Behind the news: unlocking AI opportunity in Luxembourgish

On 11 February we published the Microsoft announcement to fund SnT research to create a new type of AI model for Luxembourgish, LuxVLD. The project is carried out within the CVI2 research group at SnT, under the leadership of Prof. Djamila Aouada. Here we go behind the story to meet Dr. Nesryne Mejri, a postdoctoral researcher working on the LuxVLD project, and Eya Khamassi, a Research and Development Specialist and a UX designer for the related education platform Wingos, to understand what the project is about.

The original announcement can be found here.

Why is LuxVLD important?

Nesryne Mejri: From optimising day-to-day activities to starting a new business, AI tools with multimodal understanding equal access to opportunity.  That opportunity depends on systems that can interpret not only the language a person speaks, but also the images they share and the real-world context in which they operate. 

Today that isn’t possible for Luxembourgish, as not enough material in that language exists to train an AI model with textual and visual understanding. Data needs to be created to train a model, and that is the purpose behind the LuxVLD research project. This is also why it is one of the projects selected for Microsoft’s LINGUA programme. LINGUA supports projects that build open, high-quality datasets for under-represented languages in Europe. As part of the LINGUA programme, selected projects receive financial support for dataset creation as well as technical guidance through Microsoft’s collaboration with the APERTUS research consortium.

Yet AI tools already offer Luxembourgish, so why do we need LuxVLD?

Nesryne Mejri: When you use an AI tool in Luxembourgish today, the language is going through translation. If you ask a chatbot something in Luxembourgish, it translates it into English to understand it, develops the answer, and then translates it back into Luxembourgish for the user. 

This approach is prone to losing meaning and nuances in translation. For those who have learnt a language through study, that moment when you understand the words but not what has been said is probably familiar. In reverse, it makes it difficult for the person speaking the language to make themselves understood clearly. That turns into a limitation when using AI tools, which in turn, leads to a decline in user trust, interest, and engagement of these AI tools.  

LuxVLD aims to fix this by developing a vision–language model that can reason directly in Luxembourgish, without going through another language.

What is a vision–language model?

Nesryne Mejri: A vision–language model processes both images and text together. It can look at an image and describe it in language, or interpret text in connection with visual information. This is important because many real-world applications, such as accessibility tools, digital assistants, and educational platforms, rely on visual context. 

In our case, we want the model to interpret images and generate descriptions directly in Luxembourgish. That avoids the translation layer and preserves nuance.

Why doesn’t this exist already?

Nesryne Mejri: Luxembourgish is considered a “low-resource language”. This means that there is limited digital data available for training AI systems, unlike for languages like English. Large language models (LLMs), which are the foundation of current AI text generation tools, depend on enormous datasets to learn patterns in language. Luxembourgish simply does not have the same digital footprint as larger languages. 

This scarcity affects not only vocabulary but also cultural alignment. AI systems trained mostly on English-language data may struggle to recognise references specific to Luxembourgish society. For example, the question “Wien ass den Heng?” (in English, “Who is Heng?”), the response would be “Den Heng ass e mann” (“Heng is a man”), missing entirely the cultural reference that “Heng” in Luxembourgish refers to Grand-Duc Henri.

How will LuxVLD solve this?

Nesryne Mejri: For the LuxVLD project we are launching a national data collection campaign. We need to create a new dataset made up of image-text pairs, and to do that we will ask native speakers across different regions of the country to describe images for us. 

These image-text pairings, provided by native speakers, will allow us to create what is called a “multimodal” dataset. By combining both text and visual elements, this dataset will enable an AI model to reason directly in the language and capture much of the cultural context that is currently missing from Luxembourgish AI systems.

How are you collecting this data?

Nesryne Mejri: The data collection campaign will take place online as well as in person in multiple locations around the country. By involving contributors from different parts of Luxembourg, we will capture the linguistic diversity and regional variation that characterise the language. Practically, this means including multiple different descriptions per image.

Once you have the data, what happens?

Nesryne Mejri: This is when the possibilities expand. Both the dataset and the resulting model will be made openly available, allowing other researchers and developers to build on this foundation. For us, LuxVLD is not just a research project — it also connects to our research group’s start-up project, Wingos.

What is Wingos?

Eya Khamassi (Research and Development Specialist and a UX designer for Wingos): Wingos is an AI-powered learning platform that allows teachers to adapt learning experiences to the specific needs of their students. It is already being beta-tested in one Luxembourgish school. Using it, teachers generate curriculum-aligned exercises and provide personalised learning paths for students. Teachers remain in control as they review and adapt the AI-generated content, but the tool enables them to offer far more tailored learning experiences than before.

At the moment, Wingos relies on general language models. These are powerful, but they are not optimised for Luxembourgish, an important language within the Luxembourgish school system. LuxVLD will provide the foundation to make tools like Wingos more accurate and culturally aligned.

Why is this particularly relevant for education?

Eya Khamassi: Trust and clarity are essential in education. If students encounter awkward phrasing or mistakes in Luxembourgish, they may question the validity of their learning materials. For teachers, this means using the platform becomes more time-consuming than it should be, as they need to correct the content. 

By improving how AI understands Luxembourgish, we reduce friction for users. In turn, this makes the tool more inclusive which is particularly important in Luxembourg, where language plays a key role in education.

Looking ahead, what do you hope this project demonstrates?

Eya Khamassi: Meaningful AI innovation can start locally. Addressing the needs of a smaller language community is not a niche problem, it is part of building equitable technology. Wingos follows a human-in-the-loop approach, where AI supports teachers while they remain in control. With LuxVLD, we can ensure this collaboration truly serves the Luxembourgish community with accurate and culturally aligned learning experiences. 

Nesryne Mejri: Responsible AI development must consider linguistic diversity from the beginning. Language is not just a technical parameter; it shapes access, inclusion, and opportunity.

Read also…

• Microsoft Press Release