Advancing Wolof-French Sentence Translation: Comparative Analysis of Transformer-Based Models and Methodological Insights

Abstract

This study investigates a deep neural network (DNN) translation model for the French-Wolof language pair using the Transformer architecture, specifically the Text-To-Text Transfer Transformer (T5) and the Bidirectional and Auto-Regressive Transformer (BART) models. A Long Short-Term Memory (LSTM)-based model is included for state-of-the-art comparison. Trained on 19,571 preprocessed parallel sentences from novels and the web, the models’ performance is enhanced through Bayesian hyperparameter optimization, refined tokenization, and strategic data augmentation. We introduce a novel sampling method that combines sentence length bucketing and truncation to reflect sentence length distribution during fine-tuning. The smaller $\mathbf{T 5}$ model achieves a Bilingual Evaluation Understudy (BLEU) score of 5.03 and a Recall-Oriented Understudy for Gisting Evaluation with the Longest Common Subsequence (ROUGE-L) score of 20.37 for French to Wolof translation, and 4.82 and 22.16, respectively, for the reverse. However, there remains potential for improvement with more complex architectures and domain-specific sentences tailored to Wolof grammar. Future research should include more references for metric calculation and comparisons with the latest T5 and BART versions. Despite these limitations, our findings highlight the potential of transfer learning in enhancing translation quality for low-resource languages like Wolof.