The way we translate text has accelerated rapidly. While a decade ago we still laughed at the clunky, literal translations of early iterations of Google Translate, modern AI models now deliver texts that are barely distinguishable from human work. But in today's landscape, developers, content managers, and businesses face a new choice: do you use a specialized translation model (such as DeepL) or do you deploy a generative Large Language Model (such as GPT-4 or Claude)?
Although both systems use the Transformer architecture under the hood, their training objectives, strengths, and weaknesses differ significantly. In this article, we dive deep into the world of AI translations. We look at the nuances of the Dutch language, how to handle complex business terminology, and perhaps most importantly: how to test the quality of a translation if you do not speak the target language yourself.
Specialized Translation Models (NMT)
Specialized translation models, also known as Neural Machine Translation (NMT), are trained exclusively on one task: converting text from language A to language B. Well-known examples include DeepL, Google Cloud Translation, Amazon Translate, and open-source models such as Meta's NLLB (No Language Left Behind).
During the training process, these models are fed massive datasets of parallel texts (for example, an official EU document drafted in 24 languages by human translators). As a result, the model learns not only to translate word-for-word, but also to understand grammatical structures and context within a sentence.
The advantages of NMT models
- Speed and efficiency: Because they are optimized for a single task, NMT models generate translations in milliseconds. This makes them ideal for real-time applications such as chat translations.
- Predictability: NMT models rarely hallucinate. If you input the same sentence three times, you will generally get the exact same translation back three times.
- Costs: Translating millions of words via an NMT API is typically significantly cheaper than using heavy generative LLMs. Read more about this in our article on the costs of AI models.
The limitations of NMT models
The main disadvantage of specialized NMT models is their focus at the sentence level. They often lack the broader document perspective. If a term is introduced in a certain way in the first paragraph, the model might translate it differently in the fourth paragraph, simply because it has "forgotten" the earlier context. In addition, NMT models struggle with tone of voice. You cannot tell DeepL: "Translate this, but make it more enthusiastic and target it at young people."
General LLMs (Generative AI) for Translation
General Large Language Models (LLMs) such as GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro are not primarily trained to translate. They are trained to predict the next word in a sequence, based on a massive amount of multilingual training data from the internet, books, and articles.
This sounds like a disadvantage, but in practice, it turns out to be a huge strength. Because these models understand language at a fundamental level and can maintain context across long documents, they often outperform NMT models in terms of nuance, style, and creativity.
The power of context and instruction
The absolute advantage of LLMs is the ability to guide them via prompt engineering. You can provide the model with detailed instructions about the context of the text, the target audience, and the desired writing style. For those who want to learn more about this, our sister domain offers an excellent guide on prompt engineering techniques.
In addition, an LLM understands the context of the entire document. If you are translating a book and a character is named "Hunter", an LLM will understand that this is a name, whereas a simpler NMT model might translate it to "Jager" (Hunter) in an isolated sentence.
Disadvantages of LLMs for translation
- Hallucinations: LLMs can sometimes get creative and add text that was not in the source text, or omit parts they deemed irrelevant.
- Latency: Generating text with an LLM (especially the larger reasoning models) takes significantly longer than with an NMT model.
- Context Window Limits: Although context windows are huge nowadays, when translating entire books at once, you must watch out for 'lost in the middle' phenomena, where the model forgets instructions from the middle of the prompt.
The Quality of Dutch in AI Models
Dutch is considered a 'mid-resource language' in the world of AI. There is sufficient high-quality training data available (such as Wikipedia, government websites, and news archives) to enable models to generate excellent Dutch, but the data pales in comparison to the volume of English data.
A notorious problem when translating into Dutch, especially from English, is the "u" (formal) versus "je/jij" (informal) dynamic. The English "you" contains no indication of formality. An NMT model will often guess based on the vocabulary in the sentence (business terms often lead to "u"). An LLM, on the other hand, can be explicitly directed on this via the system prompt. For a detailed analysis of how different models perform in our language, we refer to our Dutch content comparison.
Handling Technical Jargon and Tone of Voice
In a business context, consistency is crucial. A 'hard drive' must be translated as 'harde schijf' across the entire website and not suddenly as 'vast station'. How do you tackle this?
