AbubakariSadikOsman/hausa-fake-news-xnlp

# Explainable Cross-Lingual Fake News Detection for a Low-Resource African Language **ICS4554 Natural Language Processing — Final Project** > GitHub repository: ` ` > Team: ` ` ## 1. What this project does Fake news and health/political misinformation spreads readily on African social media, but almost all automated fake-news detection research and tooling is built for English. This project builds, evaluates, and critically analyzes a misinformation detector for **Hausa** (~90M+ speakers, Nigeria/Niger/Chad/Ghana), with three things most fake-news classifiers skip: 1. **A real low-resource baseline-to-transformer comparison** — TF-IDF/classical ML vs. a fine-tuned multilingual transformer (AfroXLMR / mBERT), not just "we fine-tuned BERT and got X% accuracy." 2. **Explainability that is actually evaluated, not just visualized** — LIME token attributions plus an AOPC-style *faithfulness* check (does deleting the highlighted tokens actually change the prediction more than deleting random tokens?). 3. **Cross-lingual zero-shot transfer to Yoruba and Swahili**, using machine-translation to construct an honestly-labeled *synthetic* test set, because no labeled fake-news corpus exists yet in those languages. 4. **Ethical safeguards baked into the inference path**, not bolted on: a confidence-based human-review flag, a bias audit across content sources, and a non-negotiable disclaimer against using the model for automated takedowns. ## 2. Repository structure ``` hausa-fake-news-xnlp/ ├── README.md ← you are here ├── requirements.txt ├── LICENSE ├── data/ │ ├── external/ ← raw clone of the source dataset repo (gitignored) │ ├── raw/ ← hausa_misinformation_dataset_v1.0.csv (gitignored, fetched at runtime) │ └── processed/ ← train/val/test splits produced by src/data/preprocess.py ├── src/ │ ├── data/ │ │ ├── download.py ← fetches the dataset (git clone + fallback) │ │ └── …