"Can’t we just do that with AI?" This is a question oneword is asked regularly, and it’s a fair one. AI is on everyone’s lips at the moment: providers are making big promises, new AI models are eclipsing their predecessors almost every week, and the pressure on companies to achieve cost savings through the use of AI is mounting. Translation seems to be an obvious candidate for artificial intelligence, given that AI translation at the touch of a button has now become an integral part of our everyday lives, whether it’s the automatic translation of websites and videos in a browser, or for quick communication whilst on holiday.
However, AI is not a new phenomenon in the translation industry. Machine translation systems have been operating for years on the basis of so-called neural networks, a form of AI technology. And the process that combines machine pre-translation with professional post-editing – MTPE (short for Machine Translation + Post-Editing) – has been an established standard at oneword for around ten years and has proven its worth in thousands of client projects.
What has changed is the functionality of state-of-the-art AI technologies. Large Language Models (LLMs), such as the systems behind ChatGPT and similar applications, have fundamentally changed the way AI is perceived – whilst at the same time causing considerable confusion over terminology. All of a sudden, everything is being labelled "AI": traditional machine translation, generative language models, automated workflows. In this article, we’ll explain what’s behind this and what it means for your translation processes.
Two aspects that often get mixed up in projects: technology and process
When people talk about "AI translation" these days, there is often more to it than just one specific solution. To help you understand quotes more clearly, it’s useful to distinguish between two levels:
- Technology: Which systems are used to generate a machine pre-translation? (e.g. neural machine translation (NMT) or LLMs)
- Process: How can this pre-translation be turned into a product that you can use with confidence, both internally and externally? (e.g. MTPE)
The term "AI" on its own does not indicate how reliable the result is, how consistently terminology is applied, or what checking steps are planned following the AI translation. A clearly defined MTPE process ensures transparency, high-quality results and predictability in terms of effort, timing and the standard of output.
The technologies: A comparison of neural machine translation and LLMs
Both neural machine translation (NMT) systems and LLMs are neural networks that have been trained on the basis of (multi)lingual data to perform specific tasks. Whilst NMT models are purely translation models, LLMs can process any linguistic instructions and generate a machine translation output. Both types of model can be used for translation processes, depending on the suitability for the project and individual requirements.
Neural machine translation (NMT)
- Definition: A trained language model that has been specifically developed for translation. Depending on the system, it may have been trained to be generic, domain-specific or organisation-specific.
- Strengths: NMT is particularly effective when it comes to delivering consistent, reproducible results. Under the same conditions, it consistently produces very similar results – an advantage when dealing with large volumes and recurring texts. Furthermore, NMT can be tailored to specialist departments and terminology, which can be crucial in technical and industrial contexts.
- Limitations: NMT is less flexible when texts rely heavily on context, or when tone and communicative intent are paramount. Although domain-specific and company-specific training is possible, it usually involves a significant amount of effort – time, data resources and costs are all factors here.
Generative AI (LLMs) in the context of translation
- Definition: Models that have been trained for general language comprehension and are therefore also capable of translation, often producing results that are very close to traditional machine translation outputs.
- Strengths: LLMs really come into their own wherever context and communicative intent play a major role: the mode of addressing the audience, tone of voice, creativity. A key difference from traditional MT is that LLMs can often be significantly influenced simply by providing a few targeted guidelines (custom prompts, examples, reference material), without the need for extensive system training. This allows organisation-specific requirements to be incorporated into the AI output at low cost.
- Limitations: The output of LLMs is less reproducible than that of trained NMT systems: identical requests can produce different results. The great strength of LLMs – their general understanding of language – can also be their undoing: rather than “simply” providing translation output, they may, in some cases, add creative elements, rephrase the text, or produce output in the wrong target language. To ensure reliable practical application, LLMs therefore require the technology to be seamlessly integrated into the translation process, as well as specific checks to identify LLM-specific sources of error.
MTPE: the difference lies in the process
The technology determines how a pre-translation is produced. However, the difference in a business context is particularly evident in how this is transformed into a translation that you can use with confidence – terminologically consistent, technically accurate and in line with your specifications.
MTPE (Machine Translation + Post-Editing) is the term used to describe the workflow in which machine translation output is transformed into a reliable target text through expert post-editing and quality checks. MTPE is technology-neutral, which means that – depending on the text type, language direction and requirements – either a neural machine translation system or an LLM can be integrated as the machine translation system for pre-translation.
DIN ISO 18587 serves as the reference framework. It describes post-editing as a defined process with roles, workflows and requirements. oneword has been certified to DIN ISO 18587 since 2019 and therefore works in accordance with standardised, traceable MTPE processes.
Generative AI (LLMs) in the context of translation
- Feasibility analysis
We will assess the text type, language direction, risk profile and potential savings for you in advance, free of charge. This will provide you with a reliable recommendation for the right setup. Documents that are not suitable for MTPE (i.e. where the amount of post-editing required would be roughly equivalent to the effort involved in a translation from scratch) can instead be translated by a human translator. - Terminology and translation memories
Existing language resources are integrated directly into the process to improve MT quality and reduce costs and post-editing effort. Specified terminology can be integrated directly into NMT or LLM output in the form of a glossary. The integration of existing translation memories, in turn, ensures that existing translations are reused and lays the foundation for consistency across individual documents.
- Machine translation
The appropriate system (NMT or LLM) is selected depending on the content, languages and any additional requirements. - Post-editing & quality assurance
Professional translators check, correct and adapt the machine translation output to your specifications. On request, a revision of the translation can also be carried out by a second professional translator; multi-stage quality checks carried out by the professional translators and project managers at oneword are an integral part of the process. - Feedback & optimisation
Professional translators evaluate the machine translation across various categories. Change rates and sources of error are systematically evaluated, any necessary measures are identified, and the process is continuously improved. During regular MTPE check-ins, we’ll keep you informed about the savings you’re making through MTPE, the quality of your machine translation output and areas for improvement.
Data security is not an optional extra here, but a prerequisite. We use GDPR-compliant workflows and licensed machine translation systems and LLMs that do not store content permanently. In addition, there are clear role and authorisation frameworks to ensure that the solution can be deployed in a controlled manner within your process and tool environment.
Deviations from the standard specification are only made upon explicit request and are clearly indicated in the quote.
What really matters when it comes to AI translation
These days, the term "AI translation" can mean a great many different things. For your project, it’s not so much the buzzword that matters, but rather which technology is used and how the process is structured.
Neural machine translation systems and LLMs are different technological approaches, each with their own strengths – ranging from stable reproducibility to context- and goal-oriented adaptation. In professional translation processes, however, what ultimately matters most is that terminology, quality checks and responsibilities are clearly defined. This is precisely where MTPE comes in: as a defined workflow that transforms a machine-generated pre-translation into a technically sound product that your organisation can have confidence in.
If you’re wondering whether your documents are suitable for the MTPE process, and which technologies (NMT or LLM) and language data (glossary, translation memory) are appropriate for your content, then contact us today.
Practical experience shows that this approach doesn’t just work in theory: with oneMTPE, HELLA achieves average cost savings of 30%–40% and up to 50% faster turnaround times compared with human-only translation, whilst ensuring the quality of the machine translation output is professionally guaranteed.