Glossary in NMT
Services like the DeepL API offer the option to send a 'glossary' (word list). This is a static list of source words and their mandatory target language equivalents. This is extremely robust and literally forces the model to use specific terms, regardless of the context.
Few-Shot Prompting in LLMs
With LLMs, you work with instructions and examples. You can include a glossary in the prompt, but models sometimes tend to ignore it if they think another word fits the context better. The best way to teach an LLM corporate jargon is through few-shot prompting. You provide the model with a few examples of correctly translated sentences in the desired tone of voice.
Example of a translation prompt for an LLM:
You are an expert translator and copywriter for a Dutch tech company.
Translate the following text from English to Dutch.
Guidelines:
1. Always use the informal form of address ("je" / "jouw").
2. Maintain an enthusiastic but professional tone.
3. Strictly use the following terminology:
- "Deployment" -> "Uitrol"
- "Repository" -> "Code-omgeving"
- "User Interface" -> "Gebruikersinterface"
Text to translate:
[Insert source text here]Testing Translation Quality Yourself (Without Bilingual Experts)
One of the biggest challenges in multilingual projects is quality control. How do you know if the Japanese translation of your marketing campaign is brilliant, or if the model has just described your product as a rusty bicycle? If you do not have native speakers at your disposal, there are robust strategies to guarantee quality.
1. Back-Translation
The oldest and most reliable trick in the book. You take the generated translation (for example, in Japanese) and use an independent second AI model (preferably from a different provider) to translate this Japanese text back into your native language (Dutch or English). By comparing the original text with the back-translated text, you can immediately spot the biggest hallucinations or logical errors.
2. LLM-as-a-Judge
You can deploy an advanced reasoning model (such as GPT-4 or Claude 3.5 Sonnet) as a linguistic auditor. You provide the model with the source text, the translation, and a strict rubric (evaluation matrix). Even if you do not speak the language yourself, the model can explain to you in English whether the translation is accurate.
Evaluate the following translation from English to German.
Rate on a scale of 1 to 10 on the following criteria and provide explanations in English:
1. Accuracy: Have facts been omitted or added?
2. Fluency: Does it sound like natural, idiomatic German?
3. Style: Has the informal, persuasive tone been maintained?
Source text (English): [Text]
Target text (German): [Text]3. Technical Validation (Markup and Formatting)
In addition to linguistic quality, technical integrity is crucial. A common problem when translating web content is that AI models break or translate HTML tags, markdown, or variables (such as `{{user_name}}`). You can solve this programmatically with simple scripts that check whether the tags in the source text match the tags in the target text, before the text goes live via an API integration.
When Do You Choose Which Approach? (Cost-Benefit)
The choice between an NMT and an LLM ultimately depends on the nature of your content, your budget, and your infrastructure.
| Scenario | Recommended Model Type | Reason |
|---|---|---|
| E-commerce Product Catalog (millions of words, short descriptions) | Specialized (DeepL / Google NMT) | Low cost, high throughput, robust operation with static glossaries. LLMs are too expensive and too slow here. |
| Marketing Blogs & SEO Content (creative flair, storytelling) | Generative LLM (Claude / GPT-4) | Contextual understanding is essential. The text must not only be translated, but localized while maintaining brand voice and fluency. |
| Software UI Localization (individual buttons, menu items without context) | Generative LLM (with strict prompt) | NMTs often fail with isolated words because context is missing (is 'Home' a house or the start screen?). You can explain the context to an LLM: "This is a button in the navigation bar." |
| Legal or Medical Contracts | Hybrid (NMT + Human revision) | Absolute factuality is required. LLM hallucinations pose too great a legal risk. An NMT provides a literal base document, which must then be strictly edited by an expert. |
Conclusion
Translating with AI has shifted from a simple word-replacement task to a strategic architectural choice. For fast, predictable bulk translations, specialized NMT models remain unbeatable in cost and speed. However, as soon as nuance, context, brand identity, or complex localization come into play, general LLMs offer an unparalleled leap in quality. By making smart use of prompt engineering and automated validation techniques such as 'LLM-as-a-judge', organizations today can build fully-fledged multilingual publishing systems without relying on an army of human translators.