Picking an MT Engine When Your Language Isn't on the List

The MT Market Is Booming — and Still Leaving Millions Behind

The machine translation market keeps expanding. Google Translate covers 109 languages. DeepL covers 24 but outperforms Google on accuracy for the languages it does support — by roughly 13%, according to independent research. Microsoft Translator handles 103. These numbers are genuinely impressive, and the competitive landscape between these engines is worth understanding if you manage any localization workflow.

But here is the number that matters more for US healthcare systems, school districts, and local government agencies serving Pacific Islander communities: zero. That is how many major MT engines support Chuukese or Pohnpeian.

Not limited support. Not low-quality support. Zero.

For agencies serving Micronesian populations in Hawaii, Guam, Arkansas, or the Pacific Northwest, that gap is not a minor inconvenience. It is a compliance risk, a patient safety issue, and a community trust problem all at once.

What the MT Engine Comparison Actually Shows

Before getting into the rare-language problem, the mainstream MT comparison is worth taking seriously. If you are localizing content that does not involve Chuukese or Pohnpeian, choosing the right engine matters a lot depending on your content type and language pairs.

For Spanish, French, or Korean content, the engine choice is a real decision with measurable quality implications. DeepL's glossary feature, for instance, lets you lock in specific terminology — useful when clinical or legal precision matters. Microsoft's neural machine translation trains a single network on source and target text directly, which produces more fluent output than older phrase-based statistical systems.

These are meaningful distinctions. But they only apply if your target language exists in the engine's training data.

When Your Language Has No Engine

Chuukese is spoken by an estimated 45,000 to 50,000 people, many of whom live in the United States under the Compact of Free Association. Pohnpeian speakers number around 30,000. Neither language appears in any major commercial MT engine. Neither has a substantial digitized corpus that would make training a custom model straightforward.

This is not an obscure academic problem. Chuukese and Pohnpeian speakers are disproportionately represented in Medicaid populations in Hawaii and Oregon. School districts in Springdale, Arkansas have significant Marshallese and Micronesian enrollment. When a hospital deploys a pre-translation workflow using MT as a first pass — a common cost-reduction strategy — and that workflow has no output for a patient's language, the result is either a blank page or, worse, an inappropriate fallback to a related but distinct language.

Using Marshallese output for a Chuukese speaker is not a minor quality issue. These are separate languages. It is the equivalent of sending an Italian document to a Spanish-speaking patient and calling it close enough.

Why Hybrid Workflows Are Non-Negotiable for These Languages

For MT-supported languages, the standard argument for machine translation post-editing (MTPE) is efficiency: let the engine do the first draft, have a human editor refine it, and save 30–50% on per-word cost. That argument is sound when the engine produces a usable draft.

For Chuukese and Pohnpeian, there is no draft to edit. The workflow has to start with a human translator.

This changes the economics, but it does not change the legal or ethical obligation. Under Section 1557 of the Affordable Care Act and Title VI of the Civil Rights Act, healthcare entities receiving federal funding must provide meaningful access to individuals with limited English proficiency. "Meaningful access" has been interpreted by OCR to include written translation of vital documents. A notice of patient rights, a discharge summary, a consent form — these require accurate written translation regardless of whether an MT engine can assist.

The practical implication for a health system procurement team is this: if your language access vendor is quoting you MT-assisted rates for Chuukese or Pohnpeian, ask them specifically which engine they are using and what the BLEU score baseline is for that language pair. If they cannot answer, they are either using a general-purpose engine inappropriately or padding a fully human translation cost with an MT label.

What a Responsible Workflow Actually Looks Like

For healthcare and government clients working with Chuukese or Pohnpeian communities, the workflow that actually holds up under compliance review looks like this:

Translation memory first. A properly maintained TM for Chuukese means that standard phrases — appointment reminders, rights disclosures, medication instructions — get reused accurately across documents. The efficiency gain comes from consistency, not from MT.

Subject-matter-matched human translators. Chuukese medical translation requires a translator who understands both clinical terminology and the dialect variation between Chuuk Lagoon communities. A bilingual individual with no translation training is not a substitute, and neither is a Chuukese translator with no healthcare background.

Glossary control. The same concept that makes DeepL's glossary feature valuable for Spanish applies to human-translated rare languages. Locking in approved terminology for conditions, procedures, and legal rights language ensures consistency across a health system's document library.

Back-translation QA for high-stakes content. Informed consent documents, behavioral health intake forms, and pediatric care instructions should go through back-translation review before deployment.

None of this is exotic. It is standard localization practice applied to a language context where MT cannot shortcut the process.

The Takeaway

MT engines are genuinely useful tools, and understanding the differences between Google, DeepL, Microsoft, and others will help you make better decisions for the majority of your language pairs. But the decision framework changes completely when you are working with a language that has no MT support.

For Chuukese and Pohnpeian, the honest answer is that human translation is not a premium upgrade — it is the only option. Agencies that build language access programs around MT cost assumptions for these languages will either underserve their communities or face compliance exposure when audited.

If you are building or auditing a language access program that includes Pacific Islander populations, we are glad to talk through what a realistic, compliant workflow looks like for these specific languages.