How AI document translation works and where it goes wrong

A modern translation tool doesn't look words up in a dictionary. It predicts.¹ The usual method today is neural machine translation, which "uses an artificial neural network to predict the likelihood of a sequence of words". In plain terms, it guesses which words in the new language fit best.

That's why the result usually reads smoothly. It's also why it can be wrong without looking wrong.

Where the guessing goes wrong

The tool can't phone anyone, look at your shop, or ask what you meant. It only has the words you pasted in. Four kinds of trouble show up again and again in business paperwork.

• Words with two meanings. A reference on machine translation says that² "Word-sense disambiguation concerns finding a suitable translation when a word can have more than one meaning." A "bill" can be an invoice or a banknote. A "case" can be a box or a legal matter.
• Missing context. The same source lists as common issues² "the translation of ambiguous parts whose correct translation requires common sense-like semantic language processing or context."
• Casual or unusual wording.² Machine translation has trouble with "non-standard language" compared with standard language. A customer's text full of slang or shortcuts is a risk.
• Rarer languages.¹ The challenges are greater "with languages that have less high-quality data available". A tool that does well in one language pair can do poorly in another.

Smooth doesn't mean correct

Neural tools write fluent sentences. That's the trap. A mistake in a fluent sentence is harder to spot than a mistake in a clumsy one.¹ These systems "tend to produce fairly literal translations." A literal version of a polite request or an idiom can come out cold, odd or just wrong.

Names, amounts and dates are another weak point. The tool might change a name as if it were an ordinary word, or leave out a line. Numbers need your own eyes. Checking a translation without speaking the language covers how.

What you can do to help the tool

You can't fix the tool, but you can feed it better text.

1. Write full sentences.³ Google's help page tells you to "write your word in a phrase or sentence" so the tool can catch your meaning.
2. Cut slang and in-jokes from the original before translating.
3. Keep each sentence to one idea.
4. Use the same word for the same thing all the way through.

These steps lower the odds of an error. They don't remove it. Whether a machine draft is enough, or a person must do the job, depends on the document. AI translation compared with a human translator sets the two side by side.

Two ways to get a translation

Many tools are built just for translation. A newer route is to ask a general chatbot to translate.¹ These generative models "are not trained on the translation task". They learn to predict the next word. Their training text is, in practice, "dominated by English text".

Tools built for translation learn from pairs of texts in two languages. A chatbot learns from general text. Neither one understands your business, so both need a person to check the result.