24/02/2025

Would you choose a freshly prepared tiramisu from a chef, a decent dessert from the café or a cheap option in plastic packaging? The choice depends on budget, quality requirements and the time available – and the same applies when translating technical documentation. Here, too, you have three options: human translation, machine translation (MT) with post-editing (MTPE) or pure MT. While the cost savings are tempting, they also bring risks.

Translations of technical documentation are always caught between quality and costs. But how large are the potential savings and what risks come with using machine translation? Looking more closely at the available processes helps answer this question and helps with making decisions.

Multilingualism is by no means a foreign concept in technical documentation: according to a tekom study from 2013, almost two thirds of the industrial companies surveyed were already translating their documentation into at least ten languages. There is also a high volume of text. On average, around eleven information products are created for each product. At first glance, both look like a significant cost factor. At oneword, we conducted a survey of technical writers and asked about the share of time and costs attributed to each of the various technical documentation sub-processes. It is important to note that the companies surveyed often do not record internal expenses and costs, for example for creating and reviewing the source text. External costs, on the other hand, can be precisely quantified, for example for processing graphics and translation.

Allocation of time and costs to technical documentation sub-processes (source: oneword GmbH)

Our diagram shows the distribution of time and costs for the four sub-processes surveyed. Translation takes a little less than a quarter (23.75 per cent) of the total time, while creating the text in the source language takes almost half of the total time (47.5 per cent). However, all of the companies surveyed counted parallel processes in different languages as just one period of time for translation. Therefore, the time required for each target language is not added up. This then applies to all sub-processes in which participants work in parallel and not sequentially. The distribution of costs is similar, but not identical: as translations are usually carried out in several languages, they account for 26.75 per cent of the total costs, which is a larger proportion than their time expenditure would suggest. Unlike time allocation, the costs per language are added together, making translation one of the biggest cost factors. But that also means that this is where there is the greatest potential for savings. However, savings should not come at the expense of quality.

While with tiramisu the difference between a freshly prepared dessert and an industrial, mass-produced version is evident in the taste, the difference in quality with translations often lies in the details, which only become apparent once the translation is being used. Technical translations require the utmost precision, as incorrect or unclear formulations can have serious consequences. DIN EN IEC/IEEE 82079-1 therefore recommends using specialised translators for creating instructions for use due to the high degree of technical knowledge required and the specific features of translating technical documentation. It follows that where high-quality levels are required, there is also a constant cost pressure. It is therefore worth looking at the potential savings for each of the three translation options mentioned above, because, as shown, this represents an average of 26.75 per cent of the total cost of technical documentation.

The cost of a translation decreases when the time required and the services provided are reduced. While human translation requires the greatest effort, MTPE noticeably reduces the effort required. Pure MT is the cheapest option, but it is also the riskiest. In addition to the qualifications and skills of professional translators mentioned above, using supporting technologies efficiently makes a big difference in human translation and also offers the greatest potential for savings. Thanks to translation memories (TMs), translations can be stored and reused. Especially in technical documentation, which is frequently updated and added to, this is an important factor in reducing costs, having short throughput times and ensuring consistency within a text. Companies can decide individually whether existing translations should be checked again or whether they should be locked and simply used for context. These measures have a direct impact on costs: while matches from the TM are charged at a discounted rate, there are no additional translation costs for locked segments.

Systematic terminology work and using a terminology database can also significantly reduce the work involved in technical translation. Although defined terminology does not directly lead to discounts, it does minimise research work, and reduce queries and the need for corrections. All three factors reduce the effort and lead times involved in the translation process.

When using MTPE, the savings potential depends on factors such as the quality of the MT system used, the language combination and the specific project requirements. Depending on the starting point, it is possible to make savings of 10 to 30 per cent. As this discount is only applied to the MT output, the more new text that there is to translate in a document, the greater the cost reduction compared to human translation.

Integrating MTPE into CAT tools also has an impact. By using a TM and MT, existing translations are reused and new content is pre-translated by the machine. The work steps for matches from the TM are then identical to human translation, which is why the calculation is the same. The savings potential with MTPE is therefore particularly high for new documentation in which many segments have to be translated for the first time. The situation is different for documents that have only been updated. Here, the savings are primarily achieved by using a translation memory, while MT may only lead to marginal cost reductions.

When calculating an MTPE project, it is also important to consider the specifications for post-editing and their feasibility using features of MT systems. For example, if there is a lot of company-specific terminology that the MT system does not use natively, this means considerable rework when post-editing. However, some MT systems offer the option of integrating these specifications directly into the machine translation by including a glossary. This significantly reduces the effort involved in post-editing.

