In this work we introduce a collection of monolingual embedding models for ten South African languages in four different architectures. To determine the quality of the embedding models we evaluate the embeddings on two sequence-labelling tasks, namely Part-of-Speech (POS) tagging and Named Entity Recognition (NER). Languages are grouped into conjunctive (isiNdebele, isiXhosa, isiZulu, and Siswati), disjunctive (Sepedi, Sesotho, Setswana, Tshivenḓa, and Xitsonga), and Afrikaans to establish the influence of training data set size and typology on the quality of the different embeddings. To isolate representation effects we train BiLSTM-CRF taggers, while keeping the architecture, data splits, and training budget fixed, varying only the input imbedding representations, namely GloVe, fastText, Flair, and RoBERTa. In our experiments, GloVe lags behind fastText, Flair, and the transformer-based models, confirming that static word-level vectors are less suited to morphologically complex, low-resource languages. Subword-aware embeddings such as fastText remain a reliable and computationally efficient baseline, while Flair is the most competitive overall across both POS tagging and NER tasks.

Cross-lingual transfer learning enables building NLP systems for low-resource languages by leveraging data from higher-resource languages. A critical but understudied question for African languages is: which source languages should be selected for multi-source transfer? We present a systematic comparison of four source language selection strategies: random selection (baseline), genetic distance based on language family trees, geographic distance based on speaker locations, and embedding similarity from multilingual models. We evaluate these strategies on Named Entity Recognition, Part-of-Speech tagging, and sentiment analysis across five typologically diverse African target languages (Hausa, Yoruba, Swahili, Igbo, Kinyarwanda) using three multilingual models. We further investigate how the number of source languages affects transfer performance. Our experiments reveal that no single strategy dominates across tasks: geographic distance leads on sequence labeling tasks while embedding similarity is most effective for sentiment analysis, and all informed strategies consistently outperform random selection.

Large language models (LLMs) are trained on data contributed by low-resource language communities, including curated datasets such as MasakhaNER and MAFAND-MT, yet the linguistic knowledge encoded in these models remains accessible only through commercial APIs. This paper investigates whether strategic prompting can extract usable text data from LLMs for two West African languages: Hausa (Afroasiatic, approximately 80 million speakers) and Fongbe (Niger-Congo, approximately 2 million speakers). We systematically compare six elicitation task types: creative writing, functional text, structured knowledge, dialogue, topic-switching probes, and constrained generation across two commercial LLMs (GPT-4o Mini and Gemini 2.5 Flash). Generated outputs are evaluated on linguistic accuracy, lexical diversity, domain coverage, and code-switching rates through automatic metrics assessment. Our findings reveal that elicitation strategy significantly affects output quality and that optimal strategies differ by language: Hausa benefits from volume-maximizing tasks such as functional text and dialogue, while Fongbe requires constraint-heavy prompts that enforce monolingual output. GPT-4o Mini extracts 6–41x more usable target-language words per API call than Gemini, though Gemini achieves higher language purity for Fongbe on constrained tasks. We provide a practical framework for low-resource language communities to maximize usable data extraction from LLMs and release all generated corpora and code.

In this paper, we present part-of-speech (POS) annotated domain-specific data for nine South African languages. The data has been sourced from five different domains (two academic domains, Caps and theses, two non-academic domains, news and magazines, and one fiction domain, novels), uniformly pre-processed, automatically POS-tagged and then corrected by linguistic experts. The widely used NCHLT government data sets (Eiselen and Puttkammer, 2014) have also been re-tagged with the current tag sets and manually corrected. Both the new domain-specific data sets and the re-tagged NCHL data sets have been uploaded into a public repository. To illustrate the characteristics of the domain data in comparison to government data, we include and discuss data statistics, namely type-token ration (TTR), tokens per sentence and out-of-vocabulary (OOV) rates, as well as POS tagging results with a baseline tagger trained on NCHLT data and applied to the different domains for all languages. Both the data statistics and the POS results clearly show that the domain data is significantly different to government data: For all domains and languages, the tagging accuracy decreases significantly compared to testing on in-domain government data. Also, POS results for the two domains with the highest OOV rates for all languages (Caps and novels) are much lower than for the other domains. These findings emphasise the need for more diverse data resources which in turn will aid in the development of more domain-independent language technologies.

