What Do Large Language Models Do in Localisation? Eight Use Cases

Neural machine translation has been part of the localisation workflow for a few years now. Generative language models are a more recent layer. Large language models enter the localisation process not only by producing translations, but by improving text, writing code and supplying context. That carries the scope of the tool beyond the translation box.

Contents

The field is moving quickly, so a cautious tone is called for. The headings below draw on the use cases set out in the 2024 edition of the book and in the industry reports it cites. These are not firm predictions but observed practices and reasonable expectations.

The aim of this article is to help the buyer of translation see where these tools fit. We look separately at the tasks where they genuinely make life easier and the tasks where they do not remove the need for human oversight.

Where Do Large Language Models Enter the Localisation Workflow?

The entry point is usually inside the tools already in use. Cloud-based localisation platforms can add language models to their own interfaces as apps. The translator can then take a model suggestion beside the segment being worked on, without switching to a separate tab.

The second entry point is content creation. The model comes in while the source text is still being written, producing suggestions on readability, clarity and cultural fit. This is an approach that tries to solve at source the problems localisation has classically dealt with after the fact.

The third entry point is the engineering layer. In web localisation, tasks such as cleaning up, explaining or generating mark-up can be handed over to models. Similar uses are reported for the scripting languages that drive dynamic sites.

What these three entry points have in common is this: none of them on its own produces the text that goes live. Each yields a draft, a suggestion or an intermediate product. The final decision sits with the human, and for now it is expected to stay there.

The Difference Between a Neural Engine and a Language Model

The two technologies come from the same architectural family, but their purposes diverge. Neural engines are trained for translation alone. Once trained, their behaviour is relatively predictable; for the same segment you will mostly get the same output. That makes quality control easier.

Generative language models, by contrast, are general purpose. They produce text, summarise, change the style and write code. They can give different answers to the same prompt. Their fluency is high, but they are reported to fall behind neural engines in accuracy. This duality decides which of the two is used for which job.

Criterion	Neural machine translation	Generative language model
Purpose	Translation only	General-purpose text generation
Predictability	High, repeatable	Low, the answer can vary
Fluency	High	Very high
Accuracy	Relatively high	Variable, needs verification
Additional capability	Limited	Summarising, change of style, code generation

The practical upshot is this: in high-volume, repetitive translation work, the predictability of the neural engine is an advantage. In work that calls for creative adaptation, summarising or a change of style, the flexibility of the language model may be more useful. The two are not rivals, they are different tools.

The difference in predictability also has a management consequence. On a project running on a neural engine, quality control can be carried out by sampling, because the behaviour of the system is stable. When you work with a language model, the same prompt can give different results, so the scope of the checks has to be kept wider. That in turn affects the revision budget.

The Eight Use Cases the Industry Report Lists

The Nimdzi report (2023) cited by the book gathers the applications of language models in localisation under eight headings. The list is a snapshot of the field on that date, and it is expected to widen as things develop. Even so, most of the workflows being built today stay within these eight headings.

1. Translation generation: platforms are adding language models to their own engine pools.
2. Improving source content: suggestions on readability, clarity, accessibility and cultural fit.
3. Engineering tasks: automating part of the scripting, coding and technical support work.
4. Multilingual content creation: producing blog, social media and documentation texts directly.
5. Proofreading: bringing the text into line with the style guide and the templates.
6. Linguistic context: explaining what surrounds a segment, resolving ambiguity and suggesting alternative wordings.
7. Next-generation product integration: combining generated content with existing language assets through interfaces.
8. Audiovisual localisation: consistency and terminology checks in subtitling and voice-over workflows.

Look at the list closely and only one of the eight items turns out to be translation itself. The rest are preparation, checking, engineering and content creation. In other words, the real effect of language models on localisation shows up not at the translation step but in the work around it.

Improving Source Content and New Professional Profiles

The quality of the source text is one of the strongest determinants of localisation cost. An unclear sentence turns into a separate problem in every one of the target languages. That is why using language models to improve source content matters: it catches the problem before it multiplies.

The reports add a caution here as well: the suggestions must be reviewed by a specialist before they go live. The process requires a checking layer much like post-editing of machine translation. A model suggestion is a starting point, not an approval.

Out of this need comes the forecast of a new professional profile. The report calls it the post-creator: the person who verifies the creative content the generative models turn out and adapts it to the brand and the locale. Whether the profile takes hold remains to be seen, but the need appears to be real.

