Translation Methodology

UAIE for Microsoft Word · Published by Fiza Pathan Publishing (OPC) Private Limited · Invented by Fiza Pathan · Effective date: 2 May 2026 (v1.2.10)

Two surfaces, two methodologies

Translation in UAIE happens at two distinct surfaces, and the methodology behind each is different. The page that follows is organised accordingly.

Surface 1 — Interface translation. The static product strings — module titles, button labels, status messages, footer text, the words you are reading right now if you have switched UAIE into a non-English language. These 31 localisations are produced through the four-skill methodology described below, classified at one of four verification levels, and shown in the language selector as a small coloured badge. This is the surface most of this page covers.

Surface 2 — Document translation. When a user selects text inside a Microsoft Word document and runs the Translate module, the actual translation of the user's content is performed dynamically by Anthropic's Claude through the user's own API key. This is a different quality profile, governed by Claude's training rather than by our editorial verification. We document the relevant honest expectations for this surface in its own section near the foot of this page.

Conflating the two would be misleading. The badges and verification levels described in the next sections refer only to interface translation. Document translation has its own separate quality discussion further down.

Interface translation — the four-tier verification model

Every supported language is classified at one of four levels, shown as a small coloured badge next to the language selector inside UAIE:

• Native — English. The source of truth from which all other localisations are derived.
• Verified — machine-translated and cross-checked against the documented resource for that language family. Reasonably trustworthy for routine use; native-speaker review still welcomed.
• Best-effort — machine-translated against general-purpose tools, awaiting native-speaker review. Treat as provisional.
• Pending — placeholder; work outstanding.

The four language-family methodologies

Translation work in UAIE is organised by language family, with a documented methodology for each. The methodologies are formalised as internal skills used by the publishing house's editorial workflow and are summarised here.

Indian and classical-Indian languages

For Hindi, Marathi, Tamil, Telugu, Bengali, Kannada, and Urdu, the principal verification resource is Bhashini — the Government of India's machine-translation initiative covering 23+ Indian languages. Bhashini provides translations more responsive to Indian-standard register than international tools.

For Sanskrit, the principal resource is the IIT Kanpur Gita Supersite, which carries the Bhagavad Gita, the principal Upanishads, and the Brahma Sutras with the commentaries of Shankara, Ramanuja, Madhva, and Sayana, in Devanagari, IAST transliteration, and English. Where a Sanskrit phrase has been rendered in scriptural register, the Gita Supersite is the gold-standard reference. We do not use general-purpose translators (Google Translate, DeepL, Microsoft Translator) for Sanskrit because they are not trained on Vedic register and produce unreliable output for classical texts.

African languages

For Yoruba, Igbo, Amharic, Zulu, Hausa, Swahili, and Afrikaans, the named first-stop tool is Lekhak, a translator that specialises in African languages. Specialist agencies (Lughayangu, White Globe, Stepes) are used for documents that need a native-quality pass beyond first-draft machine translation. Government-recognised certified services (RushTranslate, ImmiTranslate, Certified Translation India, Novatia Translations for Nigerian Yorùbá) are used for any document headed for a court, embassy, or ministry.

UN official languages

For Arabic, Chinese, English, French, Russian, Spanish, German, and Portuguese — the eight languages best supported by United Nations resources — the principal terminology source is UNTERM, the public multilingual UN terminology database. The Official Document System (ODS) is used to find precedent translations of phrases that have already been rendered officially. For first-draft translations and contextual verification, DeepL is the named first stop for European pairs and Google Translate for Arabic and Chinese (its training set includes UN documents directly, which makes it unusually accurate on diplomatic phrasing in those pairs).

Middle Eastern Arabic and Urdu

For Gulf-jurisdiction Arabic, the principal authoritative sources are the UAE government portal U.ae, the Dubai government portal Dubai.ae, the Ministry of Human Resources and Emiratisation (MOHRE), and the Saudi government portal Saudi.sa. Gulf register differs from UN-Modern-Standard register; for institutional outreach in the Gulf, the government portal takes priority. Microsoft Translator is named first among AI tools because it is documented as integrated into Gulf government systems. Lekhak handles Urdu-to-Arabic with OCR for scanned material; Bluente handles legal documents where formatting must be preserved.

Per-language interface status (v1.2.10)

The classification below applies to the interface strings only — the static localisation of UAIE itself. For the Translate module's quality on these languages, see the document-translation section further down.

Document translation — the Translate module

The Translate module is a separate feature from interface translation. The user selects text in their Word document, presses Translate Selection or Entire Document, and receives the translated text inserted at the chosen location. The quality and the methodology behind this feature are different from the interface translations described above.

How the Translate module works

The Translate module sends the selected text to Anthropic's Claude through the user's own API key, which the user provides in the optional Settings panel inside UAIE. The translated text returns directly to the user's machine and is inserted into the document. The publishing house has no servers; nothing about the source text, the target language, or the translated output passes through any system controlled by Fiza Pathan Publishing. The Privacy Policy describes this in full.

The engine is therefore Claude — a large language model trained extensively on multilingual text. It is not a custom-built translation engine, and it is not an aggregator of multiple translation tools. The quality the user experiences is the quality of Claude's translation capability for the language pair in question, on that day.

