Scaling Frame Analysis to Genuinely Low-Resource Languages: A Case Study in Swahili and Tamil

Recent advances in multilingual news framing analysis have shown promise for low-resource settings through code-switching techniques. However, existing evaluations have focused on relatively high-resource languages like German, Turkish, and Arabic. This paper extends this line of research to genuinely low-resource languages—Swahili and Tamil—that face significant representation gaps in existing multilingual models. We introduce new annotated datasets for gun violence framing in these languages and systematically evaluate the code-switching approach under extreme low-resource conditions. Our results show that while code-switching provides consistent improvements over zero-shot transfer, the absolute performance gap between high-resource and genuinely low-resource languages remains substantial (15-20% F1-macro). We identify linguistic distance and morphological complexity as key challenges and propose adaptations to the code-switching method that yield 7% average improvement. Our work provides the first comprehensive analysis of cross-lingual frame detection in truly low-resource scenarios and establishes benchmarks for future research.