ninesowngoal/low-resource-mt-bias

Code and resources for my MSc dissertation on bias in AI translation models for low-resource languages. Includes preprocessing scripts, training configurations, evaluation metrics (BLEU, chrF, ROUGE, TER), and analysis tools for comparing baseline vs adapted NLLB models. # Low-Resource MT Bias This repository contains the code and resources for my MSc dissertation project: **Bias in AI Translation Models for Low-Resource Languages**. The project evaluates how machine translation models handle underrepresented languages (e.g., Bemba, Swahili) and explores techniques such as LoRA adaptation and data augmentation. ## Features - Preprocessing scripts for low-resource parallel data - Training configurations for baseline and adapted NLLB models - Evaluation scripts using BLEU, chrF, ROUGE-L, and TER - Example notebooks for analysis and reproducibility ## Installation Clone the repository and install dependencies: ```bash git clone github.com cd low-resource-mt-bias pip install -r requirements.txt ``` ## Usage ### Run the Notebook You can explore and reproduce the experiments directly from the Jupyter notebook: ```bash jupyter notebook swahili_to_english_small_model.ipynb ``` ### Run the Python Script For running the translation pipeline from the command line: ```bash python swahili_to_english_small_model.py ``` This script demonstrates how the model processes parallel data (Swahili - English) using the provided XML files. ## Data The repository includes sample XML files: - eng.xml — English source data - swh.xml — Swahili target data These files are illustrative. For larger-scale training, prepare your own dataset in a similar format. ## Results Evaluation results (BLEU, chrF, ROUGE-L, TER) will be generated by running the notebook or script. Example placeholder metrics (from test runs): - BLEU: 12.9 - chrF: 43.6 - ROUGE-L (F1): 0.38 - TER: 81.6 ## Citation If you use this reporsitory, please cite: ```bash Chishimba Chipeta, (2025). Bias in AI Translation Models for Low-Resource Languages. MSc Dissertation, Birmingham City University. ```