The History of Machine Translation: from the 1930s to ALPAC

The history of machine translation does not begin with computers. It begins with two engineers working independently in 1933, and it arrives at its first crisis within thirty-three years. What follows is the founding sequence; the full timeline carries it through to neural systems.

The 1930s: Troyanskii's system — the first step towards machine translation

In 1933 Petr Petrovich Troyanskii, a Soviet engineer, received a patent for a machine that would translate between languages by mechanical means. Georges Artsrouni filed for something similar in France the same year, but Troyanskii's design is the more interesting of the two because it contains, in rough form, the architecture the field would rediscover twenty years later.

His proposal separated the work into three stages. A human editor who knew only the source language would reduce each sentence to base forms and mark its grammatical roles. The machine would then substitute the target-language equivalents. A second editor, knowing only the target language, would turn the result into readable prose. That is analysis, transfer and generation, described before any machine existed that could perform the middle step.

The design was reviewed, judged impractical, and forgotten. It was rediscovered in the nineteen-fifties, by which point the field had arrived at the same structure independently.

1949: the first computer-based MT developments

Warren Weaver's memorandum of July 1949 turned an engineering curiosity into a funded research programme. Weaver directed natural sciences at the Rockefeller Foundation, had followed Shannon's work on information theory, and knew what wartime cryptanalysis had achieved by mechanical means. His argument was that translation might be approached the same way: a text in one language treated as a text in another, encoded.

He knew the analogy was imperfect and said so. Its value was that it reframed translation as a problem of statistics and structure rather than of literary judgement, and therefore as something a computer might attempt. Within a year MIT had a full-time researcher on it; within three, the first conference on the subject.

The full text and its four proposed lines of attack are covered on the memorandum's own page.

1950s and 1960s: the Georgetown experiment and the ALPAC report

On 7 January 1954 a collaboration between Georgetown University and IBM translated around sixty Russian sentences into English on an IBM 701, in front of reporters. The system held 250 words and six grammar rules, and the sentences had been chosen in advance. The technical write-ups said so plainly; the press coverage did not, and predictions of solved translation within five years followed.

The decade after was the field's expansion: groups in the United States, the Soviet Union, Britain and Japan, working through the direct, transfer and interlingua architectures. It was also the decade in which Yehoshua Bar-Hillel, who had convened the first conference, concluded that fully automatic high-quality translation was unreachable without knowledge of the world.

In November 1966 the Automatic Language Processing Advisory Committee reported to the National Academy of Sciences. It found machine output slower and more expensive than human translation once post-editing was counted, and no shortage of translators to justify the spending. United States funding stopped, and did not seriously return for about twenty years. The findings and their consequences are set out on the ALPAC page.