Global Voices, Local Frames: Cross-Lingual Corpus Analysis of Stance and Discourse in Social Media and News

Most stance and discourse corpora are English-centric and biased toward the Global North, limiting our ability to study how publics across the world frame and take positions on global issues. This paper introduces the Global Voices, Local Frames Corpus (GVLF-C), a large-scale, cross-lingual dataset covering South Asian, Sub-Saharan African, and Latin American news and social media texts (2015–2025). The corpus includes 12 languages and ~120K manually annotated documents, labeled for stance (pro/anti/neutral) and frames (economic, justice, identity, urgency, and others). We benchmark state-of-the-art multilingual models (mBERT, XLM-R, LaBSE, NLLB embeddings), demonstrating that few-shot annotation in low-resource languages yields substantial performance gains, while frame detection remains harder due to conceptual overlaps. Diachronic analyses reveal event-driven framing shifts around COP summits, natural disasters, and elections, with distinct regional emphases: justice and indigenous rights in Latin America, adaptation and vulnerability in Africa, and national identity/security in South Asia. Our findings underscore the need to decenter Global North perspectives in NLP resources, highlight cultural variability in framing, and propose a path forward for more inclusive, diachronic, and representative discourse analysis.