k-aff/COS760-Project

Exploring Semantic Relatedness in African Languages using Transfer Learning and Data Augmentation # COS760 Group 20 - Cross-lingual Semantic Relatedness for African Languages This project investigates whether NLP models can transfer knowledge about semantic relatedness across African languages, with a focus on Bantu language families and South African languages. ## Research Question Do African languages from the same family share enough structure to transfer semantic relatedness knowledge between them? We fine-tune AfroXLMR and XLM-R on Kinyarwanda relatedness data and test zero-shot transfer to Hausa and Amharic, then evaluate how well the models cluster South African Bantu languages they have never seen with relatedness labels. ## How to Run The script `src/complete_notebook.py` runs the full pipeline and works on both Google Colab and a local machine, no changes needed either way. It auto-detects its environment and handles paths and login accordingly. --- ### Option A: Google Colab (recommended, free GPU) **Step 1: Get the file** Go to the GitHub repo, click on `src/complete_notebook.py`, then click the download icon (or right-click Raw → Save As) to download the file. **Step 2: Open Google Colab and upload the file** Go to colab.research.google.com and open a new notebook. In the left sidebar, click the **folder icon** to open the Files panel, then click the **upload icon** and upload `complete_notebook.py`. **Step 3: Enable GPU** ``` Runtime > Change runtime type > T4 GPU → Save ``` **Step 4: Set up your Hugging Face token** Create a free account at huggingface.co and get a read token at huggingface.co. In Colab, click the **key icon** in the left sidebar, then: - Click **Add new secret** - Name: `HF_TOKEN` - Value: paste your token - Toggle **Notebook access ON** **Step 5: Run** Because the script runs as a subprocess, Colab Secrets aren't directly accessible inside it. First run this in a cell to export the token so the subprocess can see it: ```python import os from google.colab import userdata os.envi …