This paper presents a manually annotated morpho-syntactic corpus of Ògè, an under-resourced indigenous language spoken in Nigeria. The corpus consists of ten folk narratives (approximately 4,667 tokens) collected for the investigation of nominal structure. Annotation is expert-driven and includes token-level part-of-speech tagging together with a structured Determiner Phrase (DP) classification framework designed to capture language-specific nominal configurations. The scheme distinguishes between bare nouns and modified noun phrases, reflecting a central structural property of Ògè: noun forms remain morphologically stable across contexts, while modifiers exhibit formal and positional variation contributing to reference, specificity, and discourse prominence. The DP classification layer encodes both simple and complex nominal constructions, enabling systematic analysis of internal phrase structure. Designed as a reusable digital resource, the corpus supports morphosyntactic tagging, noun phrase boundary detection, and modeling of nominal structure in low-resource NLP settings. The annotated dataset will be made publicly available through the SADiLaR repository. This work demonstrates how descriptive linguistic analysis can inform annotation design and provides a replicable framework for developing structured resources for under-resourced African languages. Keywords: Ògè, low-resource NLP, annotated corpus, nominal structure, African languages

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.

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.

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.

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.

Retrieval-augmented generation (RAG) has been used to improve the accuracy and transparency of outputs produced by large language models (LLMs) by integrating external knowledge; however, applying RAG to low-resource languages presents unique challenges, including poor embedding representations, low retrieval quality, and semantic gaps caused by the scarcity of digital documents. In this research, we address these challenges for a selected low-resource language, Amharic, by using translative and debate-based RAG techniques to improve retrieval and reasoning. This paper outlines the key problems and research gaps in applying RAG to low-resource languages and introduces a method to enhance RAG performance for Amharic. Additionally, we introduce the first comprehensive Amharic Retrieval-Augmented Generation Benchmark (ARGB), designed to capture grammatical, cultural, and writing-system-specific constraints of the Amharic language. ARGB evaluates not only retrieval and generation quality, but also noise robustness, counterfactual robustness, negative rejection, and multi-source information integration, providing a holistic assessment of RAG capabilities. The dataset, which spans a wide range of categories, is evaluated using multiple evaluation metrics. Furthermore, we demonstrate that, using our dataset, translation-based and debate-based methods substantially improve various aspects of RAG pipeline assessment in the Amharic language. This work aims to improve the reliability, accessibility, and inclusiveness of AI systems for Amharic speakers while providing a scalable framework for other low-resource languages. Current progress on the code and benchmark can be found on this GitHub link: link.

Automatic Speech Recognition (ASR) for low-resource languages remains challenging due to limited annotated data and high linguistic variability. In this work, we investigate parameter-efficient fine-tuning strategies for Cape Verdean Creole ASR using the Wav2Vec 2.0 XLSR model. We evaluate the impact of structured layer freezing on model performance, training stability, and computational efficiency. Experiments conducted on a newly curated Santiago-dialect dataset show that full fine-tuning achieves the best absolute performance (WER 0.212, CER 0.120). However, several freezing configurations achieve comparable recognition performance while substantially reducing the number of trainable parameters and exhibiting more stable convergence. These results highlight a trade-off between adaptability and efficiency, showing that selective freezing can serve as an effective regularization strategy in low-resource settings. This work provides practical insights into parameter-efficient adaptation for under-resourced Creole languages.

Large language models (LLMs) are increasingly evaluated in multilingual settings, yet their inference behavior in low-resource African languages remains underexplored especially under pure prompting without fine-tuning. We present a systematic study of prompting strategies for Natural Language Inference (NLI) in Swahili, Yoruba, and Hausa using the AfriXNLI benchmark. We evaluate five prompting strategies Baseline (zero-shot), Script-Aware, Language Specific, Contrastive, and Native-Label Self-Translation (NL-STP) across two mid-sized open weight models (Llama3.2-3B and Gemma3-4B). To isolate the effect of prompt design, the effect of few-shot examples and Chain-of-Thought reasoning is eliminated in our study. We find a significant difference in performance of class wise across strategies with highly neutral class collapse and high prediction skew in some configurations. Contrastive prompting proves to be the most reliable and steadily improving strategy over language and model and has better balance of class behavior and balance of overall accuracy gains. Notably, well-constructed prompts are sufficient to beat more powerful baselines that are provided with few-shot prompts and Chain-of-Thought prompts. We have found that prompt formulation is essential to multilingual NLI with low-resource languages and that language aware decision structuring can be used to meaningfully enhance robustness in resource challenged settings.

