CodeHermez/African-Langs-For-Sentiment-Analysis-Using-AfriSenti-Datasets-Sentiment-Analysis-In-African-Langs-

A Comparative Study of Monolingual and Multilingual Transfer Learning Strategies with Code-Mixing Analysis # Group 60 — Sentiment Analysis in African Languages ## COS760 2025 | AfriSenti NLP Project **Members:** - Praises Obi (u26819661) - Bob Dlamini (u21739120) - Ishe Allen Chihobo (u22592238) --- ## Overview This project performs multilingual sentiment analysis across five African languages — **Hausa, Yoruba, Igbo, Nigerian Pidgin, and Swahili** — using the AfriSenti-SemEval 2023 dataset. We implement and compare four model architectures: | Model | Description | |---|---| | **TF-IDF + Logistic Regression** | Baseline model using n-gram features | | **LAFT** | Language-Adaptive Fine-Tuning of AfroXLMR (monolingual, per language) | | **mBERT** | Multilingual BERT fine-tuned per language | | **MAFT** | Multilingual Adaptive Fine-Tuning of AfroXLMR (joint training across all 5 languages) | Phase 4 extends the analysis with LIME interpretability, code-mixing diagnostics, and an error taxonomy based on Muhammad et al. (2023). --- ## Contents of the Zip File ``` Group60/ ├── README.md ← This file ├── group60-notebook.ipynb ← Main project notebook (all 4 phases) ├── download_data.py ← Standalone script to download the AfriSenti dataset └── requirements.txt ← Python library dependencies ``` --- ## Setup Instructions ### Requirements - Python 3.10+ - CUDA-compatible GPU strongly recommended (notebook was developed on Kaggle with a T4 GPU) - Install dependencies: ```bash pip install -r requirements.txt ``` > **Note:** The notebook also includes `!pip install` cells at the top of each phase that handle installation automatically when run in Kaggle. --- ## Running the Code This project is designed to run as a **Kaggle notebook**. Follow these steps: ### Step 1 — Upload the notebook to Kaggle 1. Go to kaggle.com and sign in 2. Click **Code → New Notebook** 3. Go to **File → Import Notebook** and upload `group60-notebook.ipynb` ### Step 2 — Attach the pre-computed assets dataset The model checkpoints, cached results …