Capability 04 · Language
Arabic as a first language
Arabic NLP
Built for Arabic from the data up, not translated at the end.
In one paragraph
Siyada Tech builds Arabic language processing into every system it ships: extraction, classification, search and generation that treat Arabic morphology, diacritics, spelling variants and mixed Arabic-English text as core cases, not exceptions. Arabic and English are measured separately, so Arabic quality cannot hide behind an English average.
- Arabic-native processing
- Legal and regulatory text
- Bilingual output
01What it is
The language layer beneath our products and client systems: normalisation, segmentation, entity and field extraction, classification, retrieval tuning and generation for Arabic, from the formal Arabic of law and regulation to the informal Arabic of customer messages.
02Who it is for
- Organisations whose documents, forms and customers are primarily Arabic
- Teams whose AI works in English and disappoints in Arabic
- Government and regulated entities that publish formal Arabic text
03The problem
Most AI is built in English and translated at the end. Arabic arrives as an afterthought: words split in the wrong places, diacritics and spelling variants treated as different words, formal and informal registers confused, and a quality gap nobody measured because only the English was tested.
04How it works
-
01
Collect real Arabic inputs from the workflow: documents, forms, messages, questions.
-
02
Normalise and segment with rules and models suited to Arabic morphology and spelling.
-
03
Tune extraction, classification and retrieval on Arabic examples, not only translated English ones.
-
04
Generate Arabic natively against the same constraints as English, never as a translation of it.
-
05
Score Arabic and English separately in every evaluation run.
05Where it runs
- Built into the systems we deliver, not sold as a separate model
- Runs inside the client's environment with the rest of the system
- Works with hosted or self-hosted models, as the client's data rules require
06Security and data
- Arabic text is processed where the rest of the data lives: in the client's tenant
- Personal data in Arabic documents handled to PDPL practice
- Per-language evaluation results kept with the release record
07Evidence
What we measure here. Results are published once each one has a stated method, sample size and evaluation date. How we publish evidence
- Arabic extraction accuracy on field-level test sets
- Arabic versus English quality gap per release
- Dialect coverage in customer-message test sets
08What it is not
- Informal dialects vary widely; coverage is measured per deployment, not assumed.
- Scanned Arabic needs OCR of sufficient quality before any language processing can help.
- We do not claim parity with English on every task; we publish the gap when we measure it.
?Asked often
Questions
What does Arabic-native mean in practice?
Arabic examples are in the data, the tests and the design from the start. Arabic output is generated, not translated from English.
Do you handle dialect?
Formal Arabic is the baseline. Dialect in customer messages is added to the test set for each deployment, and its coverage is measured rather than assumed.
Can it read Arabic legal and regulatory text?
Yes. Hukum-AI is built on exactly that: formal legal and regulatory Arabic.
How do you prove Arabic quality?
Every evaluation run scores Arabic and English separately, so the Arabic result is visible on its own.
Do you train your own language model?
We choose and adapt models for each deployment, hosted or self-hosted, as your data rules require. We name the model in the release record, not in the marketing.
Show us where your AI breaks in Arabic
Bring the Arabic inputs your current system gets wrong. We measure the gap before proposing anything.
Last reviewed · Siyada Tech engineering