The History of CAT Tools: What Happened After the ALPAC Report?

The history of translation technology is usually told as a success story. Yet the history of CAT tools begins not with a success but with a report of failure. The ALPAC report, published in 1966, gave a cool-headed assessment of where machine translation research stood at the time, and the verdict was negative.

Contents

The report closed one field but opened another door in the very same pages: the idea that the computer should assist the translator instead of replacing them. The translation memory, the termbase and the match rates we use today all came through that door.

This article explains how a decision taken sixty years ago shapes the translation quotation you receive today. You do not need to be a history enthusiast; at the end of the story sits the “repetition discount” line on your invoice.

What did the ALPAC report say, and what did it stop?

The 1950s were years of optimism for machine translation. Warren Weaver’s memorandum of 1949 had put the idea on the agenda, and the Georgetown experiment of 7 January 1954 staged a demonstration with a limited dictionary of 250 words (Pietrzak & Kornacki 2021, Chapter 3). The press presented that demonstration as the herald of a problem that would be solved within a few years.

By the end of the decade the promise had not been met. In 1966 the ALPAC report measured this picture and stated that fully automatic machine translation was, with the hardware and the linguistic models of the day, more expensive, slower and less reliable than human translation. The consequence was harsh: in the United States, the public funds allocated to the field were largely cut.

The part of the report that is often skipped over is this: it did not reject machine translation outright. What it rejected was the goal of taking the human out of the loop. The report pointed to tools that support the translator directly, such as dictionary access and text processing, as the more realistic investment.

The table below summarises the phases the field has passed through since then and the practice each phase has left behind.

Period	Dominant idea	What remains today
1949–1966	Fully automatic translation will be possible before long	The founding questions of the field and the first dictionary structures
1966–1980	Let the computer assist the translator	The logic of the termbase, narrow-domain applications
Late 1980s–2000s	Translation once done can be stored and reused	Translation memory, match rates, repetition discounts
After 2010	Let statistics and neural networks produce suggestions	Post-editing of machine translation, human oversight

The real idea out of the report: not the machine, but the machine-assisted human

In the period after ALPAC the research agenda changed. The question was no longer “can a computer translate?” but “which part of the translator’s work can a computer take on?”. It was a modest question, but a productive one.

What really ate up a translator’s day was not composing sentences: it was searching dictionaries, using the same term consistently throughout a document, and rewriting an almost identical paragraph that had been translated the month before. All of these are repetitive, rule-based tasks, the kind of work computers are good at.

This reframing changed the aim of the technology as well. The goal was not a system that produces flawless translation; it was a system that finds the flawless translation the translator produced earlier and puts it in front of them. That is the core idea of the translation memory, and it is still the same today.

The commercial result of this approach became measurable too. While the industry benchmark stood at 2,500 words a day, translators using tools were reported to average 2,684 words a day (DePalma 2012). The difference may not look striking on its own; but that average includes texts with a low repetition rate as well. In document families where repetition is dense, the gain is far more marked.

The gap years: moving terminology onto the computer

The 1970s were a quiet period for machine translation, but not for translation technology. In those years large institutions began handing the terminology problem over to the computer: databases were built in which thousands of technical terms, abbreviations and institution names were held in one place.

The choice was not random. Terminology is the layer at which an error in translation costs the most. The style of a sentence can be argued over; the name of a screw appearing in three different forms within one document, on the other hand, means the wrong part on the production line.

In the same period the first machine translation application that actually worked also appeared in a narrow field: the Météo system, which went into service in 1976, translated weather bulletins. The reason for its success was not the power of the technology but the narrowness of the field: a limited vocabulary and fixed sentence patterns. The lesson still holds today: the more predictable the language, the better automation works.

How did the translation memory idea mature?

The translation memory is the product of a simple observation: in technical texts the same sentence comes back with surprising frequency. A new version of a user manual carries most of the previous one over unchanged. The clause headings of contracts barely change at all.

The observation is supported by figures as well. In industry research, the proportion of the text translators handle that is entirely new was measured at 59.38% for in-house translators and 59.11% for freelance translators (DePalma 2012, CSA). In other words, roughly 40% of an average job carries the potential to overlap with something translated before.

Two technical thresholds had to be crossed before the idea became workable: being able to break text into units at sentence level (segmentation) and being able to measure how similar two sentences are (fuzzy matching). With the spread of personal computers both thresholds were crossed at the end of the 1980s, and commercial translation memory programs came onto the market.

Why does the history of CAT tools still explain the present?

