Table of Contents
Algerian plant teams write procedures in Modern Standard Arabic, French or English, but ask questions in Darija mixed with French technical terms, sometimes typed in Latin letters. Most AI assistants are trained and tested on standard languages, so they misread these messages, miss the right document or answer in the wrong language. Design and test for this explicitly.
Key takeaways
- Algerian Darija is written in Arabic script or in Latin letters (Arabizi). Written in Latin letters, it has no standard spelling and mixes heavily with French (Muller et al., 2020).
- In a 2025 test of seven models translating Arabizi, Algerian scored lowest of three dialects, with an average BLEU of 4.24 against 9.65 for Egyptian (Al Almaoui et al., 2025).
- Model cards rarely name Algerian Arabic. Qwen3 lists Moroccan and Tunisian Arabic among its varieties, but not Algerian (Qwen).
- Arabic LLM benchmarks still lack code-switching evaluation, including Arabic-French mixing (Alzubaidi et al., 2025).
- The practical answer is to understand Darija on input, answer in MSA or French, quote values verbatim and test with your own teams' real messages.
How do Algerian plant teams actually write and speak?
The constitution makes Arabic the national and official language, and Tamazight also a national and official language (Constitution of 2020, Articles 3 and 4). Check with counsel which language a given document must use. Day-to-day plant language is more mixed. As our guide to multilingual documents describes, manuals often arrive in English, internal procedures are often in French, and forms and correspondence may be in Arabic.
Spoken and informal written language is something else again: Algerian Darija, with French equipment names, fault terms and verbs dropped into the sentence. Researchers who studied Maghrebi broadcast speech found a relatively high presence of code-switching in Algerian Arabic compared with Morocco and Tunisia (Amazouz et al., 2017).
This register shows up wherever people type or speak freely: shift notes, maintenance requests, messages to a supervisor, and the questions they will type into an assistant. The same idea, "the pump is overheating, the bearing needs changing", might arrive as:
- MSA: «المضخة ترتفع حرارتها، يجب تغيير المحمل»
- Darija in Arabic script with French terms: «الـ pompe راهي تسخن، لازم نبدلو الـ roulement»
- Arabizi: "la pompe rahi tsakhan, lazem nbadlou le roulement"
- French: « la pompe chauffe, il faut changer le roulement »
These are illustrative examples, not quotes from a client. An assistant tested only on the first and last forms has not been tested on the middle two.
Why is Darija hard for language models?
Little written data. Algerian Darija has little presence in digital corpora (Al Almaoui et al., 2025), and the text researchers do collect comes largely from social media. DziriBERT, a model built specifically for the Algerian dialect, was pre-trained on more than one million Algerian tweets, and its authors report it beat existing models, especially on text in Roman script (Abdaoui et al., 2021). The first treebank for romanised Algerian user-generated text contains about 1,500 annotated sentences (Seddah et al., 2020). These are small numbers compared with the text available for MSA or French.
No standard spelling. Researchers describe North African Arabizi as written in Latin script "with no standard spelling and no standard transliteration of Arabic letters", with a high degree of variability across writers (Muller et al., 2020). Arabizi uses Latin characters and numbers (Al Almaoui et al., 2025), so one word can be typed several ways.
Heavy French. In the annotated Algerian Arabizi training data used in that study, French made up 33.84% of tokens (Muller et al., 2020). That was social media text; in a plant, the technical vocabulary is even more likely to be French.
Two scripts and mixed grammar. A survey of code-switched Arabic NLP notes that romanisation is non-standard and that unstandardised spelling complicates evaluation (Hamed et al., 2025). French verbs can also take Darija grammar, which standard Arabic and French analysers are not designed for.
| Feature | Illustrative example | What breaks | What to do |
|---|---|---|---|
| Darija in Arabic script with French terms | «الـ pompe راهي تسخن» | Keyword search on the MSA term مضخة misses it; Arabic stemmers mangle the mix | Glossary expansion; cross-lingual embeddings |
| Arabizi | "la pompe rahi tsakhan" | Arabic-script analysers never see it; a language detector may label it French or unknown | Detect it explicitly; test it as its own language |
| French words with Darija grammar | "ydémarri" (it starts up) | Stemming and glossary matching | Add inflected forms to the glossary |
| Spelling variants | roulement, roulman, رولمان | Exact match | Fuzzy matching and a variants list |
| Tags and codes inside the sentence | "P-201 dkhel f défaut" | Tokenisation; tags split or altered | Keep tags as exact tokens; quote them verbatim |
| Voice notes | Darija and French in one breath | Speech recognition errors on technical terms | Test with your own recordings |
What does research say about LLMs and Algerian Darija?
