Work

Understand Kamba: Building, Evaluating, and Improving Machine Translation for Kikamba, an Underrepresented Kenyan Language. (Adapting NLLB-200 to a Low-Resource Bantu Language))

Year: 2026
Type: preprint
Abstract: Understand Kamba investigates the development, evaluation, and improvement of machine translation for Kikamba, an underrepresented Kenyan language. Authored by Lawrence Maluki, the paper examines multilingual translation models, adaptation with limited parallel data, and the effects of orthographic variation on evaluation. The study documents low-rank adaptation (LoRA) of NLLB-200 distilled 1.3B using approximately 10,200 sentence pairs. Historical matched evaluations show Kikamba-to-English chrF++ improvements from 34.09 to 44.33 on FLORES-200 devtest and from 25.52 to 37.09 on held-out children’s stories, with smaller gains in the reverse direction. Supporting experiments examine speech recognition and speech synthesis for an integrated language-access application. The paper also presents an artifact audit, additional evaluation of cached int8 translations, document-cluster bootstrap confidence intervals, and sensitivity analysis for potential training–test overlap. Graphs, methodological documentation, and reproducibility guidance support inspection and further research. This research preprint distinguishes completed experiments from planned validation. Historical fine-tuning predictions and some provenance records remain unavailable. A blinded human translation-evaluation pilot is prepared, but ratings have not yet been collected. A separate voice-contribution collection is documented as supporting evidence. The work contributes a documented adaptation methodology, empirical findings, and an evaluation framework for further research on Kikamba and other underrepresented languages.
Language: English
Open Access status: green
OpenAlex ID: W7226644359
Cites:
Cited by:
Related to: 10