Localization's Dirty Secret: MT Is a Starting Point, Not the Finish Line

Everybody tells you to just run it through DeepL and be done. That's the worst advice you can get if you actually care about shipping a product that doesn't make your customers laugh — or worse, cringe. Machine translation has gotten genuinely impressive, but it's still a tool, not a teammate. Here's how we actually decide when to use it, when to post-edit, and when to keep a human in the loop.

The Scenario: You're Localizing a SaaS Product into Three Languages

Imagine you are a localization manager at a mid-sized software company. You're about to launch in German, Japanese, and Portuguese. Your dev team has already built the product in English. Your budget is tight, your timeline is tighter, and your CEO has heard that "AI can translate everything now." You've got a few options: use raw MT everywhere, use a CAT tool with MT integrated, or hire human translators for everything. But the reality is more nuanced than picking one.

The first thing to understand is the difference between machine translation and computer-assisted translation. MT is fully automated — it replaces the human during the translation phase. CAT tools, on the other hand, are machine-assisted human translation: the translator keeps control and uses a translation memory to reuse previously translated segments. That distinction is everything (Machine translation, Wikipedia).

The Numbers That Change Your Mind

Let's talk quality. When Google launched its neural MT system in 2016, it reduced translation errors by 55–85% on sampled Wikipedia and news sentences compared to the old phrase-based system (GNMT at production scale, Google Research). That's a huge jump. But here's the catch: that was on general-domain text. Your SaaS product has UI strings, marketing copy, and legal disclaimers — not exactly Wikipedia.

DeepL claims that in blind tests with professional translators, its translations were chosen roughly three times more often than Google, Microsoft, or Facebook (DeepL Translator, Wikipedia). Impressive, but note the word "professional translators" — they were judging, not translating. Raw MT output still needs human post-editing, especially for idiomatic expressions, cultural nuance, and domain-specific terminology (Machine translation, Wikipedia).

The Post-Editing Reality Check

So you decide to use MT as a starting point. Now you need to plan for post-editing. ISO 18587:2017 is the standard that specifies requirements for full human post-editing of MT output, and it even defines the competences post-editors must have (ISO 18587:2017, ISO). That's not a casual task — it's a professional job.

How do you measure the effort? Translation Edit Rate (TER) counts the edits needed to turn MT output into a reference translation. A variant, HTER, uses a human-created reference that's closest to the MT output, and it's widely used as a proxy for post-editing effort (Translation Edit Rate, ACL Anthology). In practice, if your HTER is high, you're better off having a human translate from scratch.

Here's a concrete example: you have a 2,000-word marketing page. You run it through MT and get a draft. Your post-editor spends two hours fixing it. If they had translated from scratch, they might have spent three hours. That's a 33% saving — but only if the editor is skilled. If the text is highly creative or brand-sensitive, post-editing might take as long as translating from scratch, and the quality will still be worse. So the decision isn't binary.

What the Big Players Actually Do

Look at the European Commission. They have 24 official languages and their Directorate-General for Translation is one of the largest translation services in the world. In 2022, they produced about 2.6 million translated pages (European Commission translation department). They use eTranslation, a free neural MT service for EU public administrations, SMEs, universities, and NGOs, covering all 24 official EU languages (eTranslation, European Commission). But they don't ship raw MT—they use it as a pre-translation step in their workflow, with human review.

Even WIPO, the World Intellectual Property Organization, built WIPO Translate specifically for patent and technical documents. Their first neural engine, Chinese-to-English, went into production in PATENTSCOPE in October 2016, and they expanded to 10 languages by September 2017 (WIPO Translate, WIPO). Why? Because patents have highly repetitive, formulaic language — a perfect match for MT. But they still have human review for legal accuracy.

How to Decide: A Simple Framework

Here's what we actually do. We classify content into three tiers:

• High-risk: Legal, medical, financial, or anything where a mistake could cost money or lives. Human translation with ISO 17100:2015 processes (ISO 17100:2015, ISO). No MT without full post-editing.
• Medium-risk: UI strings, help articles, product descriptions. Use MT with post-editing, following ISO 18587:2017.
• Low-risk: Internal emails, forum posts, gist of a document. Raw MT is fine.

Quick tip: For medium-risk content, always run a small pilot. Pick 100 segments, have a professional post-edit them, and measure the HTER. If the HTER is above 30%, you're better off with human translation.

Why You Shouldn't Trust BLEU Alone

You might be tempted to evaluate MT quality with BLEU scores. BLEU, introduced by IBM in 2002, compares n-gram matches between machine output and human reference translations (IBM Research). It's simple, but it's increasingly seen as unreliable because it correlates poorly with human judgment. Newer metrics like COMET and BLEURT are becoming standard because they correlate better with human judgment (IBM Research). So don't pick an MT engine based solely on BLEU — use human evaluation or at least COMET.

In our scenario, we'd evaluate the three language pairs with a mix of COMET and human spot-checks. For German, you might find that DeepL performs well. For Japanese, maybe Google's multilingual system, which supports zero-shot translation, works better. Don't assume one engine fits all.

What I'd Actually Do

Here's my recommendation: Build a hybrid workflow. Use a CAT tool that integrates MT, like those that support XLIFF and TMX. XLIFF is an XML-based standard that holds source and target segments, and it's adopted as ISO 21720:2017 (Localization file formats, Microsoft Learn). This lets you leverage translation memory for consistency and MT for new segments.

But don't let the machine run the show. Set up a post-editing step for all customer-facing content, and use a professional post-editor who understands the domain. For your SaaS launch, that means hiring post-editors for German, Japanese, and Portuguese, not just translators. And measure the effort with HTER to iterate.

Finally, don't be seduced by the latest AI models. SeamlessM4T from Meta handles up to 100 languages for speech and text, but it's not a silver bullet for your product's tone. NLLB-200 supports 200 languages and improves quality by about 44% on average, but it's still not human-level (No Language Left Behind, Nature).

The bottom line: MT is a powerful starting point, but it's not the finish line. Treat it as a raw material, not a finished product. Your users will thank you.

Sources

• Machine translation (Wikipedia) - https://en.wikipedia.org/wiki/Machine_translation
• DeepL Translator (Wikipedia) - https://en.wikipedia.org/wiki/DeepL_Translator
• ISO 18587:2017 (ISO) - https://www.iso.org/standard/62970.html
• Translation Edit Rate (ACL Anthology) - https://aclanthology.org/2006.amta-papers.25/
• European Commission translation department - https://commission.europa.eu/about-european-commission/departments-and-executive-agencies/translation_en
• IBM Research - https://research.ibm.com/blog/bleu-nlp-benchmark-anniversary