This paper presents the first publicly available UD treebank for Tswana, Tswana-Popapolelo. The data used consists of the 20 Cairo CICLing sentences translated to Tswana. After pre-processing these sentences with detailed POS (XPOS) and converting them to universal POS (UPOS), we proceeded to annotate the data with dependency relations, documenting decisions for the language specific constructions. Linguistic issues encountered are described in detail as this is the first application of the UD framework to produce a dependency treebank for the Bantu language family in general and for Tswana specifically.

In this paper we present a case study for three under-resourced linguistically distinct South African languages (Afrikaans, isiZulu, and Sesotho sa Leboa) to investigate the influence of data size and linguistic nature of a language on the performance of different embedding types. Our experimental setup consists of training embeddings on increasing amounts of data and then evaluating the impact of data size for the downstream task of part of speech tagging. We find that relatively little data can produce useful representations for this specific task for all three languages. Our analysis also shows that the influence of linguistic and orthographic differences between languages should not be underestimated: morphologically complex, conjunctively written languages (isiZulu in our case) need substantially more data to achieve good results, while disjunctively written languages require substantially less data. This is not only the case with regard to the data for training the embedding model, but also annotated training material for the task at hand. It is therefore imperative to know the characteristics of the language you are working on to make linguistically informed choices about the amount of data and the type of embeddings to use.

In this paper, we present a project where existing text-based core technologies were ported to Java-based web services from various architectures. These technologies were developed over a period of eight years through various government funded projects for 10 resource-scarce languages spoken in South Africa. We describe the API and a simple web front-end capable of completing various predefined tasks.

This paper describes the named entity language resources developed as part of a development project for the South African languages. The development efforts focused on creating protocols and annotated data sets with at least 15,000 annotated named entity tokens for ten of the official South African languages. The description of the protocols and annotated data sets provide an overview of the problems encountered during the annotation of the data sets. Based on these annotated data sets, CRF named entity recognition systems are developed that leverage existing linguistic resources. The newly created named entity recognisers are evaluated, with F-scores of between 0.64 and 0.77, and error analysis is performed to identify possible avenues for improving the quality of the systems.

Phrase chunking remains an important natural language processing (NLP) technique for intermediate syntactic processing. This paper describes the development of protocols, annotated phrase chunking data sets and automatic phrase chunkers for ten South African languages. Various problems with adapting the existing annotation protocols of English are discussed as well as an overview of the annotated data sets. Based on the annotated sets, CRF-based phrase chunkers are created and tested with a combination of different features, including part of speech tags and character n-grams. The results of the phrase chunking evaluation show that disjunctively written languages can achieve notably better results for phrase chunking with a limited data set than conjunctive languages, but that the addition of character n-grams improve the results for conjunctive languages.

In most languages, new words can be created through the process of compounding, which combines two or more words into a new lexical unit. Whereas in languages such as English the components that make up a compound are separated by a space, in languages such as Finnish, German, Afrikaans and Dutch these components are concatenated into one word. Compounding is very productive and leads to practical problems in developing machine translators and spelling checkers, as newly formed compounds cannot be found in existing lexicons. The Automatic Compound Processing (AuCoPro) project deals with the analysis of compounds in two closely-related languages, Afrikaans and Dutch. In this paper, we present the development and evaluation of two datasets, one for each language, that contain compound words with annotated compound boundaries. Such datasets can be used to train classifiers to identify the compound components in novel compounds. We describe the process of annotation and provide an overview of the annotation guidelines as well as global properties of the datasets. The inter-rater agreements between the annotators are considered highly reliable. Furthermore, we show the usability of these datasets by building an initial automatic compound boundary detection system, which assigns compound boundaries with approximately 90% accuracy.

The development of linguistic resources for use in natural language processing is of utmost importance for the continued growth of research and development in the field, especially for resource-scarce languages. In this paper we describe the process and challenges of simultaneously developing multiple linguistic resources for ten of the official languages of South Africa. The project focussed on establishing a set of foundational resources that can foster further development of both resources and technologies for the NLP industry in South Africa. The development efforts during the project included creating monolingual unannotated corpora, of which a subset of the corpora for each language was annotated on token, orthographic, morphological and morphosyntactic layers. The annotated subsets includes both development and test sets and were used in the creation of five core-technologies, viz. a tokeniser, sentenciser, lemmatiser, part of speech tagger and morphological decomposer for each language. We report on the quality of these tools for each language and discuss the importance of the resources within the South African context.