The Limits of the Research: How Do the Findings Adapt to the Translation Industry in Türkiye?

The research behind this series of articles is a case study measuring how computer-assisted translation tools have changed the work of the freelance translator. All the data comes from a single source, and that source has plain limits.

Contents

This article is the honesty framework of the series. Here we explain why every percentage we share in the other articles is given with the label “(Poland, 2019)”, and we discuss how much of it can be treated as valid under the conditions of the translation industry in Türkiye.

Let us give the short answer at the outset: for Türkiye the findings are not a benchmark but a list of questions.

What was the research? A short summary

The study rests on a survey carried out in Poland over the 2019 period. There were 141 participants: 90 of them said they used a CAT tool, 51 that they did not. The comparisons were drawn between these two groups.

The table below brings together all the verified findings we use across the series. The last column states whether comparable data exists for Türkiye.

Finding	Value	Equivalent for Türkiye
Sample	141 translators (90 users, 51 non-users)	None
Tool distribution	Trados 43.3% · memoQ 29.1% · Wordfast 10.6%	None
Licence ownership	91% own their licence	None
Compatibility problems	90% run into them, 70% of those frequently	None
Living from translation alone	Users 75% · non-users 41%	None
Reasons for not using one	No need 50% · price ≈40% · no demand ≈40% · too complex 25%	None
Higher rate for translation into a foreign language	80% (non-users)	None

That the last column is empty throughout is not a confession of shortcoming but a deliberate choice. Because there is no comparable measurement for Türkiye, we do not produce estimated figures.

Reading the table as a whole matters too. Every row belongs to the same survey, the same 141 people and the same period. These are not independent pieces of evidence confirming one another, then, but different faces of a single observation.

Limit 1: the sample comes from Poland, not Türkiye

The most basic limit is geography. Every participant was working in the Polish market, and the structure of that market differs from Türkiye at several important points.

Poland is a member of the European Union, and the translation demand of the EU institutions feeds the regional market directly. That demand is high in volume, standard in format and carries a requirement to use particular tools. The prevalence of Trados has its roots in this institutional demand as well: the European Commission tenders of 1996-97 turned the tool into a de facto standard (Garcia 2015).

There is no such central source of demand in Türkiye. Document translation tied to notarisation, export paperwork and web content carry greater weight instead. Because the mix of work is different, the pattern of tool use is to be expected to differ too.

The language dimension should be added to this. In the Polish sample the dominant language pairs lie between Polish and the major European languages. Turkish belongs to a different language family, is agglutinative and is more limited in digital resources. Match rates and term checking behave differently in Turkish.

The economic scale differs as well. The size of the global language industry in 2019 was reported at 49.6 billion dollars (CSA); for the European Union language industry a figure of 8.4 billion euros was given for 2008 (Rinsche & Portera-Zanotti 2009). These totals are not distributed evenly across countries, and there is no reliable breakdown of the share belonging to Türkiye.

Limit 2: 141 people, a small sample

The second limit is scale. For a descriptive case study 141 people is reasonable; for generalising at national scale it is small.

It helps to see what the number means in practice. In the group of 51 non-users, a proportion of 25% corresponds to roughly thirteen people. At that scale, two or three people answering differently move the proportion by several points and the table looks different.

The groups being unequal produces an asymmetry too. In a comparison of 90 people against 51, the figures for the smaller group are the more fragile. Taking percentage differences seriously down to the decimal place would be misleading.

For that reason we present the figures across the series as a ranking by weight rather than as estimates of proportion. Saying “this reason was voiced more often than the others” fits the data; saying “exactly half of all translators” does not.

The same principle shaped the form of our tables. We keep the decimal places because they are given in the source, but we point out that those places do not amount to precision. In a group of 90 there is no meaningful difference between 43.3% and 43.9%.

Limit 3: voluntary participation and self-reporting

The third limit is methodological. Participation was voluntary; no randomly selected list was approached. The translators who came across the survey answered it of their own accord.

A systematic bias is to be expected in samples of this kind. People with an interest in the subject answer a survey about technology more readily. Those who do not belong to professional networks, who work alone or who translate as a secondary occupation are under-represented.

A second methodological point is that the data rests on self-reporting. Income, tool use and the frequency of problems were recorded in the participant’s own words; none of it was independently verified. In self-reported measurements it is normal for the position to be presented somewhat favourably.

These two points do not invalidate the data. But they do make it obligatory to turn a sentence such as “91% of translators in Poland” into “91% of the translators in this sample”.

The shift looks small, but its consequences are large. The first sentence is a claim about a country; the second is an observation about a group. Across the series we try to use the second form.

