Why Low-Resource NLP Needs More Than Cross-Lingual Transfer: Lessons Learned from Luxembourgish

A new research paper published on arXiv examines the limitations of cross-lingual transfer in Natural Language Processing (NLP) for low-resource languages, using Luxembourgish as a primary case study. While cross-lingual transfer allows multilingual models to leverage high-resource language data, the authors argue it cannot fully substitute for language-specific efforts. The study synthesizes prior findings and data collection results, revealing a fundamental interdependence between these two approaches. Cross-lingual transfer significantly boosts performance but relies critically on high-quality, task-aligned target-language data. Conversely, such limited local resources alone are insufficient for strong performance but reach their full potential when integrated into a cross-lingual framework. The authors conclude that these methods are complementary rather than competing alternatives. They provide practical guidelines for balancing cross-lingual transfer with language-specific development to create sustainable NLP pipelines for underrepresented languages. This work highlights the need for nuanced strategies in AI development to ensure equitable technological representation for linguistically diverse communities.