SA-Monolingual-Corpora

Cleaned monolingual corpora for isiZulu, isiXhosa, Sesotho and Sepedi, drawn from South African government publications.

Produced for the doctoral thesis Injecting Commonsense Knowledge into Pretrained Language Models for Low Resource Languages (University of Cape Town, 2026). Code at https://github.com/sello-ralethe/SA-knowledge

Structure

One configuration per language, each with a single train split. Files are plain text, one sentence per line, UTF-8, deduplicated.

Loading

from datasets import load_dataset

zulu = load_dataset("sello-ralethe/SA-Monolingual-Corpora", "isizulu")
for row in zulu["train"]:
    print(row["text"])

Method

Encoding recovery for the Sotho-Tswana s-caron corruption, boilerplate removal, repair of words fused across inline link boundaries, and rule-based sentence segmentation.

Language identification runs at document level before sentence level. The Nguni languages share around 70% of their vocabulary, so a short sentence often cannot be assigned reliably while a whole page can. Sentences whose top label is a sibling of the verified document language are kept; only English blocks and out-of-family predictions are removed.

Licence

CC BY 4.0, subject to the terms of the originating government departments.