Arabic Computational Linguist
& AI Language Engineer
I work on Arabic speech and text data, linguistic annotation, model evaluation, and data quality across Modern Standard Arabic and Moroccan Arabic. I combine linguistic analysis with Python, SQL, and regex to build reliable language-data workflows.
Arabic expertise grounded in multilingual language work and applied AI evaluation.
I’m a native Arabic speaker with a background in translation and linguistics and professional experience evaluating multilingual AI data. My work spans Modern Standard Arabic, Moroccan Arabic, and North African dialect variation, including phonetics, morphology, diacritics, semantic ambiguity, and code-switching.
At Meta through TEKsystems and Apple through Appen, I worked on language-data quality, annotation, and model-output evaluation. I use Python, SQL, and regular expressions to extract, filter, and validate data, and apply linguistic judgment to ambiguous cases and annotation guidelines.
My independent research examines Moroccan Arabizi de-romanization through automatic metrics, native-speaker audits, and error analysis. Earlier translation and proofreading work across Arabic, French, and English continues to inform how I assess meaning, context, and language quality.
The capabilities I bring to language-data teams and independent NLP research.
Applied in Arabic language-data evaluation and linguistic error analysis.
Experience with Siri speech/text data and Arabic AI evaluation workflows.
Used in quality workflows and the four-system de-romanization study.
Professional language work across documents, documentaries, software, and games.
From multilingual language work to Arabic AI data and evaluation.
Built and evaluated Arabic training and evaluation datasets across MSA and dialects. Created, reviewed, and adjudicated annotations and golden datasets. Used Python, SQL, regex, tools, and APIs for extraction and quality checks, and provided feedback on guidelines, prompts, and evaluation metrics.
Evaluated speech and text language data for Siri command applications. Reviewed model-generated responses for meaning, wording, context, and consistency, applying phonetics, syntax, semantics, and morphology to annotation and quality review.
Reviewed language and search data for accuracy, usefulness, relevance, and contextual interpretation. Collaborated with team members to resolve ambiguous content and quality issues.
Translated and proofread French, English, and Arabic content for clients including VICE/HBO and Sensata Technologies. Work spanned documents, documentaries, localization, and video games, with research into slang, idioms, and specialized terminology.
Studied linear algebra, statistical modeling, NLP, NLU, and Transformers. Developed and applied a custom NLP pipeline for sentiment analysis.
Advanced study in linguistics, audiovisual translation, digital-content localization, and translation project management, with analysis of morphological, syntactic, semantic, and contextual challenges.
Study of phonetics, syntax, semantics, sociolinguistics, and practical applications of linguistic theory.
Research, language tooling, and multilingual contributions.
Automatic De-Romanization of Maghribi Arabizi to Arabic Script in Social Media
An empirical benchmark comparing an ordered rule mapper, MLE word mapping with regex fallback, a word-level character LSTM, and a full-passage mT5-small Transformer on the same 805-passage test set. Native-speaker audits and linguistic error analysis examine reference quality, alignment noise, and dialectal variation.
No single system leads on every metric. mT5 has the highest observed BLEU; MLE has the strongest character- and word-level form fidelity.
Accepted for camera-ready at the 9th International Conference on Arabic Language Processing. The revised manuscript has been submitted; final confirmation is pending. The camera-ready manuscript is available above. Research code and aggregate evidence are public on GitHub; dataset access remains subject to upstream conditions.
| System | BLEU ↑ | chrF ↑ | CER % ↓ | WER % ↓ | PosEM % ↑ |
|---|---|---|---|---|---|
| Rule mapper | 7.24 | 46.73 | 24.50 | 72.50 | 8.69 |
| MLE + regex | 35.44 | 69.60 | 15.17 | 42.48 | 16.76 |
| Word LSTM | 9.62 | 51.98 | 44.94 | 69.00 | 9.42 |
| mT5-small | 36.75 | 68.86 | 42.86 | 64.03 | 21.54 |
PosEM measures position-wise token agreement. Bold values mark the best observed score in each column. The mT5–MLE BLEU difference is 1.31 points (95% paired-bootstrap interval: −0.55 to 3.24; p = 0.074), which does not establish superiority at the 0.05 level or equivalence.
Dataset: 8,045 passage pairs from UBC-NLP / NileChat Arabizi-Morocco, split into 6,436 training, 804 validation, and 805 test pairs. Its Arabizi was generated from Arabic-script Moroccan material with Command R+; performance on naturally authored Arabizi remains untested. Source data is subject to upstream access conditions and CC BY-NC 4.0.
Human audits: two native Moroccan Arabic speakers independently accepted all 300 sampled test references. All-positive labels make Cohen’s κ undefined. A separate 100-pair audit found 75 aligned and 25 misaligned positional training pairs. Review of 50 apparent LSTM errors yielded 42 true errors, 5 acceptable variants/non-errors, and 3 uncertain cases; this error-conditioned sample is not an accuracy estimate.
Interpretation: mT5 uses full passages, external pretraining, and more eligible training data than the word-level systems. This is not an architecture-controlled comparison. All 78 mT5 outputs that failed to emit EOS within the generation limit remain in the primary scores. Split provenance and source-document independence retain limitations described in the manuscript. The exploratory human mT5–MLE review does not establish a human-preference winner.
An educational Python notebook exploring URL removal, custom emoji-to-text mappings, English-oriented normalization, and TextBlob sentiment scoring. Word clouds and label counts help inspect the processed sample; the project does not claim benchmarked classification accuracy.
Interested in Arabic speech and text evaluation, language-data quality, or NLP research? Let’s connect.
Tampa Bay Area, FL