Umutoniwasepie/Kinyarwanda_hatespeech_InitialSol

# Kinyarwanda_hatespeech_Detection_InitialSol ## 📖 Description The **Kinyarwanda Hate Speech Detection** project builds a machine learning system to identify hate speech in Kinyarwanda social media text, supporting safer online communities in Rwanda. It classifies tweets as negative (~35%), neutral (~38%), or positive (~27%) using: - **Logistic Regression with TF-IDF**: Baseline model, ~56–57% accuracy, ~0.57 F1-macro. - **XLM-RoBERTa**: Transformer model with class weights `[2.86, 2.63, 3.70]`, ~52% accuracy. Features: - **Preprocessing**: Removes URLs, mentions, special characters, Kinyarwanda stopwords. - **Visualizations**: Pie charts, boxplots, histograms, word clouds, n-grams (e.g., `mubi`, `neza`). - **Tokenization**: Verified with XLM-RoBERTa. - **Deployment**: Streamlit app with Kinyarwanda UI. Future work possibility: KinyaBERT, lightweight augmentation. ## 🔗 GitHub Repository github.com ## 🛠️ How to Set Up the Environment and Project ### Environment Setup - **Kaggle**: - Create a notebook with Python 3.11, T4 x2 GPU. - Upload `notebook/kinyarwanda-hatespeech.ipynb`. - Add dataset “kinyarwanda-hate-speech-data” (`/kaggle/input/data/`). - **Local**: - Use Python 3.11. - Open `notebook/kinyarwanda-hatespeech.ipynb` in VS Code/Jupyter. - Place `train.tsv`, `dev.tsv`, `test.tsv` in `data/`. - Install dependencies from `requirements.txt`. ### Project Setup 1. Clone: `git clone github.com` 2. Data: Verify `data/` has `train.tsv` (3302 samples), `dev.tsv` (827), `test.tsv` (1026). 3. Notebook: Run `kinyarwanda-hatespeech.ipynb` in Kaggle (~30–60 min). 4. App: Run `streamlit run app/app.py`, test tweets (e.g., “umuntu mubi”). ## 🎨 Designs ### Figma Mockup - **Description**: UI with 200x100px text area (“Andika ubutumwa...”), “Gusuzuma” button, prediction table, “Tanga” feedback button. White background, Roboto font. - **File**: desig …