Andrea-Mwonge/NLP-Luganda-Project

# Luganda Sentiment Analysis — Prototype A working proof-of-concept for *"A Comparative Evaluation of Transformer-Based Models for Sentiment Analysis of Low-Resource Luganda Social Media."* It demonstrates the full pipeline your proposal describes (Chapter 3) end-to-end on a small scale, so you can show your advisor a running system while you build out the real dataset. ## ⚠️ Important: the sample data is a placeholder `data/sample_data.csv` contains 40 short Luganda sentences I wrote using general vocabulary knowledge — **it has not been verified by a fluent Luganda speaker or checked for dialectal accuracy.** Do not use it as your research dataset or cite results from it in your final report. Its only purpose is to let the training/demo code run so you can see the pipeline work. Swap it out for real annotated data before you generate any results for your report (see "Scaling up" below). ## What's included ``` luganda_sentiment_prototype/ ├── data/sample_data.csv # placeholder dataset (40 rows, 3 classes) ├── src/train.py # fine-tunes a transformer + reports metrics ├── src/app.py # Gradio demo: type text, get sentiment ├── requirements.txt └── README.md ``` ## How to run it **Recommended: Google Colab** (free GPU, avoids local setup issues) 1. Upload the whole folder to Colab or clone it into a Colab notebook. 2. Set runtime to GPU (Runtime → Change runtime type → T4 GPU). 3. Run: ```bash pip install -r requirements.txt python src/train.py --model xlm-roberta-base --epochs 3 ``` 4. This trains on the 70/15/15 split, prints accuracy/precision/recall/F1 on the test set, and saves the model to `model_output/`. **Try another model** — matches your Objective 2 (BERT, RoBERTa, XLM-R): ```bash python src/train.py --model bert-base-multilingual-cased python src/train.py --model roberta-base ``` (Note: plain `roberta-base` is English-only; for a fair multilingual comparison, consider `xlm-roberta-base`, `bert-base-multilingual-cased`, …