Machine translation was announced early, defunded in the middle, and rebuilt from scratch

A public demonstration and the expectations it set

In the mid 1950s a joint demonstration between a university and a computer manufacturer translated a set of prepared Russian sentences into English before an audience of journalists. The system had a small vocabulary and a handful of rules, and the sentences had been chosen to suit it.

The coverage that followed was enthusiastic, and estimates circulated that the general problem would be handled within a few years. It is difficult to reconstruct exactly who promised what, since press accounts and later recollections disagree, but the impression created was of a problem essentially cracked.

The underlying approach was substitution plus rearrangement: look up each word, apply grammatical rules, produce output. On restricted material this works well enough to be impressive. The trouble begins the moment the material stops being restricted.

Ambiguity turned out to require knowledge of the world

Words mean different things in different contexts, and choosing between the meanings frequently depends on facts about physical reality that are nowhere in the sentence. A celebrated example given by one critic turned on a sentence where deciding which sense of a word applied required knowing the relative sizes of ordinary objects.

His argument was uncomfortable and precise: fully automatic high-quality translation would require a machine to have access to general knowledge, and nobody had any idea how to give it that. This was not a complaint about implementation. It was a claim about what the task demands.

Rule-based systems responded by adding more rules, and the rules interacted. Each new rule fixed some sentences and broke others, and the systems became unmaintainable well before they became adequate — a pattern the field would repeat with expert systems two decades later.

An assessment committee wrote down what everybody suspected

By the mid 1960s a committee convened to assess progress reported that automatic translation was slower, less accurate and more expensive than employing translators, and that there was no immediate prospect of that changing. It recommended redirecting money towards fundamental work on language and towards tools that assist human translators.

The immediate effect was severe. Funding in the country that commissioned the report largely stopped, groups dissolved, and machine translation became an unfashionable thing to work on for many years. The episode is now cited as an early example of a field losing patience with itself.

What is less often noted is that the report was substantially correct about the state of the technology, and that its recommendation to fund computational linguistics rather than translation systems shaped a generation of research productively.

Work continued where the domain was narrow enough

Outside the affected funding stream, practical systems appeared in exactly the places the difficulty was smallest. A long-running Canadian system translated weather bulletins, a domain with a tiny vocabulary, rigid sentence patterns and no ambiguity worth the name. It ran for years and did useful work.

That is the durable lesson from the period, and it is a lesson about scope rather than about optimism. Restricting the domain does not merely make the problem easier by degree; it removes the specific difficulty — the need for general knowledge — that made the general problem intractable.

Assisted translation followed the same logic. Storing previously translated segments and offering them again to a human translator is unglamorous, and it produced real productivity gains decades before automatic systems were usable for anything serious.

The moral usually drawn is not quite the right one

The episode is normally retold as a story about hype and its punishment, which is fair as far as it goes. The more useful reading concerns evaluation: the committee’s decisive contribution was to insist on measuring output against what a competent human produced, at realistic cost, on material nobody had selected in advance.

That comparison is what the demonstrations had avoided, and the habit of avoiding it did not die with the report. Any claim about a system’s ability rests on the choice of what it was tested on, and choosing the test to suit the system remains the most common way of producing an impressive result.

It is also worth recording that the critic who argued the task required world knowledge was right about the requirement and wrong about the consequence. Later systems did acquire something functioning as world knowledge, absorbed statistically from enormous quantities of text rather than written down by anyone. The obstacle was real. The route around it was one nobody had imagined.