Machine translation changed what it measures. Patent documents did not.

For anyone who uses translation in professional work, the useful question about the history of machine translation is not how the technology changed. It is: at each stage, what was the system built to be good at?

Rule-based translation: correct within what was written down

The earliest systems relied on bilingual dictionaries and grammar rules written by people. Whoever built the system had to tell the machine which word corresponded to which, how word order changed, and when an element was dropped or added.

What this approach optimised for was conformity to rules. A result counted as good when it did what the dictionary and the rule set required.

But real language is full of exceptions, ambiguity and structures nobody anticipated. Faced with a sentence outside what had been written down, the machine had only the existing rules to apply — even when the result came out stiff or wrong.

Statistical translation: pick what usually appears

The next shift stopped asking “what rules should this language follow” and started asking “in the translations we already have, how do people usually put it?”

The system learns from a corpus of parallel text: source sentences and their translations. From that it estimates which phrases tend to travel together, which orderings are common in the target language, and which candidate is more probable.

Google Translate straddles exactly this changeover: most of its languages initially ran on a rule-based engine, then moved to a statistical system Google built itself.

This generation optimised for what is common. Widely used phrasing has a great deal of data behind it; rare usage, or usage confined to one narrow field, has far less.

Neural translation: making the whole sentence hold together

Statistical translation works with words and phrases and assembles them into a sentence. Neural translation learns to represent the source sentence as a whole and produces the target sentence from the relationships across all of it. The machine is not only asking what each piece should be replaced with, but what part each piece plays.

The change produced a step you could feel: smoother sentences, more natural ordering. Google Translate moved to neural translation, and DeepL launched during the period when the method came into wide use.

The standard this generation is judged by is fluency at sentence level. Readers began to assess output by how it reads: does it flow, does it sound like something a person wrote.

Large language models: translation as a kind of writing

Large language models were not built specifically to translate. The same model answers questions, summarises, rewrites, explains — and moves text from one language into another. Asked to translate, it does so with the same text-producing ability it uses for everything else.

The result is that the capacity to make text read well is stronger than it has ever been.

Across four stages the prevailing standard drifted: from rule-conformity, to resembling what is common, to fluency and naturalness. All four are measured on ordinary prose.

A patent specification is read against a different yardstick.

Two standards at once

A patent specification is both a technical document and a legal instrument. Its terminology has to be right in both senses: right technically to a person in the field, and right legally against the scope of protection it establishes.

A general-purpose system optimising for naturalness has no way of knowing that the second standard exists.

Why “inconsistent” costs more than “wrong”

This is the counter-intuitive part, and the most important thing about translating patent documents.

A term translated wrongly is comparatively easy to deal with: you know which word is wrong, so you find it and replace it throughout.

A term translated inconsistently is not. One concept appears under three different names, scattered across the description and the claims. To fix it you first have to find all three — and you do not know what you are looking for. There is no wrong word to search for, only places that ought to match and do not.

The consequence is not merely wasted effort. In a document where the scope of protection is fixed by its wording, one concept carrying two names can be read as two concepts.

That is why consistency, rather than fluency, is the thing worth measuring first. And it is precisely what none of the four generations above was built to optimise.

”Correct” also depends on who is reading

Even once the wording is consistent, a question remains: consistent with which glossary?

Many product lines, many applicants and many IP firms keep house terminology of their own — used in earlier filings, settled in correspondence with examiners, sometimes written into a contract. For those files, “correct” is not the most common rendering. It is the one used last time.

No general-purpose system can know that glossary. It exists only on your side — which is why the translation needs a person in control, not just a better machine.

The same difficulty in miniature: how to write the names of measurement units in the target text. This is a common stumbling block for less experienced translators, and it has no universal answer — see Unit names in your language.

What Pattrans is built around

Not around producing smoother prose — the field has done that well across four generations — but around giving the second yardstick somewhere to stand:

• The termbase is the centre, not an add-on. One set of terms applied consistently across the description, the claims and related documents in the same patent family — see Keeping your termbase clean and trustworthy.
• Review the terms before translating. Fix a term once, before it is applied to the whole document, instead of hunting three variants of it afterwards — see Review terms with Deep mode.
• You decide last. A bilingual copy to review in Word, imported back when you are done — see Review the bilingual translation in Word.

The translation still needs you to read it and take responsibility for it. What changes is that you control the thing most worth controlling, and you control it before it multiplies across the document.