How to moderate hate speech and insults in Arabic and Turkish

To moderate Arabic and Turkish text, use a moderation API that lists both languages for the categories your rules cover, such as insults and hate speech, and test it on your own messages before you rely on it. Many tools were built for English first: some list Arabic for one category only, and some do not list Turkish at all. Detectivision checks Arabic and Turkish text for insults, hate speech, racism, sexism, sexual language and profanity, among 12 languages, in one API call.

Why English-first tools miss Arabic and Turkish

Arabic

• Many spoken varieties. People write online the way they speak: Egyptian, Gulf, Levantine or Maghrebi Arabic as often as Modern Standard Arabic. A word that is ordinary in one variety can be an insult in another, and a model trained mostly on formal text meets the everyday words less often.
• Arabic in Latin letters. Many people type Arabic with Latin letters and digits for the sounds Latin lacks: 3 for ع, 7 for ح, 2 for ء, as in "7abibi". A model that knows only Arabic script, or only English, reads it as noise.
• The script itself. Short vowels are usually left out, letters change shape with their position, and people stretch words with repeated letters or the ـ stroke for emphasis. A word list matches none of these unless every form is in it.
• Two directions in one line. Arabic runs right to left, while numbers, names and English words inside it run left to right. Code that cuts or cleans text without handling both can split a word in two.

Turkish

• Suffixes. Turkish adds suffixes to a root, so one insult takes many forms: aptal, aptallar, aptalsın, aptallığın. A list of whole words misses the forms it lacks. A list that matches inside words blocks harmless words that happen to contain a short vulgar one; our Turkish page shows examples from the most common Turkish words.
• Typing without Turkish letters. On a keyboard without them, ş becomes s, ç becomes c, ğ becomes g, ı becomes i, ö becomes o and ü becomes u: şerefsiz is typed serefsiz. Each spelling is a new entry for a list.
• Two letters i. Turkish has a dotted i and a dotless ı. Code that changes case without the Turkish locale turns a capital I into i rather than ı, which can turn a harmless word into a vulgar one.
• Abbreviations. Swear words are often typed as two or three letters, which a list cannot tell apart from ordinary abbreviations without the rest of the message.

Both

Messages mix languages: Arabic with English or French, Turkish with English or German. Emoji, repeated letters and symbols in place of letters are used to slip past filters. And the same word can be abuse or reporting abuse, depending on the sentence around it.

What each API publishes for Arabic and Turkish

Each company's facts are from its own pages, as read on 5 November 2026. Detectivision is not affiliated with any of them. A language in a list means the company says it covers it; it does not tell you how well. Only a test on your own messages does.

How Detectivision checks Arabic and Turkish

Send each message as it was written to POST /api/v1/moderation with "type": "text_moderation". Nothing is translated or cleaned first, and there is no language to set, so a community that writes in Arabic, Turkish and English needs one integration.

Each flag raised comes with its confidence, so your rules can treat them apart: remove hate at once, hold insults for a moderator, only warn on profanity. The models are our own: our team collects and cleans the training data, and trains and develops the models in-house.

curl https://api.detectivision.ai/api/v1/moderation \
  -H "Api-Key: $DETECTIVISION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "text_moderation",
    "input": "أنت غبي ولا أحد يريدك هنا",
    "mode": "detect",
    "flags": ["insult", "hate", "racism", "sexism", "sexual"],
    "metadata": { "commentId": "c_7310" }
  }'

{
  "id": 3184,
  "type": "text_moderation",
  "mode": "detect",
  "status": "completed",
  "flagged": true,
  "flags": [{ "flag": "insult", "confidence": 0.93 }],
  "checked": ["hate", "insult", "racism", "sexism", "sexual"],
  "unchecked": [],
  "tookMs": 640,
  "usage": { "charged": 2, "remaining": 9998 },
  "metadata": { "commentId": "c_7310" }
}

The message means "You are stupid and nobody wants you here." The Turkish "Sen salak birisin, kimse seni istemiyor." says the same and goes to the same endpoint. Values in the example are illustrative. A text can be up to 10,000 characters and costs 1 credit per started 2,500 characters by webhook, twice that for an answer at once (Text moderation docs).

Voice messages and the sound of videos in both languages go through audio moderation, which understands the same 12 languages and gives each finding with its start and end in seconds.

Test before you switch

1. Collect real messages. Take a few hundred messages your moderators have already judged in each language: removed ones and approved ones.
2. Cover how people write. Include Gulf, Egyptian, Levantine and Maghrebi Arabic if your users write them, Arabic typed in Latin letters, Turkish typed without its letters, mixed-language messages, and messages that quote abuse to report it.
3. Count both kinds of error. For each flag, count the abusive messages it missed and the harmless ones it flagged. One overall accuracy figure hides the category that matters to you.
4. Set a line per flag. A flag is raised from a confidence of 0.5 by default. Hold the band your moderators should see, and act automatically only above a line where harmless messages stop appearing.
5. Keep people who read the language. Appeals and borderline cases need moderators who read Arabic and Turkish, and know the slang of your community.

Start the free trial and run the test in the dashboard's Playground or through the API. The Perspective API alternative page maps Perspective's attributes to these flags if you are moving before it stops.