Pure machine translation without professional checking, like the tiramisu from a plastic container, is by far the cheapest of the three options. The greatest potential for savings depends on the MT provider. Depending on the provider, the client is billed on a character, word or document basis or through monthly or annual subscriptions for professional and data-secure systems. Using pure machine translation without a subscription or licence agreement is not an option in professional contexts due to the lack of data security. Pure MT can also be integrated into CAT tools as a hybrid approach so that TM matches can also be used in parallel.

One stumbling block with machine translation is the lack of reproducibility, as translation systems are constantly evolving through regular training. The same source document can therefore deliver a different result today than tomorrow. Unlike the production of tiramisu, the machines do not work to a recipe (which would have a reproducible result). Instead, they work ‘freestyle’, which can lead to different results with each attempt. This variability also has a direct effect on the subsequent steps: if documents for MTPE or pure MT are always transferred in full to an MT system, full post-editing and a full quality check must be carried out after each translation. The review, final verification and layout stages may also have to be carried out again. When updating documents, this approach achieves hardly any savings compared to creating a new document, as the entire process always starts from the beginning.

Another problem with using machine translation is terminology. How well technical terms are implemented varies considerably within one MT system, and even more so between different systems. Tests conducted with a sentence from the machining technology field show, for example, that the technical term “Räumen” was translated completely differently in three MT systems.

Translation from the machining technology field using different MT systems (source: oneword GmbH)

The follow-up costs associated with the post-editing required to make the terminology consistent would also vary accordingly.

When preparing documents for translation, it is also important to know whether all content or only modified text passages need to be translated. At first glance, selecting only new content can save time and costs, but it brings risks: if only individual text segments are translated, they may not be consistent with the rest of the document. If a text is compiled manually, the potential for errors increases considerably. This could be because changes are overlooked or a unit of text is assigned to the wrong translation in the foreign-language version. Therefore, especially when using CAT tools and translation memories, it makes more sense to submit the entire document for translation. This allows the system to recognise matches from the translation memories and automatically use existing content. This maintains consistency and reduces the amount of post-processing work required.

Most companies in German-speaking countries have their first experience of MT with the language pair German-English. English is the most frequently required target language for companies in the DACH region. Many users feel confident enough in their own language skills to judge the quality of the machine translation. If the English translation from the machine is convincing, they quickly move to using MT for all other required target languages. English is the gold standard of any MT system, as it is often the language with the largest volume of training data available, regardless of the desired language combination. For all other languages, if they are offered by the MT system, it’s like comparing apples and oranges. There can be significant fluctuations in the quality of the target languages within a system. A text that is basically well suited for MT and only needs minor corrections in English may therefore require considerably more reworking in other target languages. Even with common language combinations, there is also a phenomenon that can affect the output quality of MT systems: the use of English as a relay language. We have analysed the risks that can go hand in hand with relay languages in a separate blog post.

When using a relay language within the MT system, documents are not translated directly into Spanish or Chinese, but first into English and only then into the target language. While in human translation, this approach is usually only used for low-resource language combinations (e.g. Czech to Vietnamese) where there are few to no translators available, the method is commonplace in MT even for high-resource languages such as Spanish. It runs inconspicuously in the background, but comes with considerable risks. The following illustration uses the German word “leicht”, an adjective meaning “light” (the opposite of heavy). It shows that the machine translation has taken place via English, producing the Spanish translation “luz”, a noun meaning “light” (the opposite of darkness):

Translating via English can also mean that the content and accuracy of the translation suffer. As a low-detail language, English expresses adjectives and job titles, for example, neutrally, meaning that they are not gendered. Information that can be read from words in a high-detail language, such as German, may therefore be lost when relayed through English.

Which of these translation options is chosen for technical documentation depends on quality requirements, time and budget. Machine translation paired with human expertise in the form of MTPE can significantly speed up the translation process and, above all, reduce costs. Translation memories and terminology databases play an important role in ensuring consistency, while post-editing by experienced translators guarantees quality. A feasibility analysis can be used to check the suitability of each translation project for MT in advance, as well as to optimise the use of resources. The choice of MT system has a major influence on the output, as does the scope of specifications and project instructions. But be careful: having to correct the MT output can also quickly swallow up the time saved. Pure MT may be sufficient in a few cases, for example using internal documents to obtain information only. But for publication, or where translation errors would result in risks, MTPE or human translation should always be used. Unlimited savings are also not possible in translation if you still expect a high-quality service and high-quality result. Or, if you reach for the unchilled tiramisu, you should not be surprised when you get a stomach ache later.

Would you like to find out more about how machine translation can save you money? Then contact us for a personalised consultation.