P-Adjinsala/misinformation-detection-africa

This notebook is part of ongoing exploration into how AI and NLP can be applied to detect misinformation — a growing challenge in African digital and civic spaces, particularly around elections, health, and governance. # 🛡️ Misinformation Detection for African News Contexts ### NLP-based text classification pipeline for information integrity monitoring --- ## Overview Misinformation is one of the most pressing challenges facing African digital spaces — spreading through WhatsApp chains, social media, and informal news channels faster than fact-checkers can respond. This project builds a **reproducible NLP pipeline** that classifies text statements as *reliable* or *potentially misleading*, using the LIAR benchmark dataset as a foundation. The work is grounded in a civic tech context: rather than treating misinformation detection as a purely technical problem, this notebook frames it as a **governance and digital rights challenge**, examining both the capabilities and the limits of automated detection in African information ecosystems. --- ## What this project demonstrates | Skill | Detail | |-------|--------| | **NLP fundamentals** | TF-IDF vectorisation, text preprocessing, n-gram features | | **ML classification** | Logistic Regression baseline with cross-validation | | **Model evaluation** | Precision, recall, F1-score, ROC-AUC, confusion matrix | | **Interpretability** | Top predictive features visualised per class | | **Civic framing** | Explicit discussion of African context, language gaps, and ethical limits | --- ## Key results - **ROC-AUC: ~0.74** on LIAR test set using TF-IDF + Logistic Regression - **5-fold CV F1** consistently above 0.65 — solid baseline for an interpretable model - **Confusion matrix** and **top feature analysis** included for full transparency --- ## Visualisations The notebook produces 4 publication-quality charts: 1. Label distribution (original 6-class + binary) 2. Statement length by class 3. Confusion matrix + ROC curve 4. Top 15 predictive words for each class (Fake vs Real) --- ## Limitations & next steps This baseline intentionally uses an interpretable model over a black-box transformer — in civic contexts, **explainabili …