Machine translation has been declared almost solved twice in 72 years, and the capability it still lacks is the one it lacked in 1954.
On January 7, 1954, IBM and Georgetown University put a machine in front of reporters in New York and translated Russian into English. The system handled about 60 prepared sentences with a vocabulary of 250 words and 6 grammatical rules. The researchers said publicly that the problem would be substantially solved within “three to five years.” The prediction was not a swindle but the honest expectation of serious people, and it came directly out of the war, where Russian had been treated as a cipher and cryptography had worked.
Warren Weaver, who had run American wartime science administration, set out that assumption in a letter in 1947. “This is really written in English, but it has been coded in some strange symbols,” he wrote of an article in Russian, and federal money followed his confidence for the better part of a decade.
The reckoning arrived in 1966, when a committee convened by the national academies reported that machine translation was slower, less accurate and considerably more expensive than employing human translators. The panel, known by its acronym ALPAC, found “no immediate or predictable prospect of useful machine translation.” Government research funding was cut off within a year, and the field entered a winter that lasted roughly 20 years.
The report is routinely described as a historic mistake, and it was not. It was accurate about the systems that then existed, and it was accurate for a reason the discipline needed another 25 years to accept. Meaning is not stored inside the words, but distributed across the context surrounding them, which is why no combination of dictionary and rulebook can reach it.
The revival began around 1990, when researchers at IBM abandoned linguistics altogether and built a method that made no attempt whatever to understand a sentence. It ingested the same text in two languages, most famously the bilingual debates of the Canadian parliament, and computed which sequences tended to correspond to which. The output was frequently absurd and roughly comprehensible, and it comprehensively beat 30 years of careful grammatical engineering.
A second jump came in 2016, when the major systems moved to neural models that translate an entire sentence through a learned representation rather than assembling it out of pieces. Quality improved more in a single release than it had across the preceding decade, and the commercial market for translating manuals, menus and technical documentation has largely disappeared since.
What remains is precisely the part that was never a coding problem, because a translator is required to decide. Which of the available words carries the tone, whether a joke should become a different joke, how a character ought to sound in a language whose register system the original never possessed. Every translated sentence is an argument that somebody has settled, and a system trained on the average of all previous settlements returns the average answer, fluent and plausible, with the decision that would have made it live sanded off.
Gregory Rabassa produced the English translation that carried Latin American fiction to the rest of the world, and he did it without reading the book first.
Rabassa arrived in the trade sideways, serving as a cryptographer during the war, taking a doctorate and teaching for years before he began translating Latin American writers. Almost nobody in the English-speaking world was reading them at the time. His version of Julio Cortázar's “Hopscotch” won the National Book Award for translation in 1967 and opened the door for the boom that followed. When Gabriel García Márquez wanted an English “One Hundred Years of Solitude,” Cortázar told him to ask Rabassa, who was busy. García Márquez waited 3 years for him.
The working method is the part that horrifies people who have never attempted the job, because Rabassa did not read the novel in advance. He translated as he read, reasoning that a reader encounters a novel in ignorance of where it is going, and that his English ought to be produced by somebody in the identical condition. The procedure is indefensible as general advice, and it produced in 1970 the English version that reached millions of readers.
García Márquez said afterwards that the translation was better than his original, a remark usually quoted as a compliment. It is more interesting as a technical claim about what a second version can accomplish that a first cannot, given that the second version already knows how the book ends.
Rabassa translated something like 30 authors across half a century, was paid by the page for most of that time, and was rarely credited on a cover during the first half of his career. Late in life he wrote a memoir about the trade and titled it “If This Be Treason,” after the Italian saying that a translator is a traitor.
He said repeatedly, and without self-pity, that translation is not transfer but decision, and that there is no such thing as the same sentence in another language. There is only a sequence of choices, each of which surrenders something and purchases something. The translator, he wrote, is the person who takes responsibility for the exchange rate.
Two questions closed the month in Tonio Town: what a machine can and cannot do with a language, and who is responsible for the words when nobody signs them. The people who live between two languages answered.
Q. You use it every day. What is it good at?
A. “Everything I do not care about, which is most of a day, and I want to say that without sneering because it has genuinely given me back hours. A form, a notice from the school, an email about a delivery — it is faster than me and it does not get tired. What it gives me back is exactly the text that has no author. The moment there is a person in the sentence, it hands me something smooth that I would never say, and I cannot always tell you why, and the not being able to say why is the whole problem.”
Q. The article says the machine cannot decide.
A. “That is the sentence. In my language there are four ways to say the same request depending on who I am to you, and English has one, and the difference does not disappear — it has to go somewhere. A person moving it puts it into the rhythm, or a word order, or leaves it out on purpose and accepts the loss. The machine picks the average and the average is a stranger. Polite, correct, nobody.”
Q. They said in 1954 that it would be solved in three years.
A. “And they were not fools, they were people who had just won a war with code-breaking and assumed language was a code. It is not a code, it is a habit of a group of people, and it changes on a Tuesday for no reason. I have been in this country nineteen years. I still learn a thing every week that no dictionary has, and I learn it from being wrong in front of somebody kind.”
Q. And the translator who did not read the book first?
A. “I laughed and then I stopped laughing. Because what he was protecting was the reader’s ignorance, which is the one thing you destroy the moment you know the ending. Everybody who has explained a joke has done the opposite of what he did. I would not teach his method. I would teach the reason for it.”
The note from the school said that Sota had been “reserved but engaged” in class, and the family spent forty minutes on it at the kitchen table.
Ken put it through the phone. The phone produced a Japanese sentence that was correct and meant, approximately, that the boy was quiet and also participating.
“That is not what it says,” said Yuko.
“That is exactly what it says.”
“That is what the words say. That is not what the teacher is doing.”
Sota, who is eight and has been in the school for eleven weeks and is the only person in the room who has met the teacher, ate his rice and said nothing.
“What is she doing?” said Ken.
“She is telling us there is no problem,” said Yuko, “before we ask whether there is a problem. Nobody writes ‘engaged’ about a child who is fine. You write ‘fine’.”
They looked at the note for a while.
“Sota,” said Ken. “Do you talk in class?”
“No.”
“Why?”
Sota thought about it with the seriousness of a boy who has been asked a real question at a table where the adults have been arguing about a machine.
“I know the answer,” he said, “but by the time I have it in English, someone else has said it.”
Yuko put the note down.
“Write that,” she said to Ken. “In English. Send it to her. Do not use the phone.”
It took Ken fifty minutes and it was four sentences long, and two of them were wrong.
— The teacher replied the next morning, at length, with a suggestion about giving Sota the question the night before. It has been working since November. Ken has kept his own four sentences, mistakes and all, in the drawer with the passports.