Translating LHCb's Documentation: First experiences and ensuring maintenance

The Large Hadron Collider, the world's most powerful particle accelerator, sits beneath the border of France and Switzerland, smashing protons together to recreate the conditions of the universe just after the Big Bang. One of the experiments housed there, known as LHCb, focuses on understanding why the universe is made of matter rather than antimatter. To do this, thousands of scientists from around the globe must work together, using complex software to sift through billions of collision events. However, this software is not static; it changes constantly as the machine itself is upgraded. For new researchers joining the team, learning this ever-shifting digital landscape has traditionally been a steep climb, relying on a set of online tutorials and annual workshops held in Geneva. As the experiment entered a new phase of operation, the team faced a practical challenge: how to keep these vital training materials accurate, up-to-date, and accessible to a growing number of collaborators who speak languages other than English, particularly those in China.

The story of how the LHCb collaboration solved this problem begins with the realization that their existing training materials, known as the Starterkit, needed a complete overhaul. The experiment had just undergone a massive upgrade, and the software used to analyze data had changed so significantly that the old lessons were no longer useful. The team wrote a new version of these lessons, but they also recognized a demographic reality: China hosts the largest group of LHCb scientists outside of Europe. Many of these researchers could not easily travel to Switzerland for the annual in-person workshop due to distance and visa requirements. To ensure these scientists were not left behind, the team decided to create a parallel training event in China, which required translating the entire new set of lessons into Mandarin.

Translating hundreds of pages of highly technical physics documentation is a daunting task, especially when the text contains complex code and specialized jargon that standard translation tools often get wrong. The researchers found that if they simply asked an artificial intelligence to translate the text, the result was frequently inaccurate, with the machine inventing technical terms or misinterpreting physics concepts about 20 percent of the time. To fix this, they built a custom system that combined the speed of AI with the precision of human experts. They first trained a local AI model on a verified dictionary of physics terms, ensuring the machine understood the specific vocabulary of the experiment. The AI would then generate a draft translation, but this draft was never published directly. Instead, it was routed to a panel of bilingual physicists who reviewed the work, correcting any subtle errors and ensuring the flow of the language felt natural to a native speaker.

To keep this system running smoothly over time, the team designed a maintenance workflow that prevented the Chinese version from falling behind the English original. They realized that if the English lessons were updated, the Chinese version would need to change immediately to remain useful. They established a process where any change to the English text automatically triggered a notification to a dedicated liaison in China. This person would then use the AI-assisted tools to translate the new material and verify it. Crucially, the team also separated the computer code examples from the text itself. By storing the code in separate files that were automatically inserted into the lessons, they ensured that the code remained identical in both languages. This allowed them to run automated tests on the code every time a change was made, guaranteeing that the examples would actually work for a student trying to follow along, regardless of which language they were reading.

The success of this approach was measured by tracking how many people used the translated materials. After the Chinese lessons were launched, the team observed that roughly 22 percent of all visits to the Starterkit website involved a user accessing at least one page in Mandarin. This steady usage indicated that the translation was not just a formality, but a vital resource for a significant portion of the collaboration. The data also showed a clear pattern of engagement, with a noticeable dip in traffic during the Lunar New Year holiday, reflecting the real-world rhythms of the scientists using the site. The project proved that it was possible to maintain two versions of a complex technical document simultaneously without creating a bottleneck, ensuring that the Chinese-speaking community could participate fully in the experiment's training and analysis.

This effort has paved the way for the first "sibling" workshop to be held in China, taught entirely in Mandarin, complementing the traditional event in Europe. The system the team built is flexible enough that if other groups of scientists, such as those in Brazil, wished to hold their own local training events, the same translation and maintenance framework could be adapted for them. By combining automated tools with human oversight, the LHCb collaboration has ensured that language is no longer a barrier to entry for new scientists, allowing the global team to work together more effectively as they continue to explore the fundamental nature of the universe.