henryparfait/kinnews-topic-classifier

Kinyarwanda News Topic Classifier

castorini/afriberta_base fine-tuned to classify Kinyarwanda news articles into 14 topics (KINNEWS categories). Built as an individual NLP capstone project.

Training data

KINNEWS (Niyongabo, Hong, Kreutzer and Huang, 2020; MIT license). Exact duplicates (over half of the raw training rows) and 156 articles with conflicting labels were removed; 10% of the remaining training data was held out for validation.

Training

Max length 512 tokens, AdamW (lr 3e-5, linear warm-up and decay), class-weighted cross-entropy, early stopping on validation macro-F1.

Results (deduplicated KINNEWS test set, 3464 articles)

Metric	Score
Accuracy	0.7546
Macro-F1	0.5810
Weighted-F1	0.7509

Intended use and limitations

For research and demonstration on Kinyarwanda news text. Rare categories (tourism, fashion, environment, education) are much less reliable; class distributions differ strongly between the KINNEWS train and test splits; the model can partly rely on source names and bylines; it was not trained on social media, non-news or non-Kinyarwanda text.

Code and full report: https://github.com/henryparfait/kinyarwanda-news-classifier