meenah-imam/hausa-sentiment-codemix

# Hausa Sentiment Analysis: Code-Mix vs Monolingual This project investigates the performance of sentiment analysis models on **Hausa language data**, comparing: * Monolingual Hausa text * Code-mixed Hausa-English text * Combined datasets The goal is to understand how **code-mixing affects model performance** in low-resource African NLP. --- ## Dataset The experiment uses the **AfriSenti Hausa dataset** (Muhammad et al., 2023), accessed via Hugging Face. * ~22,000 tweets * Pre-split into train, validation, and test * Labels: Positive, Negative, Neutral --- ## Methodology The study follows a controlled experimental setup: 1. Load and preprocess Hausa tweets 2. Create: * Monolingual dataset * Code-mixed dataset * Combined dataset 3. Train transformer-based models (e.g., XLM-RoBERTa) 4. Evaluate performance using: * Accuracy * F1-score * Confusion matrix --- ## Experiments Three training conditions: * **Monolingual only** * **Code-mixed only** * **Combined dataset** All datasets are **size-matched** to ensure fair comparison. --- ## Results Results include: * Class distribution analysis * Model performance metrics * Visualizations of comparison across setups (See `/outputs/figures` for plots) --- ## How to Run ### 1. Install dependencies ```bash pip install -r requirements.txt ``` ### 2. Run notebook ```bash jupyter notebook notebooks/Codemix_vs_Monolingual.ipynb ``` --- ## Requirements * Python 3.10+ * PyTorch * Hugging Face Transformers * Datasets * Scikit-learn * Matplotlib --- ## Citation If you use this work, please cite: AfriSenti: Muhammad et al. (2023) --- ## Author Amina Imam Abubakar PhD in Applied AI (NLP) Focus: Low-resource African languages, Sentiment Analysis --- ## Notes This project is part of ongoing research on: * Code-mixing in African NLP * Low-resource language modeling * Responsible AI in multilingual systems