Limit 4: 2019, a photograph taken before artificial intelligence

The fourth and perhaps most decisive limit is the date. The survey belongs to the 2019 period. Neural machine translation was widespread by then, but large language models had not yet entered the freelance translator’s daily workflow.

In the time since, licence models have shifted towards subscription, cloud-based tools have spread and post-editing has become the default workflow on many projects. Part of the findings on attitude and expectation has therefore in all likelihood dated.

Some findings, by contrast, look more durable. Who carries the cost of the licence, and the difficulty file formats have in talking to one another, are structural problems that are not easily solved even as the generation of models changes.

We do have data from other dates for comparison: the proportion of freelance translators who had tried machine translation was reported at 55% (DePalma 2012). That figure is in all likelihood much higher today; but because we have no current measurement produced by the same method, we give no new number.

What use are these findings, then?

Once the limits have been listed, a legitimate question follows: with so many constraints, why do we use this data at all?

Because the alternative would not be using no data; it would be using claims of unknown origin. Most of the sentences circulating in the industry along the lines of “CAT tools raise productivity by such and such” have no sample behind them whatsoever. A measurement whose limits are known is always worth more than a claim with no source.

The second benefit is that it generates the right questions. The licence cost sitting with the translator in Poland prompts the same question to be asked in Türkiye. Compatibility problems being chronic is a reminder to ask your supplier about file formats.

The third benefit is time. The 2019 photograph offers a starting point to anyone wanting to measure what has changed in the period since artificial intelligence. Before a comparison can be made at all, a baseline is needed.

One warning about use, finally: these findings are not marketing material. A supplier using survey results of this kind to prove its own superiority would be loading the data with a weight it cannot carry.

The Türkiye context

There is no publicly available, regular study measuring the translation industry in Türkiye by this method. Partial studies produced by professional bodies and universities exist, but there is no comparable time series.

One reason for this is that the definition of the profession is unclear. Translation is not a regulated profession at European Union level either (Pym et al. 2012); nor is a compulsory professional certificate required in order to provide translation services in Türkiye. Sworn translation, as a separate category, runs within the framework of the Notaries Act. When it is not clear who is to count as a translator, building a sampling frame becomes difficult too.

The features of the Turkish language come on top of that. Agglutination makes word-based volume measurements impossible to compare directly with other languages; free word order affects fuzzy match rates. Even if the Polish data were carried over to Türkiye, in other words, the units of measurement themselves would not mean the same thing.

The gap in the data has a practical consequence: organisations buying translation services in Türkiye have to look at their own past projects rather than at industry averages. Your own archive offers a more reliable basis for comparison than industry data that is not publicly available.

What does this mean for you?

For the party buying translation, the real gain here is a method: you can test every figure put in front of you with the same four questions.

Where was this data collected? In your own market, or in another country?
How many people, or how many projects, does it rest on?
How was participation obtained; does the figure rest on self-reporting or on measurement?
What date does it belong to, and what has changed since?

A statistic that cannot pass these four questions is not enough to take a decision on. In such cases the most reliable course is to produce your own data: a ten-page pilot translation gives a more binding result than a survey of a hundred people.

We take the same approach ourselves: when we quote, we work from an analysis of your file rather than from general averages. If you would like an assessment measured on your own document you can use the quotation form, and for questions about the process you can get in touch with us. The other articles in the series are available in the knowledge centre.

Frequently asked questions

Can the findings of this research be applied to Türkiye?

Not directly. The sample is in Poland, consists of 141 people, rests on voluntary participation and belongs to the 2019 period. For Türkiye the findings are a source of questions, not a benchmark.

Is there comparative data for the translation industry in Türkiye?

There is no publicly available, comparable measurement for Türkiye produced by the same method. That is why we give no percentages for Türkiye in our articles.

Why is a sample of 141 people considered insufficient?

It is small for national generalisation. Once it is divided into subgroups the cells fall to ten or twenty people, and a few answers changing move the proportion by several points. For descriptive purposes it is sufficient.

Is 2019 data still meaningful today?

As a reference point, yes. Because it comes from before large language models spread through the industry, it offers a basis for comparison to anyone wanting to measure what has changed since.

References

Pietrzak, P. & Kornacki, M. (2021). Using CAT Tools in Freelance Translation: Insights from a Case Study. Routledge Focus, Chapter 6.
Pym, A., Grin, F., Sfreddo, C. & Chan, A. L. J. (2012). The Status of the Translation Profession in the European Union. European Commission.
Garcia, I. (2015). Computer-aided translation: systems. In: The Routledge Encyclopedia of Translation Technology. Routledge.
DePalma, D. (2012). Translation Productivity Benchmarks. Common Sense Advisory.