How matching works

English-only libraries miss mixed-script Maghrebi chat. Anti-Ttyah normalizes first, then looks up tokens and phrases.

raw token
  → strip tashkeel / tatweel
  → fold أ إ آ → ا  and  ة → ه
  → Franco-arabe digits (3 → ع/a, 7 → ح/h, 9 → ق/q, …)
  → collapse repeats (n*kkk → n*k)
  → drop article prefixes (el / al / ال)
  → fold separators (. * - _ ') when matchSeparators is on
  → lookup against the listed keys

The playground Explain trace shows the keys for each token.

What gets folded

Step	Example idea
Tashkeel / tatweel	Marks on Arabic letters are ignored
Alef / teh marbuta	أ/إ/آ → ا, ة → ه
Digit-speak	3 7 9 5 8 2 6 (and Latin 1 0)
Repeats	Long runs of the same letter collapse
Articles	el / al / ال prefixes
Separators	z.e.b.i and ze*bi count as one token (default)

The full wordlist is not shown here.

Tokens vs phrases

• Terms are single tokens ([\p{L}\p{N}]+, optionally joined by separators).
• Phrases match consecutive tokens (ولد الق*بة, nique ta mere).

inspect returns spans in the original string, so you can highlight or mask without guessing offsets. Hits from the default list also carry lang and category.

Categories

Default entries are tagged sexual, insult, slur, or mild. Pass categories to createFilter (or --categories on the CLI) to use only some buckets. Extra terms you add yourself are always included.

False positives

Keep the list conservative. If a token is also a product name or given name, prefer allowlist over deleting a useful insult. See Wordlist.