Abstract:
This study developed a Neural Machine Translation (NMT) model for
Dholuo-Swahili translation, addressing the lack of translation resources for lowresource
languages in Kenya. Using the OpenNMT-py framework, the model was
trained on a parallel dataset of 27,415 sentence pairs from two datasets to improve
model performance. Performance was evaluated using BLEU (Bilingual Evaluation
Understudy), METEOR (Metric for Evaluation of Translation with Explicit
Ordering), perplexity (PPL), and prediction scores, with results showing a BLEU
score of 16.29, METEOR score of 0.27, and PPL of 1.88, indicating moderate
translation accuracy. Despite achieving low PPL, fluency and lexical accuracy
challenges persisted due to limited training data and linguistic differences. Future
research will focus on expanding datasets, leveraging multilingual models like
LLaMA 3.2, and integrating hybrid approaches to improve translation quality. This
work contributes to digital inclusion, language preservation, and enhanced crosslinguistic
communication for underrepresented languages in Kenya.
