The History of Machine Translation: Timeline, Milestones and Research Archive
Seventy-six years of trying to get one language out of another by machine. The history of machine translation runs from a 1949 memorandum through the demonstration that sold the idea, the report that stopped it, and the statistical and neural methods that revived it. This machine translation research archive serves the recovered papers at the addresses scholarship already points to, and redirects the rest to the account of their subject.
The material is drawn from the published record, including the survey work of W. John Hutchins, whose papers are reproduced here as they are recovered. A profile of a renowned baccarat player sits among the general-interest notes.
The milestones in the history of machine translation
Six milestones carry most of the story. Every later argument in the field is a reply to one of these milestones.
1949The Weaver memorandum
Warren Weaver proposes treating translation as a problem of cryptography and statistics, and circulates the argument to roughly two hundred correspondents. Almost every later position in the field is a reply to it.
The Georgetown–IBM experiment
Forty-nine Russian sentences rendered into English on an IBM 701, on six grammar rules and two hundred and fifty lexical items. The press coverage, more than the result, is what funded the following decade.
Bar-Hillel's objection
A survey argues that fully automatic high-quality translation is not merely unsolved but unreachable without knowledge of the world. The objection has never been fully retired, only worked around.
The ALPAC report
The Automatic Language Processing Advisory Committee reports to the National Academy of Sciences, and United States funding stops for roughly twenty years. It is the single most consequential document in the field.
The statistical turn
IBM's Candide group returns to Weaver's 1949 framing, replacing hand-written rules with aligned parallel corpora and parameter estimation.
Sequence to sequence
Encoder–decoder networks with attention displace phrase-based statistical systems within three years of first publication.
A timeline of machine translation, decade by decade
Read in sequence, the history of machine translation is not a steady climb. Its milestones fall into two collapses and two revivals: a decade of optimism ended by the ALPAC report in 1966, a long quiet period of rule-based systems sold to institutions rather than to the public, a statistical turn in the late nineteen-eighties that returned to the 1949 framing, and a neural turn after 2014 that displaced the statistical systems in about three years.
The scroll-through timeline lays the whole sequence out with the primary source behind each entry, so a claim about what happened in 1954 or 1966 can be checked against the document that made it rather than against a summary of a summary.
What the research archive holds
Two collections sit underneath the narrative. The publications bibliography reproduces the survey and conference papers in full, each at the address that external citations already use. The compendium of translation software catalogues the commercial and research systems themselves, from the institutional installations of the nineteen-seventies through the desktop packages of the nineteen-nineties.
Writing the history of machine translation depends on those two collections staying reachable. A machine translation research archive is only worth keeping if its addresses stay put. Where a paper is cited by an encyclopaedia entry or a conference proceeding, that citation resolves here to the document itself, not to a landing page that explains the document once existed.
Why the record keeps being rewritten
Almost everything popularly known about this field comes from press coverage rather than from the papers. The 1954 demonstration is remembered as a breakthrough because that is how it was reported, not because the technical write-up claimed it; the 1966 report is remembered as a verdict that machine translation was impossible, which is not what it says. Both distortions are old enough to have been copied into textbooks, and both are correctable against documents that still exist.
That is the use of keeping an archive rather than a summary. A claim about rule counts, vocabulary sizes, funding figures or dates can be checked here against the paper that first made it, and where a later account disagrees with a contemporary one, the disagreement is visible instead of averaged away.
When did machine translation begin?
As a documented research programme, in 1949, with Warren Weaver's memorandum on translation. Proposals to mechanise translation appear earlier, in patents filed in 1933, but the memorandum is what set the research agenda and drew funding.
What did the ALPAC report actually say?
That machine translation was slower, less accurate and more expensive than human translation, and that there was no shortage of translators to justify the spending. It recommended funding computational linguistics instead. The consequence was roughly twenty years without United States support.
Was the 1954 Georgetown–IBM demonstration genuine?
The sentences really were translated by the machine, but they were selected in advance and the system held six rules and two hundred and fifty words. The demonstration was honest about its scale in the technical write-ups and much less so in the press coverage.
How are the papers here related to John Hutchins?
W. John Hutchins wrote the standard surveys of the field's history and compiled the catalogue of translation systems. His bibliography is reproduced here in full, and his papers are cited throughout as sources rather than presented as the voice of this site.