07/04/2025

The rapid change brought about by the integration of artificial intelligence is affecting almost all industries and is also revolutionising the translation industry. The industry is used to drastic changes, as the breakthrough of neural machine translation (MT) caused disruption just a few years ago. Since then, MTPE – the combination of MT and post-editing (PE) – has established itself as a method for bringing machine translations to a professional level. The emergence of Generative Artificial Intelligence (GenAI) in the form of Large Language Models (LLMs) such as GPT, Mistral and Llama again begs the question of whether and how translation processes will continue to evolve and change.

Change usually also means uncertainty, and this is precisely what the translation industry is confronted with: according to CSA Research, a technological breakthrough in AI could positively impact growth forecasts to the same extent that legal hurdles, such as lawsuits related to AI-generated translations, could have a detrimental impact on them. The only thing that is certain is that the field of machine translation will continue to develop – with a host of new opportunities and challenges.

As industry insiders, we will take a look at the latest developments in MTPE and LLMs and show you how we at oneword use these technologies in a targeted and secure manner for our customers.

Machine translation + post-editing (MTPE) has established itself as a reliable standard in the translation industry. Machine-generated translations are checked and post-edited by human professional translators to ensure the quality, accuracy and stylistic consistency of the translation. This method offers considerable advantages – especially for large volumes and companies that require fast and cost-effective translations.

Continual improvements in MT technology have also increased the quality of machine translations over the years. Nevertheless, these systems do not translate without errors and are by no means suitable for every document and every language combination. This is why it is particularly important that a feasibility analysis is conducted prior to machine translation. This involves checking whether the text type, subject area, language combinations and content are suitable for MT pre-translation and what savings can be expected from using it. This means that all those involved in the process are on the safe side, as only those projects where time and money can really be saved are put through the MTPE workflow. Combining this with other technologies, such as translation memories, content and language management systems and optimised source texts, enables smart, cost-effective and secure translation processes and clear competitive advantages.

An in-house survey conducted by oneword shows that the majority of our top customers are already using machine translation. Over 75% of companies from a wide range of industries use oneMTPE, our customised solution for using MT in the translation process. In addition to the feasibility analysis, MTPE at oneword also stands out thanks to its detailed feedback process, which enables continual monitoring and targeted improvements. Other steps, such as optimising source texts, using clean terminology databases and glossaries, and closely integrating automated workflows, help to make the process even more efficient. After all, sustainable savings can only be achieved without compromising on quality by strategically combining technology and human expertise.

Large Language Models (LLMs) herald a new era of artificial intelligence that can go far beyond conventional machine translation. In contrast to MT systems, LLMs were not developed specifically for translation, but for general language comprehension. Therefore, they are not limited to translating within the context of the sentence but also grasp broader contexts. Thanks to their broader understanding of context, they are better able to recognise and implement linguistic, stylistic and cultural nuances.

Despite linguistically impressive results, LLMs also have clear weaknesses, including hallucinations, inconsistent results and a lack of reproducibility when generating text.

MT systems have also been delivering often impressive results for years. It is therefore worth comparing the advantages and disadvantages of MT and LLMs, which also suit different applications.

The first thing to look at is the sentence structure and the style of the output: while MT systems adhere strictly to the sentence structure and only understand and apply context within a sentence, LLMs often provide more idiomatic formulations, as they can grasp broader textual contexts and also create a consistent style across the entire text. This aspect is somewhat limited when used within a CAT tool, as the tool breaks down the text into individual sentences. However, even within CAT tools, a text is not transferred to the Large Language Model sentence by sentence, but in larger blocks.

When it comes to specialised terminology, neither type of system provides consistent results. In practice, it depends both on the underlying training material and on word probabilities, which change from sentence to sentence, depending on the other words used. In both MT systems and LLMs, however, it is possible to integrate terminology as glossaries, databases or prompts to ensure that the specified terminology is used correctly and consistently.

LLMs have a clear advantage when it comes to freer translations or rewriting a target-language text according to specific instructions. The generated output can be revised again at any time and adapted to new style guidelines or a different target audience, for example.
MT systems score highly in terms of reliability: omissions or spontaneous additions, which occur with LLMs in the form of hallucinations, are now rare in traditional machine translation. As well as being more complete, MT output can also be more reliably reproduced than with LLMs, which usually deliver a different result with every prompt.

From the sources of error mentioned above and experience with machine-generated texts, it is clearly also essential that translations from large language models are post-edited. As well as the traditional review and error correction, revision in the form of prompting, for example to stylistically adjust the output, is also very important. Utilising results from error and feedback analyses also plays a decisive role. This is because, unlike MT systems, LLMs can be guided by instructions in the form of prompts. Typical sources of error can therefore be addressed and, in the best case scenario, eliminated.

As already mentioned, unlike MT systems, LLMs were not explicitly trained for translation. Through multilingual material, however, they have learned this function more or less ‘on the side’. Compared to traditional machine translation, however, they cover a much broader spectrum of tasks for the language industry, e.g. automated text processing, intelligent content creation and AI-driven translation and language assistance.

For example, LLMs are particularly suitable for:

• Content creation (blog articles, product descriptions, YouTube scripts, e-mail campaigns)
• Creative translations (transcreation) for marketing and social media
• Context-related, idiomatic and stylistic adaptations
• Reformulations and summaries of content

MT remains ideal for:

• Translations into many different languages
• Technical, legal or medical translations with high precision
• Structured texts with clear terminology
• Mass translations with a short time frame, e.g. product descriptions in e-commerce

It is becoming apparent that MT systems and LLMs will complement rather than replace each other. The decision on which system to use must be made on a case-by-case basis, particularly in view of the costs of a translation, which can be higher with LLMs due to token-based billing. While LLMs tend to excel in content creation, customisation and creative translations, MT remains the first choice for structured, precise and, above all, reproducible specialised translations. The important thing for both technologies is to integrate them into existing processes and, for example, to use them in a hybrid approach together with translation memories and terminology databases. Post-editing also remains an integral part of the process, as every system has different sources of error. Therefore, human expertise remains essential, also for assessing the general feasibility and utilising feedback.

AI in the form of LLMs has definitively arrived in the translation industry. Unlike many industries, however, our sector is experienced in dealing with disruptive changes. As a company, oneword actively participates in expert groups and tests to evaluate the use of LLMs for a wide range of scenarios and to establish secure AI-supported translation services. Our many years of experience in the MTPE sector, particularly with regard to general feasibility and evaluating machine output, play an important role here. In a rapidly changing world, we offer you AI expertise that you can trust for your language processes.