On the ability to bootstrap spellcheckers for Nguni languages via reusable model pipelines

Mahlaza, Zola and Keet, C. Maria and Magwenzi, Tadiwa (2026) On the ability to bootstrap spellcheckers for Nguni languages via reusable model pipelines, Proceedings of Annual Research Conference of South African Institute of Computer Scientists and Information Technologists (SAICSIT 2026), 13-15 July 2026, Cape Town, South Africa.

Text SAICSIT_2552.pdf - Accepted Version Download (582kB)

Abstract

Automated spellchecking is one of the core human language technologies. It is investigated intermittently for African languages, and then mostly for the most resourced among the low-resourced languages. Reusing spellchecker infrastructure to bootstrap spellcheckers for very low-resourced language remains an elusive goal. The aim of this work is to systematically investigate reusability of a pipeline to develop spellcheckers for all Nguni languages, and whether reusing a language model in one language for spellchecking in another may be feasible as an alternative bootstrapping option. We used the KenLM language model toolkit to train new models with texts in the four Nguni languages using n-grams (1 to 5), refined with a dynamic threshold. We then created a baseline with automatically wrecked words on their own language model, and assessed precision, recall, F1, and accuracy cross-lingually. Error detection accuracies are 86% for isiZulu, 83% for isiXhosa, 87% for isiNdebele, and 82% for siSwati. Average F1 score degradation is sometimes small (e.g., isiZulu text on isiNdebele model, isiXhosa on isiZulu), and sometimes substantial (siSwati on all three others). We therefore conclude that the approach may thus be feasible.

Item Type:	Conference paper
Subjects:	Computing methodologies > Artificial intelligence > Natural language processing
Date Deposited:	20 Oct 2026 07:53
Last Modified:	20 Oct 2026 07:53
URI:	https://pubs.cs.uct.ac.za/id/eprint/1826

Actions (login required)

View Item