How it works
PostBlendr Moderation wraps a fine-tuned XLM-RoBERTa classifier trained on Arabic and English abusive-content data — Modern Standard Arabic plus Iraqi, Levantine, Gulf, Egyptian and Maghrebi dialects, Arabizi, and code-switched text. The model outputs a probability for each of eight classes; it never sees your account, only the text.
The eight classes
SAFE
No abusive content detected.
PROFANITY
Crude or vulgar language, not aimed at a person.
INSULT
A personal insult or demeaning remark toward someone.
HARASSMENT
Targeted, repeated or intimidating hostility toward a person.
HATE
Hostility toward a protected group or its members.
DEHUMANIZATION
Language that denies someone's humanity (comparisons to vermin, disease, etc.).
THREAT
An expression of intent to cause harm.
INCITEMENT
Encouraging or calling for violence or harm against others.
The policy layer
The model classifies; a separate, deterministic policy decides what to do. Each class has a calibrated confidence threshold and an ALLOW / REVIEW / BLOCK mapping, plus a severity override that can escalate borderline-confidence but high-severity cases. Operators edit this policy in the admin panel — changes are logged and take effect without a redeploy.
Known limitations
- DEHUMANIZATION is the weakest class and is treated as review-only regardless of confidence — the boundary with HATE and INSULT on severely degrading language is genuinely fuzzy.
- HARASSMENT rests on a small evaluation set; treat its calls as provisional.
- Short text, heavy sarcasm, and niche slang are harder. The model gives a probability, not a verdict — a human should confirm anything consequential.
Data & privacy
- The text you submit is stored. That is the point: submissions build the corpus that trains future models. Don't paste anything you wouldn't want retained.
- For a link, we fetch the public post text server-side (OpenGraph / oEmbed). We never sign in or bypass access controls; if a site blocks preview bots you're asked to paste the text yourself.
- We store a one-way salted hash of your IP for rate-limiting and abuse handling — never the raw address. Each visitor gets a small number of free checks before a short cooldown; a first-party cookie tracks that count per browser.
- Moderators review results in the admin panel and confirm the correct label — accepting the model's call or correcting it. Any submission may be picked for review, and those human-confirmed labels — never raw model guesses — are what feed retraining.