Arabic Dialect Detection - Developer Platform

NAWA classifies Arabic text into four major dialect groups using a deterministic marker matcher, not a language model. That means the result is auditable: every classification can be traced to the specific words that produced it, and it costs nothing and adds no latency.

Supported dialects

Dialect	Code	Regions	Key markers
Gulf	`gulf`	Saudi Arabia, UAE, Kuwait, Qatar, Bahrain, Oman	وش، شلونك، وايد، زين، أخوي، طال عمرك، أبشر، بغيت
Egyptian	`egyptian`	Egypt	أوي، ده، دي، عايز، ازاي، كده
Levantine	`levantine`	Lebanon, Syria, Jordan, Palestine	هلق، كتير، شو، هيك، منيح
MSA	`msa`	Pan-Arab (formal)	إن، الذي، بالإضافة، يجب، لذلك

Detection is marker-based, so its accuracy depends on whether a comment actually contains dialect-specific vocabulary. A short or formally-worded Arabic comment carries no markers and correctly returns msa with a low confidence. Treat dialect_confidence as the signal for how much to trust the label.

Example classifications

Gulf Arabic

Comment: “وش رايكم بالمحتوى الجديد؟ أنا شايف إنه وايد حلو”Translation: “What do you think of the new content? I think it’s very nice”Markers: وش (what - Gulf), وايد (very - Gulf), حلو (nice - Gulf)

{
  "dialect": "gulf",
  "dialect_confidence": 0.97
}

Egyptian Arabic

Comment: “الفيديو ده جامد أوي، عايز تاني كده”Translation: “This video is amazing, I want more like this”Markers: ده (this - Egyptian), أوي (very - Egyptian), عايز (I want - Egyptian), كده (like this - Egyptian)

{
  "dialect": "egyptian",
  "dialect_confidence": 0.98
}

Levantine Arabic

Comment: “كتير حلو الفيديو، بس شو القصة ورا الأغنية؟”Translation: “Very nice video, but what’s the story behind the song?”Markers: كتير (very - Levantine), شو (what - Levantine)

{
  "dialect": "levantine",
  "dialect_confidence": 0.95
}

Modern Standard Arabic (MSA)

Comment: “يجب أن نشجع هذا النوع من المحتوى الهادف”Translation: “We should encourage this type of meaningful content”Markers: يجب أن (must - MSA formal), هذا النوع من (this type of - MSA structure)

{
  "dialect": "msa",
  "dialect_confidence": 0.93
}

How it works

How the pipeline works

Dialect detection runs inside NAGL (NAWA AI Gateway Layer) and does not involve a language model:

Script and language detection: identifies the input language. Arabic written in Latin letters and digits (Arabizi, for example shlonkm ya shabab) is transliterated to Arabic script first, so it is classified on meaning rather than spelling.
Marker matching: the text is scored against weighted vocabulary sets for each dialect. Only dialect-specific words count. Words shared across every dialect, such as متى or مرحباً, are deliberately excluded, because they would otherwise pull ordinary Gulf and Egyptian comments toward MSA.
Confidence scoring: a score from 0 to 1 derived from how many markers matched and how strongly they distinguish one dialect from the others.

Because it is deterministic, the same input always produces the same dialect, and a misclassification is fixed by correcting a marker list rather than by retraining anything.

The `dialect_confidence` field

The dialect_confidence score ranges from 0 to 1:

Range	Interpretation
0.90–1.00	High confidence - strong dialectal markers present
0.70–0.89	Medium confidence - some dialectal features detected
0.50–0.69	Low confidence - text may be code-switched or ambiguous
< 0.50	Very low - text may be too short or use minimal dialectal features

Short comments (under 10 words) often have lower dialect confidence because there are fewer linguistic markers. For critical applications, consider filtering on dialect_confidence > 0.7.

Improving accuracy with feedback

If NAWA misclassifies a dialect, submit feedback to improve the model:

curl -X POST https://api.trynawa.com/v1/feedback \
  -H "Authorization: Bearer nawa_test_sk_xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "request_id": "req_abc123",
    "field": "dialect",
    "expected_value": "levantine",
    "comment": "This is Lebanese Arabic, uses هلق and كتير"
  }'

RLHF feedback is incorporated into model fine-tuning cycles, continuously improving accuracy across dialects.

​Supported dialects

​Example classifications

​How it works

​How the pipeline works

​The dialect_confidence field

​Improving accuracy with feedback

Supported dialects

Example classifications

How it works

How the pipeline works

The `dialect_confidence` field

Improving accuracy with feedback