This is a timeline of machine translation from the patents of 1933 and the memorandum of 1949 through to translation inside general language models. A machine translation timeline is only useful if it is checkable, so each entry names the document or system it rests on rather than summarising a summary.
A Machine Translation Timeline: from 1949 to Neural MT, Decade by Decade
Seventy-six years in five movements. Read in sequence it is not a steady climb but two collapses and two revivals, and the idea that revived it last was the one written down first.
The proposal, 1933–1954
1933
Two patents, independently
Georges Artsrouni in France and Petr Troyanskii in the Soviet Union each patent a mechanical translation device. Troyanskii’s design anticipates an interlingua; neither is built at scale.
1949
The Weaver memorandum
Warren Weaver circulates a proposal to about two hundred correspondents, framing translation as cryptography and statistics. It sets the agenda and the funding follows.
1952
The first conference
Yehoshua Bar-Hillel convenes the first conference on machine translation at MIT. The participants disagree about almost everything except that the problem is tractable.
1954
The Georgetown–IBM demonstration
Sixty Russian sentences into English on an IBM 701, on 250 words and six rules, in front of the press. Funding expands sharply.
Doubt and retrenchment, 1959–1966
1959
Bar-Hillel changes his mind
The man who convened the first conference argues that fully automatic high-quality translation is unreachable without knowledge of the world, using the sentence "the box was in the pen" to show that disambiguation needs facts, not grammar.
1966
The ALPAC report
The committee finds machine output slower and costlier than human translation, with no translator shortage to justify it. United States funding stops for roughly twenty years.
The rule-based decades, 1968–1985
1968
Systran founded
Peter Toma builds on the Georgetown work. Systran will run for the United States Air Force, then the Commission of the European Communities from 1976, and in some form for fifty years.
1976
Météo goes live
The TAUM group in Montreal puts a system into daily production translating weather bulletins. A narrow domain with fixed vocabulary is where rule-based translation genuinely works.
1978
Transfer and interlingua mature
Japanese and European projects settle the architecture: analyse the source, transfer at some level of abstraction, generate the target. The debate is how deep the transfer should go.
The statistical turn, 1988–2006
1988
IBM’s Candide group
Aligned parallel corpora and parameter estimation replace hand-written rules. The move from rule based to neural machine translation begins here, with the statistical step in between.
1993
The IBM alignment models
Models 1 to 5 formalise word alignment and become the shared baseline for a decade of work.
1999
The summer workshop
Johns Hopkins reimplements the IBM models openly, and statistical machine translation becomes something a research group can start without IBM’s resources.
2006
Phrase-based systems at scale
Public web translation moves to statistical phrase-based systems. Quality is uneven and coverage is enormous, which turns out to be the trade users prefer.
The neural turn, 2014–present
2014
Sequence to sequence
Encoder–decoder networks translate without an explicit alignment stage. Attention arrives the following year and removes the fixed-length bottleneck.
2016
Production systems switch
Large public services replace phrase-based pipelines with neural ones. The change in fluency is immediate and the change in error type is more interesting: mistakes become plausible rather than broken.
2017
The transformer
Attention without recurrence. The architecture that follows dominates translation and then leaves translation behind for general language modelling.
2020s
Translation inside general models
Translation stops being a separate product category and becomes one capability of systems trained for everything. Weaver’s statistical framing, seventy years on, is the one that held.
Why the 1949 framing took sixty-five years to win
The statistical idea was in the 1949 memorandum and then ignored for forty years. The reason is not stubbornness. Estimating translation from data requires parallel text in quantity and the compute to count it, and neither existed until the late nineteen-eighties. The rule-based decades were not a detour taken in error; they were the only approach the hardware of the time could carry, which is why a timeline of machine translation reads as alternating eras rather than as accumulation.
What the statistical decades changed
The statistical turn replaced the question. Rule-based systems asked what the correct translation is. Statistical systems asked which translation is most probable given everything previously seen. That reframing is what made the move from rule based to neural machine translation possible, because a neural model is a far better probability estimator over the same question.
How to read this machine translation timeline: the neural decades
A machine translation timeline that stops at quality scores misses the more useful observation. Rule-based output failed visibly: word salad, untranslated tokens, obvious breakage. Neural output fails fluently, producing well-formed sentences that are wrong. The practical problem moved from producing readable text to detecting confident error. Any machine translation timeline written from here on has to account for that shift.
When does the machine translation timeline actually begin?
1933, if patents count: Artsrouni in France and Troyanskii in the Soviet Union filed independently. 1949 if a research programme is the criterion, because that is when funding and an agenda appeared together.
Why is 1966 treated as a break rather than a slowdown?
Because United States federal funding did not taper, it stopped. Groups dispersed and most of a research generation left the field. Work continued in Canada, Europe and Japan, so the break is national rather than global.
Was the statistical turn a rejection of linguistics?
In practice yes, in principle no. Candide's results came from counting rather than from grammar, and the field's centre of gravity moved accordingly. The linguistic work did not stop; it stopped being the thing that shipped.
How quickly did neural systems displace statistical ones?
About three years from first publication to production deployment at scale, which is fast by the standards of anything else on this timeline. The rule-based-to-statistical transition took closer to fifteen.