SA-Knowledge

This repository collects corpora and evaluation data for four South African languages: isiZulu, isiXhosa, Sepedi and Sesotho. The resources were developed for the doctoral thesis Injecting Commonsense Knowledge into Pretrained Language Models for Low Resource Languages (University of Cape Town, 2026). Each subset corresponds to a thesis chapter and can be used independently.

• Point of contact: Sello Ralethe
• Supervisor: Dr. Jan Buys, Department of Computer Science, University of Cape Town
• Languages: isiZulu (zul), isiXhosa (xho), Sepedi (nso), Sesotho (sot), English (eng)
• Code: github.com/sello-ralethe/SA-Knowledge

This card documents all six subsets in the repository: parallel_sentences, monolingual, knowledge_Bases, QuestionAnswering, multi_hall_SA and generics.

Scope of the corpora

The corpora released here are drawn from the government domain alone. They do not incorporate the SADiLaR integration described in Chapter 3. Any statistics computed over these files will therefore fall below the corpus totals reported in the thesis, which combine all four collection streams.

Orthographic considerations

isiZulu and isiXhosa use conjunctive orthography, in which morphemes are written bound to the stem. Sesotho and Sepedi use disjunctive orthography, in which the same morphemes appear as separate words. For content of equivalent meaning, Sotho-Tswana text is therefore considerably longer in words than both its English source and its Nguni equivalent, and its type-token ratio is correspondingly lower, as the additional tokens are predominantly high-frequency function words.

parallel_sentences

This subset contains sentence-aligned English to target-language bitext, with one file per language pair together with a combined archive.

The source material consists of South African government publications, namely citizen service pages and Cabinet statements published in parallel across the official languages. Because these documents are indexed by language in their URLs, we identify translation pairs through structural matching rather than content-based alignment.

Sentence alignment follows the two-pass strategy described in Chapter 3.

Each pair carries a similarity score and a confidence label. Following Table 3.2 of the thesis, pairs scoring above 0.85 are classified as high confidence, pairs between 0.65 and 0.85 as medium confidence, and pairs below 0.65 are discarded. Splits are assigned at document level, so that no document contributes sentences to more than one split.

monolingual

This subset contains cleaned monolingual text for each of the four languages, formatted as one sentence per line. The material is drawn from the target-language side of the same government collection, with duplicate source files removed by content hash before processing.

Language identification proceeds at document level before sentence level, and this ordering is deliberate. The Nguni languages share approximately seventy percent of their vocabulary, and a short sentence frequently cannot be assigned reliably to isiZulu rather than isiXhosa, while a complete document can. Applying identification per sentence and discarding every sentence assigned to a sibling language removes substantial quantities of genuine isiZulu text. We therefore verify each page as a document first, and thereafter remove only clear contaminants at sentence level, namely untranslated English blocks and out-of-family predictions. Out-of-family predictions are additionally checked against the noun class prefix and subject concord discriminators described in Section 3.5.2 before any sentence is discarded.

Cleaning includes recovery of the character encoding corruption affecting the Sesotho and Sepedi š character, removal of scraper headers and boilerplate, and the same inline-link and clause-number repairs applied to the parallel data.

knowledge_Bases

This subset contains the structured knowledge projected from English into the four target languages using LeNS-Align, the multi-level alignment approach introduced in Chapter 4. Two source knowledge bases were projected: ConceptNet, which supplies commonsense relations, and DBpedia, which supplies encyclopedic relations over named entities. The two are released separately because Chapter 8 finds that they contribute complementary rather than interchangeable knowledge.

The subset is organized by source knowledge base, and within each by representation. Under triples, the projected knowledge is given in relational form, one file per language. Under verbalized, the same knowledge is rendered as natural language sentences, again one file per language. The verbalized form is the representation consumed by the mT5+KG condition in Chapter 5, where knowledge is supplied to the model as text rather than as graph structure, and it is released alongside the triples so that both conditions can be reproduced without regenerating the verbalization.

QuestionAnswering

This subset contains translated multiple-choice evaluation data. Questions are presented in fill-in-the-blank format following Yasunaga et al. (2021), which is the format used for evaluation in Chapter 5.

commonsenseqa

These files contain translations of CommonsenseQA into Sepedi, isiXhosa, isiZulu and Sesotho. Two translation provenances are present, consistent with Section 5.6.1: manual translation for isiZulu and Sepedi, and machine translation using Google Cloud Translate for the remaining languages.

openbookqa_translated

These files contain translations of the opening portion of the OpenBookQA test split into all four languages, distributed in OpenBookQA JSONL format with identifiers, choice labels and answer keys preserved from the English source.

multi_hall_SA

This subset is the cross-lingual hallucination detection benchmark introduced in Chapter 7, released as one file per language pair.

The benchmark is built from parallel English and target-language government service documents. Facts are extracted from aligned document pairs, and each fact is then perturbed into a hallucinated variant belonging to one of four error types: temporal shift, entity error, numerical inaccuracy and location mistake. The four types are balanced exactly, each accounting for a quarter of the items in every language.

Section 7.3.3 explains why the facts are extracted in parallel rather than written in English and translated. Translating hallucinated statements into the target languages would introduce translationese artifacts that a detector could exploit, so that measured detection would partly reflect unnatural phrasing rather than the factual error itself.