What this means for translation quality

Claude's translation capability is generally strong, often very strong, but it is not uniform across languages. Honest expectations for the Translate module, organised by language family:

• UN official languages and major European languages (Arabic, Chinese, English, French, Russian, Spanish, German, Italian, Portuguese, Dutch, Polish, Greek). Claude's translation between these and English is broadly excellent, suitable for routine institutional correspondence, study, and draft work. Edge cases involve idiomatic register and specialist terminology, where a UN-trained tool such as UNTERM or DeepL may still be a useful cross-check for material going to a serious destination.
• Major Indian languages (Hindi, Marathi, Urdu). Strong, particularly for general-register translation. For the more Persianised and Sanskritised registers — Urdu shifting toward Pakistani-standard or Hindi shifting toward Sanskrit-derived formal vocabulary — quality varies, and Bhashini remains the named verification source for any Indian-language translation that must be institutional in tone.
• Smaller Indian languages (Tamil, Telugu, Bengali, Kannada). Variable. Claude has substantial training on these, but less than for Hindi. Use the Translate module for comprehension and draft; verify through Bhashini or Anuvadini AI for material with weight.
• East Asian (Japanese, Chinese). Strong for both. Japanese in particular is a language Claude handles well across registers.
• Sanskrit. Claude has rich training on classical and Vedic texts in translation — the Bhagavad Gita, the principal Upanishads, the Brahma Sutras, the major commentators — but generation in Sanskrit itself is uneven. For comprehension of a Sanskrit passage, the Translate module is useful; for production of new Sanskrit content of scholarly quality, the IIT Kanpur Gita Supersite and traditional commentaries remain the right reference, not the Translate module.
• Middle Eastern Arabic and Urdu in Gulf register. Modern Standard Arabic is strong; Gulf-jurisdiction register varies. For documents headed for UAE or Saudi institutions, the Tier 1 government portals (U.ae, Saudi.sa) remain the authoritative reference — Claude's output is a starting point, not a final answer.
• African languages (Swahili, Afrikaans, Hausa, Yoruba, Igbo, Amharic, Zulu). Variable, with the smaller-corpus languages weakest. Swahili and Afrikaans are reasonable; the smaller Sub-Saharan languages should be treated as draft-quality and verified through the named African-language tools (Lekhak, Lughayangu, the certified services) for any institutional or pilot-school material.

When to use the Translate module — and when not to

The Translate module is appropriate for:

• Comprehension and draft. Reading a passage in a language you do not know well; producing a first-pass translation to be edited by hand.
• Internal correspondence. Translating an email, an internal memo, a personal note, a study summary.
• Educational use. Helping a student access material in a language they are learning; producing reading material in plain language across languages.
• Working drafts of any document for which the user is going to do a human pass before sending.

The Translate module is not appropriate for, and must not be relied upon for:

• Court submissions, immigration documents, embassy material, certified educational records, marriage certificates, residency documents, or any document carrying legal weight in any jurisdiction.
• Final-form publication in a language the user does not personally read at native level. AI translation of a manuscript chapter is a draft; the published edition needs a human translator.
• Material in a language Claude is documented as weak in (the smaller African and Indian languages above) when accuracy matters and a native speaker is not available to verify.

The escalation routes for document translation

For any document that needs professional or certified translation, the same Tier 2 and Tier 3 routes from the four interface-translation methodologies apply:

• UN languages: Translators without Borders for humanitarian and non-profit work; a paid professional via ProZ or Smartcat for commercial or institutional work.
• African languages: Lughayangu, White Globe, or Stepes for pilot-grade native-quality work; RushTranslate, ImmiTranslate, Certified Translation India, or Novatia for certified legal work.
• Middle Eastern Arabic and Urdu: a translator approved by the receiving Ministry of Justice for any document going to a Gulf court, embassy, or ministry. The certification regime is jurisdiction-specific (UAE-MoJ-approved is not Saudi-MoJ-approved).
• Sanskrit: traditional commentary plus academic translation; not a Tier 3 problem in the certification sense, but a scholarly one.

What the methodology page does not cover

The verification badges on the language selector refer to the interface translation only. There is no equivalent badge for the Translate module's output, because the Translate module's quality is dynamic — it depends on the language pair, the source register, the length of the passage, and Claude's current capabilities. A user receiving a translation from the Translate module is responsible for judging whether that translation is fit for the purpose at hand, with the guidance above as the honest reference.

How to suggest a correction (interface strings only)

Every panel of UAIE for Microsoft Word carries a "Suggest a translation correction" link in its footer. The link opens a pre-filled email to uaie@gmail.com with the language already named and a template body asking for the current text, the suggested correction, and any notes on register or dialect.

The feedback link is for interface strings — the localisations classified by the four-tier verification model above. It is not a channel for reporting individual document-translation outputs from the Translate module, because those are dynamic and not stored on our side. If a user finds the Translate module weak in a particular language, the right response is to use one of the escalation routes above for that document, and to take the weakness as a signal that the language is in the variable band rather than the strong band.

Native-speaker corrections to interface strings are read by Blaise Martis (Technical Director) and applied in the next patch release. Where a contributor wishes to be credited, the credit appears in the changelog of the release that adopts the correction. Where the contributor prefers to remain anonymous, no credit is attached and the change is recorded simply as "native-speaker correction adopted."

What is not changed by this methodology

Three things are independent of the localisation work and are unaffected by translation status:

• The product's behaviour — every feature operates the same way regardless of selected language. Translate, Plain Language, Multiple Intelligences, the Reading Companion, the Braille module, the Dyslexia Reading Font, and so on are all functionally identical across languages.
• The product's privacy posture — see the Privacy Policy. UAIE has no server, transmits no telemetry, and stores no user content. The translation methodology applies only to the in-product strings the user reads.
• The product's intellectual property — UAIE™ is a trademark of Fiza Pathan Publishing (OPC) Private Limited. The patent applied for in respect of the underlying engine is unaffected by interface language.