Building an Arabic NLP Portfolio That Can Be Reviewed

A useful Arabic NLP portfolio can begin with a small text-review task: preserve the supplied wording, identify a field when it is explicit and ask for clarification when it is not. You can demonstrate those decisions with synthetic examples before choosing a model.

Keep two fields in your project: original_text, which never changes, and a separate comparison_text for any transformation you are testing. That separation makes a normalization decision inspectable.

Start with ambiguity you can explain

Use these author-created examples as a miniature review set. The expected response belongs to this exercise; it is not a measured result from an AI system.

Supplied text	What the example tests	Expected handling
أرسل التقرير إلى علي.	A recipient is named	Retain علي as the supplied recipient
أرسلها إليه.	The text omits the thing and recipient	Ask what to send and to whom
الموعد غداً.	A relative date lacks a reference date	Preserve غداً; ask for the date/context before creating a dated entry
عَلَم / عِلْم / علم	Distinctions carried by vowel marks	Keep original forms; do not silently select a vocalization for علم

For the second example, a suitable clarification is: ما الذي تريد إرساله، وإلى من؟ For the third, ask ما التاريخ المقصود بغداً؟ if the task needs a calendar date. Include an empty response or an invented recipient among your failure cases so the review has something concrete to reject.

Separate Unicode normalization from editorial changes

Unicode Standard Annex #15 defines normalization forms, including NFC and NFKC. NFC handles canonical equivalence; NFKC also handles compatibility equivalence. Choosing one does not define your application's policy for removing Arabic vowel marks or replacing letters.

This JavaScript example can be run locally. It makes no model or network call: