Self-Supervised Contrastive Learning for Low-Resource Classification Tasks

Abstract

The scarcity of annotated training data represents one of the most pressing bottlenecks for the deployment of artificial intelligence systems in Nigeria and across sub-Saharan Africa. This challenge is particularly acute in domains such as Nigerian indigenous language processing (covering Yoruba, Igbo, Hausa, Efik, Tiv, and over 500 other languages), healthcare diagnostics at resource-limited facilities, and agricultural disease detection in smallholder farming communities. In this paper, we present SSCL-LR (Self-Supervised Contrastive Learning for Low-Resource), a novel unified framework that combines momentum-based contrastive representation learning with task-adaptive prototype alignment to achieve strong classification performance under severely constrained annotation budgets. Our approach introduces two key technical innovations specifically designed for the Nigerian and broader African low-resource context: (1) a Hierarchical Augmentation Strategy (HAS) that generates semantically consistent positive training pairs across linguistic, visual, and medical modalities; and (2) a Dynamic Prototype Alignment (DPA) mechanism that progressively refines class decision boundaries using limited labeled examples combined with abundant unlabeled data. We evaluate SSCL-LR on eight benchmark datasets, including four Nigerian-specific corpora: the Hausa-NLP News Classification dataset, the Yoruba Sentiment Corpus (YorubaSenti), the Nigeria Centre for Disease Control (NCDC) clinical notes dataset, and the Cassava Leaf Disease dataset sourced from smallholder farms in Benue and Nasarawa States. Across all benchmarks, SSCL-LR achieves 4.1%--9.3% improvements over state-of-the-art baselines under 1-shot, 5-shot, and 10-shot settings. Our code, pre-trained models, and the newly curated NaijaLowRes benchmark suite are publicly released at https://github.com/acair-unilag/sscl-lr.