Four thousand years of writing in two languages

Clay: two columns of words

Scribes in Mesopotamia kept word lists in two columns, pressing Sumerian on one side and Akkadian on the other. By the early second millennium BC Sumerian was no longer spoken, but it was still the language of learning and religion, so the lists taught one language through the other. Lugal is king in Sumerian; šarru is king in Akkadian. They are among the oldest bilingual dictionaries we have.

Stone: the same decree three times

In 196 BC priests at Memphis issued a decree in honour of Ptolemy V and had it carved in three scripts: hieroglyphs for the temples, Demotic for daily use and Greek for the government. French soldiers found one copy at Rashid in 1799. Because the Greek could be read, the stone became the key to the other two, and Jean-François Champollion announced his decipherment of the hieroglyphs in 1822. A translation made for readers in 196 BC turned out to be most useful to readers two thousand years later.

The quill: translation as scholarship

In twelfth-century Toledo, scholars worked in teams to turn Arabic science and philosophy into Latin, often through a spoken intermediate language. Monasteries copied and translated scripture by hand, a page at a time. A mistake copied once could be copied for centuries, so reviewers checked the work against the source: an early form of quality assurance.

The pen and the typewriter: translation as business

Trade and diplomacy made translation a paid profession. Letters, contracts and patents crossed borders in carbon copies. In 1954 Georgetown University and IBM showed a computer translating more than sixty Russian sentences into English, and machine translation went from fiction to a research programme.

Memory: never translate the same sentence twice

In the 1990s translation memory tools became standard on translators’ desktops. Every approved sentence was stored with its translation, so the next time it appeared the match was waiting, with the differences marked. Termbases did the same for product names and technical terms. Consistency stopped depending on the translator’s memory.

AI and people

In 2016 neural machine translation replaced phrase-based systems in the main public engines, and fluent drafts became cheap. Large language models followed, able to take instructions about tone, audience and length. Speed is no longer the hard part. The hard part is knowing that a translation is right: that it uses the client’s terms, fits the space, keeps the tags and says what the source says. That still takes rules, checks and a person who signs it off.

That is the part LanguageOps is built for. AI drafts with your memories and glossary, QA checks every line, and a linguist approves it before it goes back to you in the format you sent.