The translation industry is currently undergoing a profound transformation, driven by a technology that is setting new standards across all areas of language processing: Large Language Models (LLMs). In this expert interview with Jourik Ciesielski, a leading language technology specialist, you'll learn what makes these models so powerful and why they are revolutionizing the way we translate.
Our Expert: Jourik Ciesielski
Jourik Ciesielski is CTO at Yamagata Europe and founder of C-Jay International. With a master’s degree in translation and a postgraduate certificate in specialized translation, he is a recognized expert in language technology.
His focus is on machine translation, large language models, audiovisual translation, and translation management systems. With his deep understanding of both worlds—translation and technology—he is the ideal interview partner for the topic of LLMs in the translation industry.
What Are Large Language Models?
Large Language Models (LLMs) are highly advanced neural networks trained on vast amounts of linguistic data. Unlike traditional machine translation systems (MT), LLMs are not limited to a single task but can handle a wide variety of language-related tasks.
“A large language model is a neural network trained on very large amounts of linguistic data. During training, it learns to recognize patterns and relationships in the data and generates contextually relevant responses in production by predicting the most likely output for a given input.”
As Ciesielski explains, LLMs differ from the language models typically used in the translation industry (e.g., machine translation models, text-to-speech models) in several key ways:
- They can process not just text, but also images, audio, and video
- They are not task-specific but can be optimized for specific purposes
- They can be “trained” for particular tasks using natural language instructions—a process called “prompting”
How They Differ from Traditional Machine Translation Systems
| Feature | Traditional MT | LLM-Based Translation |
|---|---|---|
| Adaptability |
Traditional MT systems require large amounts of parallel corpora in the target language pairs. These must be carefully aligned and cleaned. For use in specific fields, the training data must also include domain-specific terminology and expressions. Training traditional MT engines is therefore time-consuming. |
In contrast, LLMs can often use pre-trained models and adapt them to specific translation tasks via prompts—without the need to create a specialized dataset. This allows for fast and flexible customization to user needs. |
| Creativity | Low, literal translations. Traditional MT systems are typically trained for accuracy, which may not be ideal for creative content like marketing. | High, more natural formulations. Especially in marketing, the flexibility of LLMs is advantageous—they don’t just translate, they adapt content creatively for the target culture. |
| Multifunctionality | Translation only | Translation, QA, post-editing, and more |
Use Cases of LLMs in the Translation Industry
LLMs offer diverse and ever-expanding use cases in the translation industry. Ciesielski identifies three main applications that are already production-ready:
- Automated translation using style guides and glossaries
- Automated post-editing of existing translations based on style guides and glossaries
- Automated quality assurance (QA) using style guides and glossaries
Automated Translation with Style Guides and Glossaries
When style guides and glossaries are used as prompts, LLMs can produce translations that are not only accurate in content but also comply with stylistic requirements and terminology. This is especially valuable for marketing materials, which often require audience-specific localization. For example, the LLM can be prompted to use gender-neutral language or an informal tone.
Benefits for E-Commerce and Marketing Companies: Targeted brand messaging in multiple languages without quality loss, faster time-to-market for international campaigns, increased conversion rates through culturally adapted content.
Automated Post-Editing of Existing Translations
For technical documentation where traditional MT already delivers satisfactory results, LLMs can be used via prompts to enhance terminology consistency and overall quality.
Benefits for Process-Oriented Organizations: Shorter turnaround times, improved quality, and reduced overall costs.
Automated Quality Assurance (QA)
Traditional QA checks in TMS systems like Trados and MemoQ often produce a large number of false positives and false negatives. Translators and project managers have to sift through hundreds of false alerts to find actual issues. LLMs can greatly enhance these processes through contextual understanding and adaptability.
Benefits for LSPs and Regulated Industries: More accurate detection of critical errors, transparent and traceable quality control for high compliance requirements.
“I firmly believe in the third use case—quality assurance—because I think reimagining QA is a high priority for our industry. QA checks in TMS systems typically generate large numbers of false positives and negatives.”
Limits and Challenges
Despite their enormous potential, LLMs are not a silver bullet. Ciesielski points out several key limitations:
- Misconceptions and inflated expectations: “When talking to clients or colleagues about LLMs, I often have to emphasize that LLMs cannot (directly) handle DTP tasks or modify PDF documents,” says Ciesielski.
- Hallucinations: LLMs can occasionally produce completely false outputs. These so-called “hallucinations” may become less frequent as the technology matures, but they are still a serious issue.
- Not optimal for all content types: As Ciesielski explains: “Certain types of content still benefit more from traditional MT. [...] For technical documents, traditional machine translation is often the better choice because it stays close to the original and doesn’t need to be creative.”
Case Study: Technical Documentation in Mechanical Engineering
A mid-sized German mechanical engineering company uses a hybrid approach for its technical documentation—combining traditional MT with LLM-based post-editing. The MT engine provides accurate, source-faithful translations of technical terms and instructions, while the LLM improves readability and ensures consistent terminology. This reduced manual post-editing by 38% and significantly sped up product launches in international markets.
Document-Based LLM Translation in TMS Systems: An Innovative Approach
One of the most promising strategies for overcoming the limitations of traditional MT systems in TMS environments like MemoQ or Trados is document-based LLM translation. Ciesielski has developed a prototype based on this very approach.
In TMS systems, texts are typically segmented into translation units to facilitate bilingual editing and better leverage translation memories and glossaries. While this segmentation improves consistency and QA, it can also restrict MT processing. The problem: machine translation in TMS is sentence-based, removing each sentence from its textual context—precisely the context needed for high-quality results. Ciesielski’s prototype addresses this issue.
The prototype reads XLIFF files in the TMS system, identifies translatable units (excluding full TM matches), processes them as a coherent text, and sends them in one go to an LLM for translation. The LLM returns the translated content, which is then resegmented and mapped back to its original segments in the XLIFF file. Translators can intervene as usual—editing units where needed. The key benefit: they work from a higher-quality machine translation baseline.
According to Ciesielski, this approach offers several major advantages:
- The model UNDERSTANDS the relationships between sentences (whereas with limited context, it only PREDICTS)
- This results in more accurate grammatical and semantic decisions—even for low-context prepositions and references
- Problems caused by poor segmentation are resolved automatically
- API overhead is reduced, since one document is processed in a single API call
- The framework outputs segmented results—fully compatible with translation memories
The Impact on the Translation Industry
Integrating LLMs into translation workflows has far-reaching effects on the entire industry. Language service providers (LSPs) can enhance and expand their services. Translators can focus on more creative and complex aspects of translation, while repetitive tasks become increasingly automated.
“I truly believe that LLMs have the potential to revolutionize many aspects of traditional translation workflows—from automated translation to QA and workflow management. [...] Their universal nature requires individuals to develop their own use cases, which in turn demands new skills, optimized processes, and custom integrations.”
Ready for the Future of Translation Technology?
At tolingo, we integrate advanced technologies like LLMs into our translation solutions to meet your specific needs. Discover how LLM-based translation can take your international projects to the next level.
Request a free consultationConclusion
Large Language Models are revolutionizing the translation industry with their deep contextual understanding, flexibility, and ability to generate human-like text. Jourik Ciesielski’s document-based LLM translation is a particularly promising solution to the limitations of traditional MT systems within standard TMS platforms.
While LLMs optimize existing translation workflows, they remain a tool that complements—rather than replaces—human expertise. Successfully integrating this technology requires new skills, tailored workflows, and a solid understanding of its strengths and limitations.