This history has three consequences that bear on the present. The first is pricing. The per-word rate is no longer calculated on the raw word count but on a word count weighted against the memory. On long-running work the difference is not to be brushed aside.

The second is consistency. Corporate language staying the same across a stack of documents is the achievement of a database, not of human memory. Brand names, product codes and legal definitions are held steady this way.

The third is a warning. Convenience feeds carelessness as well. Bowker (2005) points out that translators can develop a “blind faith” in the memory: when the system shows a high-percentage match, the translator may accept that segment without questioning whether it is genuinely correct in the new context. The benefit of a technology is real only in so far as the person using it keeps their scepticism.

Read together, these three consequences produce the following picture: CAT tools bring costs down, raise consistency and, in return, add a new checking burden. The buyer’s job is to make clear in the contract where that burden sits. When it is not settled at the outset who will hold the memory, how often it will be cleaned and whether high-match segments fall within the scope of revision, the gain from the discount can later turn into an argument about quality.

Is today’s artificial intelligence debate on the same pendulum?

There is a structural resemblance between the optimism of the 1950s and today’s debate about large language models. In both periods the demonstrations are impressive, the expectations immoderate, and the measurement comes along behind.

The differences are real too. Unlike the systems of the ALPAC era, today’s systems mostly produce intelligible output on everyday texts, and the proportion of translators who have tried machine translation is already high: 44.02% among in-house translators and 55% among freelance translators (DePalma 2012). The technology is no longer a curiosity but part of the workflow.

The real lesson ALPAC left behind is not a note in a history book but a method: measure the result, not the claim. Whether a tool works is established not through its marketing but by measuring correction time and error rate on a particular document type and in a particular language pair.

The Türkiye context

Türkiye’s translation market joined this history late but quickly. CAT tools entered the country not through the chain of institutional research between 1966 and 1990, but directly as commercial programs in the 2000s. The result was that “how” the technology is used was learned while “why” it exists mostly went undiscussed.

The structure of Turkish makes part of this inheritance harder. In an agglutinative language a single root produces dozens of inflected forms. For a translation memory working at sentence level, this means that two sentences can be identical in meaning and yet look different at character level; the match rate may come out lower than expected.

Against that, the document types translated most heavily in Türkiye – notarised papers, visa files, tender specifications, ISO documents – are extremely formulaic texts. The repetition rate, that is, is high, and investment in a memory pays for itself quickly here. The real bottleneck in Türkiye is not technology but the fact that organisations have never put their own terminology in writing.

What does this mean for you?

For the buyer of translation, this history comes down to a single sentence: no sentence that has already been translated should be translated again at the full rate. You have the right to ask for this, and asking requires no technical knowledge.

1. When you ask for a quotation, ask for the weighted word count alongside the raw word count.
2. On your second job with the same supplier, ask in writing what the repetition rate came out at.
3. Keep your past translations as files; they are an asset that can be turned into a memory.
4. Write your organisation’s 30–50 critical terms into a table and attach it to every job.
5. State in the contract that the memory and the term list belong to you.

Four of these five items are completed on your side and cost nothing. In organisations with a continuous need for translation, the memory built up by the end of the first year pulls the budgets of the following years visibly down. If you would like to see how it is calculated, you can look at our pricing page and request a quotation for your own document.

Frequently asked questions

What exactly did the ALPAC report stop?

The 1966 report found that fully automatic machine translation was, under the conditions of the day, more expensive and slower than human translation. The public funds allocated to machine translation research in the United States were then largely cut.

What is the difference between a CAT tool and machine translation?

Machine translation translates the text itself. A CAT tool leaves the translating to the human; it stores the translations that person has made before, suggests them for repeated passages and checks terminological consistency. The two belong to different families of technology.

Why was the translation memory idea born after ALPAC?

The report did not reject automation as a whole; it pointed to assistive systems with the human at the centre. The idea that the computer should support the translator instead of replacing them matured in this period and opened the road to the translation memory.

Why does this history concern me as a buyer of translation?

Because the repetition discount in the quotations you receive today is a direct product of this inheritance. If your text contains passages that overlap with earlier translations, they are not charged at the full rate; the basis of the discount is the logic of assistive systems developed after ALPAC.

References

• Pietrzak, P. & Kornacki, M. (2021). Using CAT Tools in Freelance Translation: Insights from a Case Study. Routledge Focus, Chapter 3.
• DePalma, D. A. (2012). Translation Productivity Benchmarks. Common Sense Advisory.
• Bowker, L. (2005). “Productivity vs Quality? A Pilot Study on the Impact of Translation Memory Systems”. Localisation Focus, 4(1), pp. 13–20.