What is Machine Translation? — one analogy

1The analogy

Think about subtitling a foreign film. Swap each word for its dictionary equivalent and the meaning comes through faintly, but nobody laughs. A joke that only lands in the original country has to be swapped for a different joke the local audience already knows, or the laugh never comes.

Subtitles have their own constraints too — a character limit per line, a time the line has to stay on screen — so a long line of dialogue has to be shortened while keeping the meaning. Dubbing adds one more: it has to match how long the actor's mouth is moving, so even the same meaning gets carried in a different sentence.

A good subtitle isn't one that's faithful to the original wording — it's one that gets the original's job done in this new place too. That's exactly what machine translation is trying to do.

2In detail

It doesn't swap words

It's easy to picture a translator looking up a dictionary and swapping in words. Build it that way and French's "il pleut des cordes" — an idiom for pouring rain — comes out word-for-word as "it's raining ropes," which means nothing at all in English.

Today's translators read a sentence all the way through before starting to translate. First the whole original gets turned into a set of numbers carrying meaning, and while looking at that set, the target-language sentence gets written fresh, one piece at a time. That's why a word in the original and a word in the translation don't line up one to one.

This approach shows its strength especially between languages with completely different word order. Korean puts the verb at the very end; English puts it near the front. Reading all the way to the end before starting is what makes flipping the sentence around possible.

From rulebooks to statistics to now

Early translators ran on grammar rules and dictionaries written by hand. No matter how many rules got added, exceptions piled up faster, and the system broke down the moment a sentence got even slightly complicated.

The next generation used statistics. It gathered a huge pile of documents translated side by side in two languages, counted which chunk tended to turn into which chunk, and stitched together the highest-probability combination. Far better than rules, but you could tell the chunks had been stitched together.

Now it reads the whole sentence and writes it fresh. With a structure that weighs every surrounding word all at once, subject and verb stopped drifting apart even in long sentences. The reason today's translated sentences suddenly read naturally traces back to this shift.

Without context, it has to guess

Translating a single English sentence into Korean forces the translator to decide things the original never specified — whether to address the listener formally or casually, whether there's one listener or several, whether that person is male or female.

When the original carries none of that information, the translator guesses whichever is most common. If formal address showed up more often in similar situations during training, formal address is what comes out. That's why tone can wobble from sentence to sentence when only one line gets fed in at a time.

Feeding in a whole paragraph or conversation at once produces better results. Once who's being addressed becomes clear in an earlier sentence, the tone of the sentences after it falls into line. Telling it the relationships between people, or the right terms up front, steadies things further.

Skill varies a lot by language pair

A translator learns from text written side by side in two languages, so it does well on pairs with a lot of that material stacked up, and poorly on pairs with little. English and French have centuries of paired documents; languages with fewer speakers often have almost none.

Low-resource pairs sometimes get routed through English as a stopover, translated twice. Meaning wears down a little with each hop, so going through two hops drifts further from the original.

Even the same pair splits by field. News articles and everyday conversation come out fine, but writing dense with specialized terms — legal documents, medical instructions — can come back with a confidently wrong term slipped in.

There are ways to check

The simplest check is translating back. Run the translated sentence back into the original language and the spots that wobbled become visible. That said, a clean round trip isn't proof the translation was right — both hops can drift the same way.

Setting key terms in advance also helps a lot. Company names, product names, job titles — anything that has to be written exactly one way — stay consistent throughout a document once they're supplied as a list. For anything that really matters, it's safer to leave the final check to someone who actually knows the language.

3More precisely

Today's translators split into a side that reads the original and turns it into a set of numbers carrying meaning, and a side that writes the target language one piece at a time while looking at that set. This two-sided structure is called an encoder-decoder. Now that large language models handle translation as well, the line between a dedicated translator and a general-purpose model has blurred, and a general-purpose model can often be steered toward a house style just by describing it in the instructions.

The analogy has a limit. A subtitler watches the whole film, knows the characters' personalities and how the story ends, before settling on a single line — a translator usually only sees the sentence or paragraph it was handed. That's why a form of address settled in an earlier scene can flip later on. A subtitler also looks things up or asks when unsure, while a translator never says "I don't know" — it hands back whichever sentence sounds most plausible instead. The automated scores used to grade translation quality also just check overlap with a human translation, so they're weak at catching a smooth sentence with a flipped meaning.

4Try it yourself

5Common misconceptions

It's easy to think a translator looks up a dictionary and swaps words, but actually it reads the whole sentence, carries the meaning across, and writes the target-language sentence fresh.
It's easy to assume a smooth translation is a correct one, but actually a translation with a flipped condition or a dropped negation often reads even more naturally.
It's easy to think every language pair performs about the same, but actually quality varies a lot depending on how much paired material exists for that pair.

7One-line summary

In shortMachine translation isn't swapping out words — it's rewriting a sentence so it does in another language the same job it was doing in the original.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02