Mistakes Large Language Models Make in Translation
When we wrote about the typical mistakes made by large language models, or LLMs, such as ChatGPT some time ago, the weaknesses of machine translation were quite obvious. Models often left out parts of the text, hallucinated or made basic grammar and terminology mistakes.
Today, technology has improved significantly. Modern models rarely make serious grammar mistakes. At first glance, the texts they produce are fluent, clear and grammatically correct.
However, the problems have not disappeared. They have simply become more subtle. Because the mistakes are less obvious, casual readers may find them harder to notice. Even so, they can still change the meaning, reduce accuracy or make the text sound unnatural.
Below are some of the most common problems we find in translations produced by modern language models and correct during professional post-editing.
1. Source-Language Sentence Structure
Although language models understand the grammar of many languages quite well, they often keep the sentence structure of the English language. As a result, they may transfer not only the meaning, but also the word order, sentence length and grammatical patterns of the original text. The translation may be technically correct, but it can sound stiff, unnatural or difficult to read.
- Example: Models often copy English word order and use passive forms too frequently. They may translate “The software allows the user to easily customize the dashboard” very literally.
- Better solution: An experienced translator will restructure the sentence so that it sounds natural in the target language, for example: “You can easily customize the dashboard in the software.”
A good translation should sound as though it was originally written in the target language.
2. Repetitive Style and Generic Expressions
When translating marketing or general-interest texts, language models tend to use very predictable vocabulary. They often rely on the same generic words and phrases, which can make very different texts sound alike.
Common signs of unedited machine-generated text include the excessive use of words such as “key,” “transformation,” “seamless,” “innovative” or “in-depth.” Models also frequently use filler phrases such as “It is important to mention…” or “In conclusion…”. A human translator can give the text a more distinctive voice and adapt it to the tone, audience and communication style of a particular company.
3. Poor Understanding of the Wider Context and Purpose
Models are often very good at translating individual sentences. However, they may struggle to connect information across an entire document. They may also fail to understand where and how the text will be used. A translation for a user interface, technical manual, medical device or advertising campaign may require very different wording, even when the original sentence is the same.
Example: ChatGPT translated the phrase “NEW BORN” as “newborn baby,” even though the text referred to the Cupra Born car.
Without a clear understanding of the wider context, even a grammatically correct translation can be completely wrong.
4. Over-Translating Established Terms
Because language models are designed to produce a translation for almost every part of the text, they sometimes translate terms that should remain unchanged.
When they come across established international, technical or industry-specific terms, they may try to replace them with newly created local-language versions.
These translations may be linguistically possible, but professionals may not use them in practice. This can make the terminology confusing and the entire document look unprofessional.
A professional translator knows when a term should be translated, when an established equivalent should be used and when the original term should remain unchanged.
Need a translation?
Contact us and we'll make you a personalized offer.
Fill out the form and we will prepare a customized offer for you. For more information you can also call us at +386 1 620 26 16.