Two older contents pages sit alongside this one and are kept at the addresses that cite them: the introduction to machine translation and machine and manual translation.
The history of automatic translation is usually told as steady progress. It is not. It is two collapses and two revivals, and the idea that revived it most recently was the one written down first, in 1949. Start with the full timeline for the sequence, or take the turning points individually below.
Why the turning points are the ones listed
Nine entries is a choice, and the criterion is narrow: each one changed what researchers did next, and the change is visible in the record rather than asserted afterwards. A patent that described analysis, transfer and generation before any machine could perform the middle step qualifies. A memorandum that moved the question from literary judgement to statistics qualifies. A demonstration of sixty sentences qualifies not because the system was good but because the coverage of it set the funding for a decade.
Several famous episodes are deliberately absent. The individual rule-based systems of the nineteen-seventies matter enormously to the people who built them and very little to the shape of the field, because each was a better execution of an agreed architecture rather than a change of direction. The same applies to most benchmark results: a score that improves on the previous score is progress within a paradigm, not a turning point.
The shape this history actually has
Told properly it is not a climb. Optimism runs from 1949 to about 1959, doubt from 1959 to 1966, retrenchment from 1966 to roughly 1976, a long period of careful rule-based engineering to the late nineteen-eighties, and then two changes of method in twenty-six years that each displaced the previous one within about three years of first publication. Anyone reading only the first and last decades would conclude the problem was solved steadily, and would miss that the idea which eventually worked was written down in the first year and then abandoned as impractical for forty.
The two collapses are also not symmetrical. The first was a funding decision taken on measured evidence about cost and speed, and it was defensible on those terms. The second, the fading of rule-based work in the nineteen-nineties, was nobody's decision: the corpora and the machines arrived, the statistical results overtook the hand-written grammars, and the field moved without anyone announcing it. Only one of the two produced a report worth arguing about, which is why the first is remembered as the crisis and the second is barely remembered at all.