Three Lingala-speaking validators judged 310 samples from the three corpora: 88 audio segments, 91 document images and 131 text examples. Each sample received three judgements, 930 in all, under a counting rule fixed before the campaign.
What was found
- Meaning is preserved in 96.5 % of cases. Substantive errors, those that change the meaning, do not exceed 3 %.
- Strict validity (“correct” by majority) ranges from 43 to 84 % depending on the corpus. The corrections requested bear mostly on spelling.
- Agreement between validators, measured by Krippendorff's alpha, stays below the 0.60 threshold set in advance. These results are therefore published as indicative.
Why publish weak agreement
Weak agreement does not mean the data is bad. It says that the boundary between “correct” and “partially correct” is not the same for everyone, largely because Lingala spelling is not settled. That is useful information for the whole community, and it is what prompted the spelling harmonisation of the corpora.
Validator responses are kept by CDS under pseudonym. The published reports contain no name.
By The Maloba team, September 29, 2026.