Every row carries four statements: the true fact and its hallucinated counterpart, in English and in the target language. This is what allows the same file to support monolingual detection, cross-lingual detection where the evidence is in one language and the claim in another, and paired-contrast evaluation in which a model sees the true and false variants of a single fact. The fields are pair_id, file_id, fact_index, hallucination_type, english_factual_statement, english_hallucinated_statement, and the corresponding two fields named for the target language.

generics

This subset contains the generic statement probes used in Chapter 8 to measure the generic overgeneralization effect, that is, the tendency of a model to treat a property true of a distinctive minority of a kind as though it licensed universal quantification.

The probes are divided into two files. The minority-characteristic set contains generics whose property holds of a distinctive subset of the kind, such as statements about plumage patterns or predation behaviour that apply to some members and not others. The majority-characteristic set contains generics whose property holds of the kind broadly. The contrast between the two is what the probing task exploits, since a model that has represented genericity correctly should respond differently to them, and a model that has not should treat them alike.

The statements are drawn from GenericsKB and are confined to the animal domain, which supplies a large inventory of well-attested generics with clear prevalence distinctions and avoids the confound that arises when generic statements concern social groups.

Loading the data

from datasets import load_dataset

parallel = load_dataset("sello-ralethe/SA-Knowledge", "parallel_sentences")
csqa     = load_dataset("sello-ralethe/SA-Knowledge", "commonsenseqa")
obqa     = load_dataset("sello-ralethe/SA-Knowledge", "openbookqa")

# Hallucination benchmark, one split per language pair
halluc   = load_dataset("sello-ralethe/SA-Knowledge", "multi_hall_sa", split="isizulu")

# Generic statement probes, one split per prevalence class
probes   = load_dataset("sello-ralethe/SA-Knowledge", "generics")

The monolingual corpora and the knowledge bases are retrieved directly:

import pandas as pd
from huggingface_hub import hf_hub_download

def fetch(path):
    return hf_hub_download("sello-ralethe/SA-Knowledge", path, repo_type="dataset")

sentences = open(
    fetch("monolingual/sa_monolingual_zu_isiZulu.txt"), encoding="utf-8"
).read().splitlines()

# Projected triples. Header casing differs between languages, so normalize it.
triples = pd.read_csv(fetch("knowledge_Bases/conceptnet/triples/zul_conceptnet_triples_v1.csv"))
triples.columns = [c.strip().lower() for c in triples.columns]

# Verbalized knowledge, as consumed by the mT5+KG condition in Chapter 5
facts = open(
    fetch("knowledge_Bases/conceptnet/verbalized/zul_conceptnet_verbalized_v1.txt"),
    encoding="utf-8",
).read().splitlines()

Limitations

The corpora represent the government domain, and therefore the formal register, administrative vocabulary and policy discourse characteristic of it. Conversational language, social media text, and specialist legal, medical and technical writing are absent. Models trained on this material will perform better on government service language than on other registers.

The scale of these corpora remains limited. They represent a threshold for neural approaches rather than an abundance, and as noted above they exclude the SADiLaR, educational and news media streams that contribute to the totals reported in the thesis.

The knowledge bases carry a limitation of a different kind. They are projections of English resources, and their coverage is therefore the coverage of ConceptNet and DBpedia. Concepts without English equivalents, among them kinship systems, traditional governance structures, ceremonial practice, and agricultural and medicinal knowledge specific to the region, are structurally absent, and no improvement in projection accuracy will recover them. Gaps of this kind should be read as a property of the method rather than as an error in its execution.

The generic statement probes are confined to the animal domain and are released in English only. They diagnose a specific representational failure and are not a general test of commonsense reasoning, and results obtained on them should not be generalized to generics about other kinds without further evidence.

The corpora represent standard written forms. Sepedi encompasses several dialect groups differing in vocabulary and in certain grammatical constructions, while Sesotho varies between the South African and Lesotho standards, and the two standards are not applied consistently across every file released here. Dialectal variation is partially captured but is not annotated.

Ethics

The source documents are public South African government publications. No personal data concerning private individuals is included. Names appearing in the corpora belong to public officials acting in an official capacity, and the membership tables in which such names predominantly occur were removed during cleaning.

Licensing

The repository does not carry a single licence, as each subset inherits obligations from its upstream sources.

Artefact	Upstream source	Licence
Pipeline code	Original	Apache 2.0
parallel_sentences	South African government sources	CC BY 4.0, subject to source terms
monolingual	South African government sources	CC BY 4.0, subject to source terms
knowledge_Bases	ConceptNet 5, DBpedia	CC BY-SA 4.0
QuestionAnswering	CommonsenseQA, OpenBookQA	CC BY-SA 4.0
multi_hall_SA	Original, derived from government publications	CC BY 4.0, subject to source terms
generics	GenericsKB	CC BY 4.0

The projected knowledge bases and the question answering translations both derive from share-alike resources, and that obligation propagates to any work derived from them in turn. This applies to the verbalized files as well as to the triples, since verbalization is a reformatting of the projected knowledge rather than a new work.

Citation

@phdthesis{ralethe2026injecting,
  title   = {Injecting Commonsense Knowledge into Pretrained Language Models
             for Low Resource Languages},
  author  = {Ralethe, Sello},
  school  = {University of Cape Town},
  year    = {2026},
  address = {Cape Town, South Africa},
  note    = {Supervised by Dr. Jan Buys}
}

Acknowledgements

This work was made possible by the generous funding of the Hasso Plattner Institute for Digital Engineering, through the HPI Research School at the University of Cape Town. We also thank Google for their support through the Cloud Research Credits program, and the University of Cape Town's ICTS High Performance Computing team, whose computational facilities were essential to carrying out this research.