Transfer Learning for Multilingual Classification of Scientific Texts and Skills to the Sustainable Development Goals: A Systematic Review and Research Agenda

The United Nations Sustainable Development Goals (SDGs) have become a central framework for assessing the societal relevance of research. Automated classification of scientific texts and workforce skills against the SDGs is increasingly required by funding agencies, universities, and policymakers. However, existing systems face two critical limitations: a strong English-language bias and methodological fragmentation between scientific text classification and skill-SDG mapping. This systematic review, conducted according to PRISMA 2020 guidelines, analyzes 47 peer-reviewed studies published between 2018 and 2025 that apply transfer learning to multilingual SDG classification. Results show that multilingual BERT (mBERT) and XLM-RoBERTa dominate the technical landscape, with XLM-R-large achieving state-of-the-art performance (F1 = 85.66%) on Indonesian scientific texts. Nevertheless, language coverage remains highly skewed: pure multilingual studies without English dominance constitute only 8.5% of the literature, while African, South Asian, and Middle Eastern languages are virtually absent. Parallel research streams on scientific text classification and skill mapping operate in methodological silos, preventing cross-fertilization of techniques such as FAISS-based retrieval and zero-shot cross-lingual transfer. Social SDGs (e.g., Goals 5, 10, and 16) consistently underperform environmental and technical goals due to their reliance on implicit, context-dependent language. In response, this review proposes a five-pillar research agenda centered on unified multilingual frameworks, parameter-efficient adaptation for low-resource languages, semantic enhancement of social SDGs, standardized cross-lingual evaluation protocols, and human-AI collaborative systems. Advancing this agenda requires interdisciplinary collaboration across computational linguistics, sustainability science, and information science to develop truly inclusive and interoperable SDG classification systems.