Tahitians know what a missing diacritic marker says about you

A London reseller sent the order on 19 August 2026 under a deadline of 2pm UK time on Friday 21 August. Its four instructions told a proofreader to check the hotel group's in-house Tahitian without preferential changes and to fill the cells marked yellow. Numbering the country entries from row 4 to row 94 came third. The list ended with two edits, a word swap in the chief executive's letter and a two-word deletion from the closing message.

The supplier quoted against 471 words, translation plus proofreading by a second linguist, but London answered with an itemised objection. Proofreading amounted to 197 words, translation to 24 words with the numbering and 15 without it. A word swap in the first tab and a deletion in the second tab accounted for the remainder. Since country and region names already have official equivalents, the reseller argued, little linguistic labour remained.

Item 21019 of the disputed workbook asks hotel employees whether leadership uses their feedback to make improvements. The drafters put two further items to those employees, 1330 on a manager who encourages new skills and 1450 on a manager who builds an inclusive culture. Last comes item 7550, an open field on the year behind them. Those three questions go out to a Tahitian-speaking workforce once a version clears review.

Across 93 Tahitian cells the July column shows no ʻeta and two macrons, against 26 ʻeta and 85 macrons in the version delivered in its place. Tahitian marks the glottal stop with the ʻeta and vowel length with the macron, two phonemic contrasts. An employee handed the string without them would guess at the word the writer meant.

That guess has a measurable duration. Four dropdown entries take 4.1 seconds in either version, 21 syllables in the July text beside 20 in the delivered text, on consonants matching almost exactly. Four glottal closures and ten long vowels are the whole difference between the two lines, which a Tahitian speaker in Papeʻete hears inside the first word.

The July column did produce four entries needing no correction. Aotearoa for New Zealand heads them, a knowledgeable choice ahead of an adaptation built off the English. Completing the four are Hong Kong, Puerto Riko and Singapour. Eight further entries were sound to the consonant and short of diacritics, Raparata becoming Raparāta and Peru becoming Pērū. The in-house team had a model capable of the job.

In Papeʻete the July column writes the territory's own name as Polynésie Farani, a French head on a Tahitian tail. The delivered version writes Pōrīnetia Farāni. The column produces Te mau Hau Amui no Marite for the United States. The delivered Hau ʻĀmui nō Marite drops that article and classifier. Three diacritics separate the two spellings, an ʻeta and two macrons.

Though numbered, nine of the 91 country rows kept an empty cell beside the figure, Fiji included. Thirty-three keep the English untouched, Kazakhstan and Vietnam among them. Fifty-two use letters outside the Tahitian alphabet. Ten stay in source form in the delivered version, three of them Kosovo, Uganda and Hong Kong, leaving the linguist eighty-three to write out.

Onto bare proper nouns the July text hangs articles and classifiers, writing Te mau fenua Maldives and Te mau Philipino. The delivered Maratīvi and Firipino drop them. The machine produced Aiphiti for Egypt, on a ph digraph the alphabet does not use. The linguist replaced that spelling with ʻAifiti. The July column gives the United Kingdom Te Basileia Tahoê, a form whose head word is scriptural Greek.

Instruments of this kind are standard across Tourism and Hospitality, since a group fields a questionnaire in 23 target languages and then compares the results by country. Item 1601 asks corporate staff whether senior leadership sets out a direction and a way forward. Its Tahitian says approximately that the manager is easy to approach, which is the content of item 1462. The drafters wrote item 1610 for non-management roles and item 1627 for guest-facing roles, yet a Tahitian string common to the pair answers for them.

Item 7550 invites the employee to speak freely, though the July column shows no Tahitian against it. Asked by item 21019 whether leadership uses their feedback, that employee would have had the draft's own answer in front of them. The group commissioned a supplier to catch exactly that. The thank-you sentence exists in 22 of the 23 target columns, though the Tahitian column ends on a blank. Researchers on Hawaiʻi Island gave 204 education staff 27 statements to rank, with zero votes for the statement asserting Native Hawaiian grounding.

Reviewing a Tahitian text, a Huri linguist rejected tamāhine and wrote pōtiʻi in its place. Pōtiʻi names a girl past childhood and short of womanhood, while tamāhine leaves her age open. Since the English source gave no signal either way, it was the reviewer who heard the social band and made the change.

Kreutzer and a team of co-authors audited web-crawled multilingual datasets and found systematic mislabelling across the lower-resource entries, alongside content useless for language processing. Their audit also names the corpus behind most low-resource models. Te Basileia Tahoê brings that corpus into a hotel spreadsheet. Kornai called the outcome digital language death in PLOS ONE in 2013.

In French Polynesia the ISPF census of 2017 recorded 25.1% of families mainly speaking a Polynesian language at home. The August 2022 count put the whole territory at 278,786 inhabitants. Volansky and co-authors catalogued the residue under the name translationese, the pattern a machine produces when it oversimplifies or overcomplicates a sentence. Tahitian is the sole Polynesian entry among the 23 target languages in the sheet.

As Naous and co-authors showed, models tuned for an individual non-English language still favour Western entities. Isbister raised the question directly in 2021, on whether monolingual models should give way to machine translation altogether. Once their training data comes from earlier model output, models degrade, a finding Nature published from Shumailov in 2024. A published Tahitian survey becomes corpus. Too little Tahitian exists online to dilute a bad file with better material. A separate byline grades engine output in Tahitian line by line.

A buyer pricing Pacific text off an English word count bids against a tokenizer with very different arithmetic for the two. The reseller sent the supplier a second spreadsheet of 469 words on 22 August. The supplier answered that the new file also needed rebuilding from scratch.

After the client decided on 25 August, the order shrank to two tabs and 237 words, with the chief executive's letter struck out. That letter still opens on a hotel founder's declaration about hospitality, put into a verb meaning to cut. The correspondence records no Tahitian rebuild for it. For the two tabs the group did commission, the delivered Reo Tahiti closing line is ʻUa oti roa te Tītorotororaʻa Mero Pupu o te Ao 2026 iā ʻoe ʻe ʻua tuʻu hia atu tā ʻoe mau pāhonoraʻa.