Multilingual Assets and Resources for Modular Translation

2026-07-22

Recent innovations in language technologies have introduced large language models that industry players often pitch as one-size-fit-all solutions. These systems are however rarely reusable in novel contexts, nor sustainable given their massive parameter counts, nor scalable across the thousands of languages spoken in the world. In response, we intend to build upon the findings of ERC funded project FoTran, and develop modular architectures in which parts of a network are designed to be specialized for a given task, and specialized components can be added throughout the lifetime of a model. This approach has the potential to significantly reduce the growing carbon footprint of generative AI, and will allow us to reach a broad language coverage and target under-served languages. Our MARMoT projects focuses more specifically on two goals: (i) converting our research prototypes of scalable modular architectures for language modeling and translation into production-ready software with broad applicability and (ii) developing pretrained models that can readily be used and adapted to a wide range of languages, scenarios and real-world applications. We will emphasize the support of under-resourced languages and domains in order to push for inclusiveness and language service equality. This PoC will focus on developing open tools that can be used not only in computer-assisted translation, but also in related language services such as summarization, paraphrasing and text simplification. We will produce customizable language processing assets based on a modular design that will put the emphasis on scalability, reusability, and sustainability, thereby making them efficient during inference and easily deployable by key stakeholders, from industry to government agencies. This platform will deliver language technology solutions that can be customized at the task level so as to fit a broad array of language and application contexts.

Finland

2026-01-01