The History of Machine Translation: From Troyanskii to the Present

Machine translation looks today like an ordinary feature on your phone. You press a button and the text crosses into the other language. Behind that simplicity, though, lies roughly ninety years of searching, and most of that search is made up not of successes but of cancelled budgets and lowered expectations.

Contents

The history of machine translation is a good illustration of technology failing to advance in a straight line. The field went through two great waves of optimism and was braked hard twice. For an organisation that wants to understand what today’s tools can and cannot do, knowing that cycle is directly useful.

Below we sum up the line running from the 1930s to the present in six stops: Troyanskii’s machine that stayed on paper, the Weaver memorandum, the Georgetown experiment, the ALPAC Report, the Météo system, and the move from statistics to neural networks.

Where does the history of machine translation begin?

At the base of machine translation lies a simple but ambitious assumption: the passage between languages, if defined well enough, is an operation that can be reduced to rules. The moment that assumption is accepted, translation ceases to be a problem of art and becomes a problem of engineering.

The idea was born before computing machines became widespread. Dictionary lookup, the application of grammatical rules and the reordering of words all looked like tasks that could be carried out mechanically. All that was missing was a machine fast enough.

What history went on to show was this: what was missing was not speed. What was missing was that meaning cannot be defined independently of context. Roughly seventy years of the field’s development are largely the story of wrestling with that single problem.

For that reason the history of the field is good for more than being a chronological list. Each period gave a different answer to the question “what can we hand over to the machine?”, and each answer ran into a limit of its own. When an organisation decides today where machine assistance should stop, it is in fact answering the current form of the same question. Knowing where past attempts stumbled takes that decision out of guesswork and ties it to a measure.

Troyanskii: the machine that stayed on paper

One of the first detailed designs for automatic translation belongs to Petr Petrovich Troyanskii, who worked in the Soviet Union in the 1930s. Troyanskii designed a mechanical apparatus to work across four languages: the source sentence would first be analysed with grammatical markers, then the machine would find the word equivalents, and finally a human would correct the output.

The striking thing about this design is that it places a human at the beginning and at the end of the process. Troyanskii never claimed that the machine could translate on its own. The logic of the workflow now known as “pre-editing” and “post-editing” was already there in that very first design.

The design found no response in the conditions of its day and was forgotten for many years. Troyanskii’s work was reassessed only in the decades that followed (Hutchins & Lovtskii 2000). The reason he is counted among the founding figures of the field is not that he produced a working machine but that he broke the problem into the right parts.

The Weaver memorandum: the codebreaking analogy

The real starting signal for the field is the text written by Warren Weaver in 1949 and generally known as the “Weaver memorandum” (Pietrzak & Kornacki 2021, Chapter 3). Drawing on the codebreaking work that came of age during the Second World War, Weaver proposed that a text in a foreign language could in fact be seen as an “encoded” text.

The memorandum offered no technical solution; it offered a framework. The ideas that ambiguity of meaning could be resolved through the immediate context, that languages might share a common logical core, and that statistical methods might prove useful all came out of it.

The effect of the text was swift. Within a few years machine translation groups had been set up at more than one university and the field had come into a substantial public budget. The cautious phrasing in Weaver’s own text was largely disregarded.

The Georgetown experiment: a 250-word promise

On 7 January 1954 Georgetown University and IBM staged a joint demonstration. A computer translated Russian sentences into English and was presented to the press. The demonstration shaped public perception of the field for decades to come.

The real scale of the demonstration, however, was rather small: the system ran on a vocabulary of about 250 words and a very limited number of grammatical rules (Pietrzak & Kornacki 2021, Chapter 3). The sentences translated had been selected in advance and prepared to suit what the system could handle.

Even so, the press gave the impression that the problem would be solved outright within a few years. That gap between expectation and capability was also the underlying cause of the field’s next crisis. The distance between a demo and a product is one of the oldest traps in the history of technology.

The ALPAC Report: a long stagnation

The ALPAC Report, published in 1966, assessed the results of more than a decade of investment. Its conclusion was harsh: machine translation as it then stood was no faster, no cheaper and no more accurate than human translation (Pietrzak & Kornacki 2021, Chapter 3).

After the report, public funding for the field in the United States of America was largely cut off. In the literature the period is often referred to as “the lost years”; machine translation research was confined to a narrow circle for around twenty years.

It should be added, though, that the report was not wholly negative. ALPAC recommended investment in tools that assist humans and in computational linguistics. The intellectual origin of the computer-assisted translation tools we use today rests largely on that recommendation.

Météo: the first system to work in a narrow domain

The field’s first clear commercial success came with the Météo system, brought into service in Canada in 1976 (Pietrzak & Kornacki 2021, Chapter 3). The system translated English weather bulletins into French and stayed in actual production for years.