On the corporate side, this profile is most often an existing role widening its remit. A content editor, a language specialist or a marketing manager takes on the job of verifying model output. What matters is that ownership of that job is set down in writing; a verification step left to nobody in particular tends in practice to be skipped.

Engineering and Context Tasks

On localisation projects a considerable share of the time goes not into translation but into preparation: parsing files, preserving tags, repairing broken mark-up. Language models bring a marked easing to part of this work.

Supplying context, on the other hand, touches translation quality directly. With short interface strings the biggest problem is the absence of context. A model that explains what surrounds a segment, or suggests where it might be used, can make it easier for the translator to decide correctly. We take up the context problem in a separate article in our knowledge centre.

The risk these uses share is that the explanation produced has to be verified too. A model does not really know where a string appears; it produces a guess based on probability. Using a guess as though it were a certainty can enlarge the very problem you hoped to solve.

If Search Habits Change, What Happens to Visibility?

The book points out that language models may affect localisation indirectly as well. If users begin reaching information through language models rather than search engines, the way web content is put together may change. That in turn affects web localisation and multilingual visibility work.

It is too early to make a firm prediction today. Reasonable preparation is still possible, though: building the structure of the content with clear headings, answering questions directly and producing text that stands on its own in each language all pay off under either scenario.

Another form of preparation is keeping language assets in order. Term lists, style guides and an approved translation memory hold their value whatever technology arrives. Without these assets, no model can reproduce the voice of your organisation consistently.

In short, the technology may change but the substance of the preparation stays largely the same: clean source text, defined terms and a clear stylistic framework. These three are the ground that decides quality, in the classical translation process and in model-assisted flows alike. Putting them first in the investment queue is the safest choice in most organisations.

The Türkiye context

For language models, Turkish is a lower-resource language than English. That means the texts produced may look fluent and yet still need more correction in terminology choices and in fitting the language of the organisation. The difference between fluency and accuracy has to be read more carefully in Turkish.

The structure of the language plays a part too. Because it is agglutinative and its word order is free, the emphasis in a sentence can easily shift in model output. In genres such as official correspondence, legislative text and technical instructions, that shift can affect meaning directly.

Data is a separate heading. Sending internal documents to a language model can mean transferring data to a third party. Texts containing personal data should be assessed under the KVKK, and the data retention terms of the service used should be confirmed in writing. This article is for general information only and is not legal advice.

What does this mean for you?

If you want to bring language models into your workflow, trying them in a narrow area is a safer route than spreading them everywhere at once. The sequence below makes the trial measurable.

1. Choose a single low-risk content type and set the scope of the pilot down in writing.
2. Update your term list and your style guide; let the way you steer the model rest on them.
3. Always put the output through the approval of a language specialist and record the correction time.
4. Keep files containing personal data entirely outside the scope of the pilot.
5. At the end of a three-month period, compare the quality, time and cost figures and take the decision again.

This approach makes both the gain and the risk visible. If you are not sure where to start for your own content type, you can request a quote together with sample files, and we can define the scope of the pilot with you.

Frequently asked questions

Are a large language model and neural machine translation the same thing?

They are not. Neural engines are developed for translation alone, behave more predictably and generally produce more accurate output. Language models are general purpose, can give different answers to the same prompt, and their fluency can run ahead of their accuracy.

Which jobs can language models take on in localisation?

The areas listed in the industry reports are translation, improving source content, engineering tasks such as scripting and coding, multilingual content creation, proofreading, supplying context to the linguist, next-generation product integrations and audiovisual workflows.

What does post-creator mean?

It is a new profile forecast in the industry reports. It describes the person who verifies the creative content produced by generative models, corrects it and brings it into line with the brand. It can be thought of as the counterpart of post-editing extended to content creation outside translation.

Can I publish language model output directly?

It is not advisable. The reports stress that model suggestions should pass through specialist review. Because models can produce fluent but at times inaccurate text, nothing should go live without human approval, particularly in legal, medical and financial content.

References

• Jiménez-Crespo, M. A. (2024). Localization in Translation. Routledge, pp. 249-251.
• Nimdzi Insights (2023). ChatGPT and LLMs: separating fact from fiction for localization.
• Gur, I. et al. (2022). Understanding HTML with large language models. arXiv.
• Esselink, B. (2022). Localization and translation. Routledge Handbook of Translation and Technology.