The most direct evidence is a 2025 study that asked seven models, including GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3.3 70B, Mistral Large and Jais, to translate Arabizi from Lebanese, Egyptian and Algerian sources. Algerian ranked third, with an average BLEU of 4.24 and chrF of 23.21, against 9.65 and 34.64 for Egyptian. The authors link this to heavy code-switching with Berber, French and Spanish and to the dialect's thin presence in digital corpora. Their conclusion: "Models struggle significantly with Arabizi", although out-of-the-box models were still usable for getting the gist (Al Almaoui et al., 2025).
Getting the gist is not enough when the message contains a torque value or an isolation step.
Benchmarks do not fill the gap yet:
- DialectalArabicMMLU, accepted at LREC 2026, covers Syrian, Egyptian, Emirati, Saudi and Moroccan dialects, not Algerian (Altakrori et al., 2025).
- A 2025 survey of Arabic LLM benchmarks lists code-switching evaluation, including Arabic-French mixing, among the critical gaps (Alzubaidi et al., 2025).
- Algerian-French is one of the most studied code-switching pairs in research papers (Hamed et al., 2025), but the resources are small: one speech corpus of Algerian Arabic and French code-switching totals about four hours from eight speakers (Amazouz et al., 2017).
Model cards are similarly vague. At the time of writing, the Llama 4 card lists "Arabic" without naming dialects (Meta), and Qwen3 lists Standard, Moroccan, Tunisian and other Arabic varieties, but not Algerian (Qwen). Our post on choosing an open-weight LLM covers the other selection criteria.
Dialect-specific training can help. Atlas-Chat, a family of models fine-tuned for Moroccan Darija, reports a 13% performance gain for its 9B model over a larger 13B model on its DarijaMMLU benchmark (Shang et al., 2024). Moroccan and Algerian Darija are related but not the same, so such results do not transfer automatically.
Should the assistant answer in Darija?
In most plants, no. A sensible split is:
- Understand Darija on input, in Arabic script and in Arabizi, including French terms and tags.
- Answer in MSA or French, whichever the user chooses, and say so in the instructions given to the model. Language models sometimes answer in the wrong language, more often with complex prompts and high sampling temperatures (Marchisio et al., 2024).
- Quote values, steps and warnings verbatim from the released document in its original language. Darija has no standard spelling, so a procedure paraphrased into it cannot be checked against anything.
This keeps the controlled document as the reference, which is the same principle as in our post on operator assistants and work instructions.
Two side effects need checking. Token counts for the same content can differ by up to 15 times between languages (Petrov et al., 2023), so measure samples of your own mixed and romanised messages with each candidate's tokenizer before sizing hardware. Speech needs its own test: Whisper's documentation says its "performance varies widely depending on the language" (OpenAI Whisper).
How do you test an assistant on Darija and technical French?
Public benchmarks will not tell you, so build your own test. Our guide to evaluating an LLM on your own documents covers scoring; the Darija-specific steps are:
- Collect real messages. Take questions and notes your technicians actually wrote or said, in every form they use: MSA, French, Darija in Arabic script, Arabizi and voice. Remove names and personal details.
- Build a term glossary. For each key piece of equipment and fault, list the MSA term, the French term and the Darija forms in both scripts, with common spelling variants.
- Write reference answers. For each question, an engineer records the correct document, revision and passage, and the expected answer language.
- Score each written form separately. Measure retrieval and answer accuracy for MSA, French, Arabic-script Darija and Arabizi questions on their own, not as one average.
- Check answer language and verbatim values. Every answer should be in the requested language, with tags, units and values matching the source exactly.
- Test voice input on site. Record operators in real conditions and check how technical terms and tags are transcribed.
- Review with fluent engineers. Have someone who speaks Darija and knows the process read a sample of answers.
- Re-run after every change to the model, prompt, glossary or index.
