Bridging the Resource Gap: Fine-tuning a Machine Translation System for English-Swahili Medical Translation in Kenya's Low-Resource Context

In Kenya, the predominance of English in healthcare contexts creates a significant barrier for Swahili speakers, compromising access to information for most Kenyan citizens. This thesis aims to research ways to address this gap by fine-tuning a specialized English-Swahili machine translation system for the medical domain. The study involved the creation of the AFYA corpus, an original parallel dataset of approximately 15 000 segments, used to fine-tune an NLLB machine translation model. The performance was evaluated against Google Translate through a hybrid framework that combined automated metrics (COMET, SacreBLEU and chrF2) and a human evaluation task based on MQM metrics. Results revealed that while Google Translate offers higher Fluency, NLLB achieves superior semantic alignment and statistical parity in clinical safety (Verity). Crucially, human analysis revealed that the performance of Google Translate was inconsistent and resulted in a negative Inter-Annotator Agreement, often masking terminological errors. In contrast, NLLB’s errors were more obvious and transparent, thus easier to spot and correct through post-editing. This study argues for the need to preserve linguistic diversity in specialized fields, providing a framework for more accessible healthcare-related information in Kenya.