Warren Weaver and the 1949 Memorandum on Translation that Began Machine Translation

The Weaver memorandum of 1949 is the founding document of the field, and it is a proposal rather than a result. Warren Weaver was a mathematician who had spent the war coordinating research on fire control and had read Shannon's work on information theory as it appeared. The memorandum asks a single question: if a mechanical device can be built to break a cipher, what exactly stops one being built to translate a language?

What Warren Weaver proposed in the memorandum

The document sets out four lines of attack, and it is worth separating them because they had very different fates. Weaver suggested exploiting context to resolve ambiguity, argued that language has formal and logical structure amenable to mechanical treatment, pointed to the cryptographic analogy, and raised the possibility of linguistic universals underlying all languages.

The cryptographic framing is the one everybody quotes. Weaver put it as a thought experiment: a Russian text might be regarded as an English text encoded in a strange script, and the task of translation as decoding. He knew this was not literally true and said so. The point was that it made translation a problem amenable to statistics, and therefore to machines.

The Rockefeller Foundation and why the memorandum carried weight

Weaver was not writing as an academic with a hypothesis. As Director of the Division of Natural Sciences at the Rockefeller Foundation he controlled research funding at a scale that could create a field from nothing, and he had already done so once: the term "molecular biology" is his, and so was much of the money behind it. A memorandum from that desk was an announcement that money would follow.

It did, and quickly. The Weaver memorandum of 1949 was circulated in July; within a year there was a full-time researcher at MIT; within three years the first conference on machine translation was held there; within five years the Georgetown group had put a demonstration in front of the press.

The 1949 argument's long afterlife in statistical translation

The connection between Warren Weaver and machine translation is usually told as a founding anecdote and then dropped. It should not be. The rule-based systems of the nineteen-fifties through the nineteen-eighties took Weaver's second and fourth lines, formal structure and universals, and ignored the statistical one as impractical on the hardware of the day.

When IBM's Candide group returned to the problem in the late nineteen-eighties with aligned parallel corpora and parameter estimation, they were not inventing a new approach. They were implementing the one Weaver had described forty years earlier, on machines that could finally carry it. The statistical era, and by descent the neural one, is the cryptographic analogy coming good.

What the memorandum left out

For a document that founded a field it is strikingly free of method. There is no algorithm, no worked example, no estimate of how much text or computing a system would need. Weaver was not proposing a design; he was arguing that a design was possible and that the problem belonged to mathematics rather than to letters.

The omission mattered. Because the memorandum named four directions without ranking them, each group that picked it up chose differently, and the field spent a decade discovering by experiment which choices were affordable on the hardware of the day. The cryptographic line, the one that eventually won, was the least affordable of the four in 1949 and was therefore the first to be set aside.

Read now, the memorandum is less a blueprint than a bet about where the difficulty lies. Weaver guessed that translation was hard for statistical reasons rather than literary ones. Forty years of rule-writing suggested he was wrong, and the forty years since have suggested he was right.

Next: the 1954 demonstration that turned the proposal into a funded programme, or the full timeline.