If a model fails on the Arabic-script Darija and Arabizi rows but succeeds on MSA and French, start with the glossary and retrieval before changing models. Our guide to measuring hallucination covers the answers that look right but are not.
What we have seen at KABAS AI
DALILE, our AI knowledge assistant, runs on the client's own servers without internet access, organises documents into searchable knowledge libraries and answers questions with exact sources. That matters for this topic: real Darija messages, voice samples and the term glossary built from them reflect how a site works, and with DALILE they stay on site alongside the documents.
We have not published results on Darija. Our practical advice is to start with a language inventory of the informal channels as well as the documents: shift logs, maintenance requests and the questions people already send to colleagues. That inventory shows how much Darija and Arabizi the assistant will meet, and it supplies the first test set. The same applies to shift reports and handover notes, where operators' own wording is the input.
For background on why Arabic industrial vocabulary needs attention, see our post on LLMs and Arabic industrial terminology. Our agentic AI team deploys DALILE and connects it to your document management systems and databases.
Sources
- Constitution of Algeria, 2020 (Constitute Project)
- Amazouz, Adda-Decker and Lamel, Addressing Code-Switching in French/Algerian Arabic Speech, Interspeech 2017
- Muller, Sagot and Seddah, Can Multilingual Language Models Transfer to an Unseen Dialect? A Case Study on North African Arabizi (arXiv 2005.00318, 2020)
- Seddah et al., Building a User-Generated Content North-African Arabizi Treebank: Tackling Hell, ACL 2020
- Abdaoui et al., DziriBERT: a Pre-trained Language Model for the Algerian Dialect (arXiv 2109.12346, 2021)
- Hamed et al., A Survey of Code-switched Arabic NLP: Progress, Challenges, and Future Directions, COLING 2025 (arXiv 2501.13419)
- Al Almaoui, Bouillon and Hengchen, Arabizi vs LLMs: Can the Genie Understand the Language of Aladdin? (arXiv 2502.20973, 2025)
- Altakrori et al., DialectalArabicMMLU: Benchmarking Dialectal Capabilities in Arabic and Multilingual Language Models (arXiv 2510.27543, 2025)
- Alzubaidi et al., Evaluating Arabic Large Language Models: A Survey of Benchmarks, Methods, and Gaps (arXiv 2510.13430, 2025)
- Qwen Team, Qwen3: Think Deeper, Act Faster
- Meta, Llama 4 Scout 17B-16E Instruct model card (Hugging Face)
- Shang et al., Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect (arXiv 2409.17912, 2024)
- Marchisio et al., Understanding and Mitigating Language Confusion in LLMs (arXiv 2406.20052, 2024)
- Petrov et al., Language Model Tokenizers Introduce Unfairness Between Languages (arXiv 2305.15425, 2023)
- OpenAI, Whisper (GitHub)
Frequently asked questions
Can general AI assistants understand Algerian Darija?
Partly. In a 2025 study, seven well-known models translated Algerian Arabizi worse than Egyptian or Lebanese, and the authors concluded that models struggle significantly with Arabizi, while still finding them usable for getting the gist. Test any candidate model on real samples of your own team's messages before relying on it.
What is Arabizi, and why does it matter for a plant assistant?
Arabizi is Arabic, often dialect, written with Latin letters and digits. Many people use it in messages and chats. A search index or language model tuned for Arabic script may not recognise it, and there is no standard spelling, so the same word appears in several forms.
Should the assistant reply in Darija?
Usually not. Let the assistant understand questions in Darija, in Arabic script or Latin letters, and answer in the language the user chooses, typically Modern Standard Arabic or French. Darija has no standard spelling, and values, steps and warnings should be quoted from the released document in its own language.
Do we need to fine-tune a model on Darija?
Not as a first step. Start with a glossary of the French and Darija terms your teams use, cross-lingual retrieval and a test set built from real messages. Dialect-specific training has helped elsewhere, as with Atlas-Chat for Moroccan Darija, but it needs data and evaluation of its own, so consider it only if tests show retrieval and prompting are not enough.
Does voice input work with Darija mixed with French?
Test it before promising it. Whisper's own documentation says its performance varies widely depending on the language, and published Algerian Arabic and French code-switching speech corpora are small. Record real samples from your operators, with shop floor noise, and measure the transcription errors on technical terms and tags.



