Moderating a Developer Community

Lê Thành Trung

A filter that matches single slurs will miss Vietnamese toxicity. “chăn rau” alone is harmless. “Đi chăn rau khu Cầu Giấy, cho xin số điện thoại em nhé?” is not.

That is why openmod-vn starts with context, not a word list.

Rules before tools

The project is an open Vietnamese moderation dataset and a four-layer pipeline.

• L0 normalizer handles diacritics, teencode, and leetspeak.
• L1 lexical matcher checks 435+ terms with variant spellings.
• L2 compositional detector reads 13 skip-gram rules and scores suspicion.
• L3 decision engine returns an action, a severity, spans, and a rationale.

The layers exist because the same word can be a slur, a joke, or a legitimate technical term. A keyword list will fail. The label set has to be specific enough for two annotators to agree, and the policy has to be specific enough for a moderator to apply.

Policy profiles

A community is not one audience. openmod-vn ships policy profiles: kid_safe, standard_social, and relaxed_gaming. Each profile sets its own thresholds.

That is the part most moderation tools skip. A rule that is right for a gaming server is wrong for a school group. The policy is a product decision, not a model setting.

The bot is not the moderator

The pipeline returns an action, not a verdict. A message can be ALLOW, WARN, or BLOCK. A WARN can delete the message and reply with the reason. An ALLOW can leave it alone because the context is benign.

The output includes the span and the category, so the moderator can see why. “Blocked” is not enough. “Blocked because this phrase targets a person in this sentence” is a decision someone can review.

The appeal path

A disputed label needs a review path. The dataset is CC-BY-SA-4.0 with citations, and a label change goes through a review batch. The point is not to defend the model. The point is to make the label inspectable.

A community needs the same path for a moderation action: a private review, the rule, the evidence, and a recorded outcome. The appeal path protects the person from a bad call and the moderator from the assumption that every call is final.

What I would change

I would write the policy with the community, not for it. I would test the bot against the last batch of reports before turning it on. And I would publish the appeal outcome, even when the answer is “the action stands.”

The label, the span, and the rationale are the product.