The reason for Météo’s success was not the superiority of the technique but the narrowness of the domain. Weather bulletins use a limited vocabulary, formulaic sentence structures and a language almost entirely free of metaphor. In other words, the ideal text for a machine.

The lesson still holds today: machine translation works better the more predictable the text is. The difference between a product catalogue and a literary text arises less from the quality of the engine than from the nature of the text.

From rules to statistics, from statistics to neural networks

From the 1990s onwards the field changed direction. Instead of writing grammatical rules out by hand, statistical models that learned probabilities from large bodies of human-translated text came to the fore. For roughly twenty years this was the industry’s main line.

From the middle of the 2010s, models based on artificial neural networks became widespread. Because these models process a sentence as a whole rather than piece by piece, the fluency of the output rose markedly. In recent years the spread of large language models has strengthened the trend further.

The underlying tension nevertheless remains where it was. A fluent sentence does not mean a correct sentence. The most dangerous error in machine output is the sentence that looks flawless while being wrong, because only someone who knows the subject will catch an error of that kind.

Industry data also show the usage side of this transition. In a 2012 industry survey, the proportion of companies saying they had at least tried machine translation was measured at 44.02 per cent, and the proportion of freelance translators at 55 per cent (DePalma 2012). In other words, the tool was on the industry’s agenda even before neural models became widespread; what changed was not the existence of use but its intensity and the role granted to it.

Period	Defining event	What remains today
1930s	Troyanskii’s mechanical design	The idea of human pre-editing and post-editing
1949	The Weaver memorandum	The seed of the statistical approach
1954	The Georgetown experiment	The lesson of the gap between demo and product
1966	The ALPAC Report	Computer-assisted translation tools
1976	The Météo system	The power of narrow-domain specialisation
After 1990	Statistical, then neural models	Data-driven translation and the post-editing profession

The Türkiye context

The whole of this history was written in a world where the source and target languages were mostly English, French and Russian. Turkish did not sit on the main line of that trajectory. The consequence that carries through to today is that the processed language resources available for Turkish are more limited than they are for the major languages.

To this is added the structure of the language. Turkish is agglutinative; a great many inflected forms can be derived from a single stem. Word order is also relatively free; the element to be emphasised shifts position within the sentence. For statistical methods, those two features increase the number of patterns to be learned considerably.

In practice this means that the same engine may not deliver in the English-Turkish pair what it delivers in the English-German pair. The commonest problems in Turkish output are mismatched suffixes, wrong person assignments arising from dropped subjects, and broken predicate agreement in long sentences. These are qualitative observations, but they are consistent with experience in the field.

What does this mean for you?

For the party buying translation, this history has a single summary: machine translation is not a switch but a setting. The decision about which texts you turn it on for and which you turn it off for determines the outcome directly. You can follow this order when deciding:

Sort your text by purpose: is it to be read for comprehension, is it to be published, or will it carry legal consequences?
Measure how predictable the text is; formulaic and repetitive texts suit the machine better.
Establish the confidentiality position; ask in writing whether the engine can be switched off.
Write the level of human oversight into the contract at the outset: a light check or a full correction?

With us this distinction is drawn at the start of the project and stated openly in the quote; at what stage machine assistance is used is not kept from you. If you are not sure which category your text falls into, you can upload your document and ask for an assessment, or look through our knowledge centre for details of the process.

Frequently asked questions

When was the idea of machine translation first put forward?

The first concrete designs for automatic translation go back to the 1930s. The Russian engineer Troyanskii drew up the design for a mechanical translation apparatus working in four languages; that design, however, was never put to widespread use.

Why did machine translation research stop for a period?

The ALPAC Report of 1966 concluded that machine translation as it then stood was no faster and no cheaper than human translation. After the report, public funding for the field in the United States of America was largely cut off.

Which was the first machine translation system that genuinely worked?

The Météo system, brought into service in Canada in 1976, is generally given that title. The system translated weather bulletins and nothing else; the reason for its success was precisely that narrow and repetitive domain.

Where does today’s machine translation stand in that history?

The field moved from rule-based systems to statistical models and from there to neural networks. Fluency has risen markedly in recent years; but because fluency and accuracy are not the same thing, human oversight remains essential for texts that carry responsibility.

Sources

Pietrzak, P. & Kornacki, M. (2021). Using CAT Tools in Freelance Translation. Routledge Focus, Chapter 3 (Machine translation).
Hutchins, W. J. (1986). Machine Translation: Past, Present, Future. Ellis Horwood, pp. 21–46.
Hutchins, W. J. & Lovtskii, E. (2000). Petr Petrovich Troyanskii (1894–1950): A Forgotten Pioneer of Mechanical Translation. Machine Translation, 15(3), pp. 187–221.
ALPAC (1966). Language and Machines: Computers in Translation and Linguistics. National Academy of Sciences, Publication 1416.