Parameter-Efficient Fine-Tuning For Equitable Named Entity Recognition In African Languages: A Lora-Based Approach

Abstract

Named entity recognition (NER) remains a challenge for African languages, where limited annotated data, typological diversity, and the computational cost of full fine-tuning restrict practical deployment. This research investigates whether LoRA combined with balanced multilingual sampling can simultaneously address computational efficiency and cross-language equity for African language NER. Using the MasakhaNER 2.0 dataset, we fine-tuned the pre-trained XLM-RoBERTa model using LoRA adapters with a rank of 32 covering Hausa, Yoruba, Swahili, Igbo, Kinyarwanda, and Nigerian Pidgin. A balanced sampling strategy equalises training-set contributions across languages at each gradient step. Statistical reliability is assessed through joint multi-seed evaluation across three random seeds, and sampling-strategy effects are isolated in a controlled balanced versus unbalanced ablation. The LoRA model surpasses the full fine-tuning baseline on all six languages across all three seeds. The cross-language paired t-test rejects the null hypothesis (t(5) = 2.948, p = 0.016, d = 1.203), with a mean F1 gain of +1.275 points (95% CI: [0.163, 2.387]). Joint training reduces seed-to-seed variability substantially relative to single-language training, with Yoruba's standard deviation falling from 1.256 to 0.039 and Igbo's from 1.111 to 0.379. Training completes in 5.22 GPU-hours on a single NVIDIA T4 at 12.0 GB peak VRAM. These results establish that parameter-efficient fine-tuning with LoRA achieves statistically reliable improvements over full fine-tuning baselines for African NER across typologically diverse languages, that joint multilingual training with balanced sampling substantially stabilises per-language performance across random seeds. The framework provides a scalable and computationally efficient blueprint for democratizing African NLP systems.