CaitlinCalder/Detecting-African-MGT

# MGT Detection for Bantu Languages - COS760 Group 45 > **Detecting Machine-Generated Text in low-resource Bantu languages (isiZulu, isiXhosa, Siswati) using a three-phase NLP pipeline: TF-IDF baseline → AfroXLM-R fine-tuning → SADiLaR morphological feature fusion.** --- ## Table of Contents 1. Project Description 2. Core Features 3. Tech Stack 4. Prerequisites 5. Google Drive Access 6. Installation 7. Environment Variables 8. Usage 9. Project Structure --- ## Project Description This research project tackles the problem of **Machine-Generated Text (MGT) detection** in under-resourced Bantu languages, specifically **isiZulu**, **isiXhosa**, and **Siswati**. Given that modern large language models (LLMs) - including ChatGPT-4o, Claude, and Gemini 2.5 Pro - can now generate fluent text in these languages, the need for reliable automated detection is growing, particularly for academic integrity and media verification contexts. The system is structured as a **three-phase pipeline**: - **Phase 1 - Baseline:** A character n-gram TF-IDF + Logistic Regression classifier trained on isiZulu and isiXhosa, then evaluated zero-shot on Siswati. - **Phase 2 - Transfer Learning:** Fine-tuning of `Davlan/afro-xlmr-base` (an XLM-RoBERTa variant pre-trained on African languages) on the same multilingual corpus, with Optuna hyperparameter search. - **Phase 3 - Feature Fusion:** A Random Forest ensemble that combines the AfroXLM-R predicted probabilities with morphological features derived from the **SADiLaR** morphological resource (lexical diversity, morphological coverage, bigram repetition, etc.). A **Streamlit web application** (`app.py`) exposes all three phases as an interactive demo, including SHAP-based explainability visualisations. Human text is sourced from the **Vukuzenzele** government newsletter corpus (via Hugging Face's `dsfsi/vukuzenzele-monolingual` dataset). Machine-generated text was produced by prompting ChatGPT-4o, Claude, and Gemini 2.5 Pro, and is sto …