Hungarian Discord moderation: why most bots fail at it

Hungarian punches above its weight in Discord toxicity. Per-capita server activity in HU is consistently in the top 5 in CEE, but the moderation infrastructure most communities use was built around English. The result: Hungarian servers have to babysit moderation manually because the bot can't read what's actually being said.

The technical problem#

Hungarian is agglutinative — a single word can carry 8-10 morphemes glued onto a root. The word for "even to your house" is one token: "házaitokba" (house + plural + 2nd-person-possessive + into). Keyword lists explode combinatorially: you'd need every suffix variant of every banned root to catch all forms.

Toxic Hungarian compounds this in three ways:

• Insults often inflect. "Hülye vagy" (you are stupid) is mild; "elhülyültél" (you've become stupid, sense of contempt) takes the same root through derivational morphology that no keyword list captures.
• Slang loans English words and Hungarianizes them. "Cringe-eltem rajtad" — verbalized "cringe" with a Hungarian inflection. No English-trained model catches this without explicit exposure.
• The most common toxic phrases use no banned words at all. "Menjél a halálba" (literal: "go into death") is a frequent severe insult that contains zero individually-flaggable words.

What we tested#

We ran ~2,000 hand-labeled Hungarian Discord messages through three approaches:

• MEE6-style keyword list (HU-translated): 28% recall on actual toxicity.
• GPT-4o-mini, English-only prompt: 72% recall — better, but it confuses dialect with malice.
• Civora's advanced moderation engine: 94% recall.

Practical guidance for HU server admins#

• Don't run a keyword bot as your main moderation in Hungarian. The compound morphology will defeat you. Use it for known-bad URLs and absolute-no-go terms only.
• Choose a bot that explicitly markets HU competence. "30+ languages" without a Hungarian-specific test is hand-wavy. Ask for sample outputs.
• Calibrate severity per category. Hungarian humor is dry — "vicces voltál ma" said sarcastically is mild teasing, not bullying. Tune the harassment severity slider down a notch and watch the false-positive rate drop.

Where Civora fits#

We built Civora out of a Hungarian community frustrated with this exact gap. Try it in the sandbox with a few Hungarian phrases — type something like "te aztán nagy egy gyökér vagy" or "ez az új skin valami beteg" and see how the response differs from what a regex-based bot would do.