In 1933, Soviet scientist Peter Troyanskii presented “the machine for the selection and printing of words when translating from one language to another” to the Academy of Sciences of the USSR. Soviet aparchnicks during the Stalin era declared the invention “useless” but allowed Troyanskii to continue his work. He died of natural causes in 1950 – a noteworthy accomplishment for a professor during the Stalin era – but never finished his translation machine.
Early IBM Machine Translation
In the US, during the Cold War, Americans had a different problem: there were few Russian speakers. Whereas the Anglophone countries pushed out countless media to learn English, the Soviet Union produced far less. Furthermore, spoken Russian was different than the more formalized written Russian. As the saying goes, even Tolstoy didn’t speak like Tolstoy.
In response, the US decided the burgeoning computer field might be helpful. On January 7, 1954, at IBM headquarters in New York, an IBM 701 automatically translated 60 Russian sentences into English.
“A girl who didn’t understand a word of the language of the Soviets punched out the Russian messages on IBM cards. The “brain” dashed off its English translations on an automatic printer at the breakneck speed of two and a half lines per second.
“‘Mi pyeryedayem mislyi posryedstvom ryechyi,’ the girl punched. And the 701 responded: We transmit thoughts by means of speech.’
“‘Vyelyichyina ugla opryedyelyayetsya otnoshyenyiyem dlyini dugi k radyiusu,’ the punch rattled. The ‘brain’ came back: ‘Magnitude of angle is determined by the relation of length of arc to radius.'”IBM Press Release
Georgetown’s Leon Dostert led the team that created the program.
Even IBM notes that the computer cannot think for itself, limiting the usefulness of the program for vague sentences. Apparently, nobody at Georgetown or IBM ever heard real Russians speak or they’d know that vague is an understatement with a language that has dozens of ways to say the same word. Furthermore, the need to transliterate the Russian into Latin letters, rather than typing in Cyrillic, no doubt further introduced room for enormous error.
In 1966, the Automatic Language Processing Advisory Committee, a group of seven scientists, released a more somber report. They found that machine translation is “expensive, inaccurate, and unpromising.” The message was clear: the best way to translate to and from Russian, or any other language, is to learn the language.
Progress continued, usually yielding abysmal results. Computers would substitute dictionary words in one language for comparable words in another, with results oftentimes more amusing than informative.
Towards Less Terrible Translations
One breakthrough came from Japan in 1984, which favored machine learning because few Japanese people learned English. Researcher Mankoto Nagao came up with the idea of searching for and substituting phrases rather than words. This yielded far better, but still generally terrible results.
Eventually, in the early 1990s, IBM built on Nagao’s method by running accurate manual translations and building an enormous database analyzing word frequency. The translations became slightly less horrible. This led to “statistical translation” that was significantly less terrible.
As the World Wide Web shrunk the world the need for automated translations grew and the vast majority of these were some type of statistical translation. Subsequently, they continually improved to the point where Google Translate could pretty much help decipher, say, a bill.
Finally, in 2016, neural networks and machine learning (artificial intelligence) started to produce vastly superior machine translations. All the sudden, translations were actually readable. As of 2019, the best online translation engine, German-based DeepL, is entirely AI-powered.