Banjiola/dissertation

Automatic diacritic restoration of Yoruba Ife dialect using mT5-small and byT5-small # Evaluating a Multilingual Transformer Model for Automatic Restoration of Yorùbá Diacritics: A Study on Ifẹ̀ Dialect ## 1. Overview A diacritic is a mark appended above or below a letter to modify its pronunciation from the same letter if left unmarked. Despite the importance of these diacritics in the Yorùbá language to mark tones and differentiate words, they continue to be omitted. This omission of diacritics in Yoruba text online leads to a difficulty in understanding especially in non-standard dialects such as Ifẹ̀. This project seeks to evaluate two multilingual transformer model (mT5-small and byT5-small) for the automatic restoration of Yoruba diacritics, focusing on the Ifẹ̀ dialect, by fine-tuning and testing the model on an existing Yoruba corpus and assessing its performance using character accuracy, and word error rate. ## 2. Technologies Used - PyTorch - Transformers - datasets - Jupyter Notebook - Google Colab (T4 GPU) ## 3. Project Structure ``` project_root/ │ ├── data/ # Not tracked │ ├── processed/ │ ├── raw/ │ ├── notebooks/ ├── 01_exploratory_data_analysis.ipynb ├── 02_preprocessing_model_setup.ipynb │ ├── checkpoints/ # Not tracked │ ├── bt5-small-v1/ │ │ ├── checkpoint-1287/ │ │ ├── checkpoint-1430/ │ ├── mt5_small-v2/ │ │ ├── checkpoint-1287/ │ │ ├── checkpoint-1430/ │ ├── mt5_small-v3/ │ │ ├── checkpoint-2860/ │ ├── reports/ │ ├── figures/ │ │ ├── exploratory_data_analysis/ │ │ ├── model_evaluation/ │ ├── src/ │ ├── config.py │ ├── data_preprocessing.py │ ├── data_splitting.py │ ├── finetuning_utils.py │ ├── model_evaluation.py │ ├── .gitignore ├── LICENSE ├── README.md ├── requirements.txt ├── ife_undiacritised_and_diacritised.csv ``` ## 4. How to run ## 📥 Dataset Setup The dataset used in this project is **not included in the repository** due to size and licensing restrictions. To reproduce the results: 1. **Download the Yorùbá text dataset** from its ori …