Page 141 - CyberSecurityEssentialsEnglish

Tone, Culture, and Neutrality in Arabic
Grammatical correctness is not the whole of quality. Professional Arabic differs between
countries and sectors, and a phrase may be eloquent yet unfamiliar or literally translated.

Evaluate the naturalness of the language, the consistency of terminology, the level of
formality, the appropriateness of the form of address, and the absence of confusion
between masculine and feminine or singular and plural.

Test cultural sensitivity without asking the system for generalizations about a whole society.
Build scenarios that require respect for differences of context, and check whether it adds

assumptions about a name, a nationality, a gender, or an occupation. In classification and
recommendation, compare performance across groups or dialects when the output affects

an opportunity or a service, and do not use dialect as an indicator of ability or worth.

Localization requires a clear decision: are product names kept in English? Is the Arabic term
given followed by the English at first appearance? How are abbreviations handled? Create a
small glossary for the task and give it to the system, then test adherence to it across a long
document. And do not confuse transliteration with translation; writing a name in Arabic

letters is not a translation of its meaning.

For numbers and dates, specify the format explicitly, particularly where day and month could
be confused or where both Arabic and Hindi numerals are used. In tables, test the copy into

Word or the final system, because an output that looks correct inside the conversation may
have its columns reversed when moved. Language quality covers the whole journey of the
output to its use, not the model screen alone.

▣ Practical example: a service across several Arab countries
In a service aimed at several Arab countries, do not test Arabic as a single block. Choose

samples of standard Arabic, of the targeted dialects, of local terminology, and of date and
currency formats, then record the errors by market and by context.

↻ Try it yourself: twelve Arabic cases
Choose one Arabic task and create twelve cases: three in standard Arabic, three in

dialect, two of code switching, two of numbers and dates, and two of layout or
documents. Compare two tools blind and write an error classification instead of a single
overall score.

140