Lungisanikhan/Enhancing-Semantic-Relatedness-for-Low-Resource-African-Languages-

Enhancing Semantic Relatedness for Low-Resource African Languages via Transfer Learning and M2M-100 Data Augmentation # Semantic Relatedness for Low-Resource African Languages This repository contains the implementation and results for the project: **_Enhancing Semantic Relatedness for Low-Resource African Languages via Transfer Learning and M2M-100 Data Augmentation_** The project investigates scalable and effective approaches for improving Semantic Textual Relatedness (STR) in low-resource African languages through transfer learning and multilingual machine translation–based augmentation. ## Key Features - Three African Languages: Hausa, Kinyarwanda, Afrikaans - SemRel2024 Dataset: Standardized benchmark for semantic relatedness - Transfer Learning: Fine-tuning African-centric models (AfriBERTa, AfroXLMR) - M2M-100 Augmentation: MAFAND-MT back-translation pipeline - Significant Gains: Up to **1167% improvement** over baseline models ## 📚 Research Questions **RQ1:** Which transfer-learning methods (AfriBERTa vs AfroXLMR) yield the best STR performance for African languages? **RQ2:** How effective is M2M-100 back-translation in improving model accuracy and robustness under low-resource constraints? ## 📊 Results Summary | Language | Baseline (XLM-R) | Fine-tuned (AfroXLMR) | M2M-100 Augmented | Improvement | |--------------|------------------|-------------------------|-------------------|-------------| | Afrikaans | 0.4016 | 0.4085 | 0.6452 | +60.66% | | Hausa | 0.1247 | 0.6518 | 0.6389 | +412.51% | | Kinyarwanda | 0.0425 | 0.5390 | 0.6456 | +1417.76% | **Metric:** Spearman Correlation Coefficient (ρ) ## 📁 Project Structure ``` ├── COS802_Project_Code.ipynb ├── COS_802_Proposal_Final_u25743695.pdf ├── IEEE_Paper.tex ├── README.md │ ├── results/ │ ├── all_results_mafand.csv │ ├── statistical_tests_mafand.csv │ └── final_comparison_table_mafand.csv │ ├── visualizations/ │ ├── comparison_bar_chart_mafand.html │ ├── improvement …