NLP Summative Project, African Leadership University.
Classifies Kirundi news articles into six topics: business, entertainment, health, politics, religion, sports.
MasakhaNEWS, Kirundi subset (run), from Hugging Face.
1,117 train / 159 validation / 322 test articles.
| Model | Accuracy | Macro-F1 | Weighted-F1 |
|---|---|---|---|
| TF-IDF + Logistic Regression (baseline) | 0.863 | 0.776 | 0.859 |
| Word2Vec + BiLSTM | 0.839 | 0.744 | 0.835 |
| AfriBERTa (fine-tuned) | 0.916 | 0.853 | 0.915 |
| AfroXLMR (fine-tuned) | 0.901 | 0.829 | 0.899 |
notebooks/: data exploration, baselines, AfriBERTa, AfroXLMR, error analysisresults/: figures, metrics, predictions, error listapp/: web app codeIn progress: full methodology and run instructions coming.
NLP Summative Project, African Leadership University.
Classifies Kirundi news articles into six topics: business, entertainment, health, politics, religion, sports.
MasakhaNEWS, Kirundi subset (run), from Hugging Face.
1,117 train / 159 validation / 322 test articles.
| Model | Accuracy | Macro-F1 | Weighted-F1 |
|---|---|---|---|
| TF-IDF + Logistic Regression (baseline) | 0.863 | 0.776 | 0.859 |
| Word2Vec + BiLSTM | 0.839 | 0.744 | 0.835 |
| AfriBERTa (fine-tuned) | 0.916 | 0.853 | 0.915 |
| AfroXLMR (fine-tuned) | 0.901 | 0.829 | 0.899 |
notebooks/: data exploration, baselines, AfriBERTa, AfroXLMR, error analysisresults/: figures, metrics, predictions, error listapp/: web app codeIn progress: full methodology and run instructions coming.