends a single message; the target receives hundreds.

Pre-moderation. Reviewing content before it is published. This is how confession pages work, and it is why the page's admins are so clearly responsible for what appears.

Post-moderation. Publishing first and reviewing afterwards, usually when something is reported. This is how comment sections normally work.

Proof of work. A small computational task that a browser must complete before a form will accept a submission. It is imperceptible to one person and costly to someone sending thousands. It limits abuse without identifying anybody.

Rate limiting. Restricting how often one source can perform an action within a period of time. The main defence against flooding.

Recall. In machine learning, the proportion of all harmful items that a check manages to catch. High recall means that little gets through; it usually comes with more false positives. (Separately, some submission tools use "recall" to mean a submitter withdrawing their own post.)

Precision. The proportion of flagged items that really were harmful. High precision means that flags can be trusted; it usually comes with more misses.

Republication. Repeating a statement that someone else originally made. In defamation law, the person who repeats a statement is responsible for it as though they had written it.

Safe messaging. Established guidance on discussing suicide and self-harm in ways that do not increase the risk to vulnerable readers: no methods, no sensational language, always with sources of help. See 当提交内容提及自残时.

Shadow banning. Limiting the visibility of a user's content without telling them. Controversial because it is, by design, not transparent.

Sockpuppet. A second account or identity used by one person to appear to be several. On an anonymous form, every flooder is effectively operating sockpuppets.

T to Z

Takedown. The removal of published content, usually in response to a request or complaint. See 如何处理下架请求.

Threshold. The score above which a classifier's output is treated as a flag. Lower it and you catch more and wrongly flag more; raise it and the reverse happens. There is no correct value, only a trade-off suited to your purpose.

分类处理。 在处理传入项目之前，根据其紧急程度和难度进行分类，以便将注意力集中在最需要的地方。

信任与安全。 大型平台用来称呼处理滥用、有害内容和用户保护的团队和政策的名称。一个志愿者管理团队正在以较小的规模做着同样的工作。

两人原则。 要求在采取敏感行动之前，必须有两人同意。在一个小团队中，这是防止单个人因一时失误而犯错的主要保障。

用户生成内容 (UGC)。 由服务用户而非其运营者创建的材料。一个告解页面完全由管理员选择发布的UGC组成——正是这种选择，让他们对此负责。

替代性创伤。 版主反复接触令人痛苦的材料所产生的累积效应。这是真实存在的，它影响着志愿者和专业人士，也是轮换职责和缩短工作时间的原因。

Using the vocabulary

了解这些术语不如提出其背后的问题重要。当一个工具声称“99%准确”时，请问这个数字指的是精确度还是召回率，以及它是在哪些语言上测量的。当有人说一个帖子“没有指名道姓”时，请问那个人是否可以被识别。当大量提交的内容都说同样的事情时，请问你看到的是社区的意见还是有组织的行动。

常见问题

内容审核中的“误报”是什么？

无害内容被检查错误标记——这是你从恼火的提交者那里了解到的错误。它的反义词是“漏报”，即有害内容未能被检查标记，这是你从投诉中了解到的。

什么是“组队攻击”（brigading）？

A group organising in one place in order to descend on another — submitting, commenting or reporting en masse. The volume is meant to look like consensus.

什么是“人肉搜索”（doxxing）？

发布某人的身份识别或私人信息——地址、工作地点、电话号码、身份证件——以使他们遭受骚扰。自2025年起，在马来西亚，如果意图或效果如此，这已构成刑事犯罪。

预审核和后审核有什么区别？

预审核在内容发布前进行审查，这就是告白页面运作的方式，也是为什么其管理员对出现的内容负有明确责任。后审核是先发布后审查，通常是在内容被举报后进行，这是评论区通常的运作方式。

什么是“拼凑识别”（jigsaw identification）？

通过拼凑几个细节来识别一个人，每个细节本身都是无害的。一门课程、一个年份和一个社团角色，综合起来，可能就指向了某一个人。

“精确率”（precision）和“召回率”（recall）对审核工具意味着什么？

召回率是检查能够捕获的所有有害内容的比例；精确率是被标记的内容中确实有害的比例。高召回率通常伴随着更多的误报，而高精确率则伴随着更多的漏报，因此，声称“99%准确”意义不大，除非你知道它指的是哪一个。