Kaberia-Timo/Multi-Task-Classification-of-Kiswahili-Online-Misinformation-and-Harmful-language

\# Multi-task Misinformation Detection for Low-Resource Swahili A reproducible deep learning pipeline for multi-task learning on \*\*Swahili misinformation detection\*\* and \*\*Swahili hate speech detection\*\* using shared transformer representations. This repository accompanies the research project investigating whether \*\*joint multi-task learning\*\* can improve misinformation detection by leveraging knowledge from a related hate speech detection task in a low-resource language. \--- \## Overview The project develops and evaluates two multi-task transformer models: 1\. \*\*Baseline Multi-task Model\*\* - Equal task weighting - Joint learning of misinformation and hate speech 2\. \*\*Misinformation-Tuned Multi-task Model\*\* - Increased optimisation emphasis on misinformation detection - Same architecture as the baseline - Different task-loss weighting strategy The experiments use: \- \*\*PolitikWeli\*\* (Swahili misinformation dataset) \- \*\*AfriHate\*\* (Swahili hate speech dataset) Both datasets are harmonised into a unified multi-task training dataset while preventing data leakage and ensuring reproducibility. \--- \## Repository Structure ```text multitask-misinformation-detection/ │ ├── data/ │ ├── raw/ │ └── processed/ │ ├── notebooks/ │ ├── 01\_politikweli\_prepare.ipynb │ ├── 02\_afrihate\_prepare.ipynb │ ├── 03\_build\_multitask\_dataset.ipynb │ ├── 04\_baseline\_training.ipynb │ └── 05\_misinformation\_tuned\_training.ipynb │ ├── models/ │ ├── baseline/ │ └── misinformation\_tuned/ │ ├── reports/ ├── docs/ └── src/ ``` \--- \## Workflow The notebooks should be executed in the following order. \### Notebook 1 \*\*01\_politikweli\_prepare.ipynb\*\* Purpose \- Load hydrated PolitikWeli datasets \- Clean and harmonise records \- Construct binary misinformation labels \- Perform stratified train/validation/te …