Research efforts aimed at detecting unsafe dialogues have resulted in creation of benchmark datasets and models for evaluation. The benchmarks mostly exist in English and other high resourced languages. In order to address the challenge of unavailability of dialogue safety evaluation dataset in Hausa and Yorùbá, we repurporse DiaSafety dataset to develop HaYo dataset, by providing contextualised human annotation of dialogues in DiaSafety. We provide dialogues in Hausa and Yorùbá, obtained by human translation of dialogues in the DiaSafety dataset, to raters who are native speakers. The dialogues are annotated as Unsafe or Safe. We evaluate seven models with moderation, conversational or multilingual capabilities in terms of F1 Score. Using McNemar test, we observe that the predictions of GPT-4.1 and Gemma-3-12b-it on HaYo are statistically significant at p < 0.05. In our evaluation with instructions in English, we observe lower F1 scores in six out of the seven models, comparing the performance on DiaSafety and HaYo labels. The model predictions were inconsistent with the labels in the HaYo dataset when instructions and dialogues were provided in Hausa and Yorùbá. Compared to providing instructions in English, the issues range from responses in unspecified languages to underperformance in terms of F1 score. We plan to release the HaYo dataset to the public to promote dialogue safety research, especially in under-resourced languages.

Multilingual NLP has expanded rapidly through large-scale pretraining and cross-lingual transfer, yet this progress remains structurally uneven across writing systems. This survey reframes multilingual NLP around scripts rather than languages, arguing that writing systems constitute an under-theorized axis of computational inequality. Focusing on African scripts — Indigenous (Vai, Ge’ez, Tifinagh), modern (ADLaM, N’Ko), and adapted Arabic-based (Ajami)—we analyze how script properties interact with digital infrastructure, tokenization, and downstream task performance. We organize the literature across four analytical layers: infrastructural (Unicode and input systems), representational (segmentation efficiency and vocabulary allocation), functional (task-level disparities), and epistemic (evaluation bias and the “low-resource” framing). Synthesizing evidence from 47 studies, we show that performance gaps across scripts arise primarily from engineering design choices rather than intrinsic linguistic complexity. We conclude by outlining a research agenda for native multiscript foundation models, including script-aware scaling laws, tokenizer equity metrics, and evaluation reform. We argue that multiscript equity is not a peripheral concern but a structural precondition for genuine multilingual inclusion

Models that can estimate the readability of a given text automatically are a valuable resource for any language. There are however many languages for which such models do not work well or simply do not exist yet. In this paper, we lay the groundwork for developing a high-quality application for Afrikaans by having encoder-only language models (LMs) complete a set of cloze tests already completed by humans. Strong correlation between the cloze-test performance of humans and an LM is an indication that the LM could possibly serve as a proxy for human participants. We show that the output of models trained on (some) Afrikaans correlates reasonably well with human answers, underscoring the potential of LMs to be used in automatic readability assessment. A more fine-grained analysis confirms that the correlation is not driven by only a few strongly correlating word classes, but spread relatively evenly over all word classes. We further establish by means of a manual evaluation that, in cases where the cloze-test performance of humans and an LM correlate strongly because both were wrong, LM answers tend to be further off than human answers for the same cloze items. It is noteworthy that the model with the best correlation, afRoBERTa (r=0.62; Spearman’s ρ=0.62), is neither the most accurate nor the largest model, but a model trained on Afrikaans only, showing the benefit of small, monolingual LMs compared to large, multilingual models for specific purposes.

The development of Natural Language Processing (NLP) resources for African indigenous languages remains constrained by limited data availability, fragmented expertise, and a lack of sustainable, locally grounded infrastructures for enabling language research. While much existing work focuses on producing discrete resources such as corpora or lexicons, less attention has been paid to the social, institutional, and methodological conditions that enable such resources to be created, maintained, and sustained. This paper presents the Hundzula Retreat for NLP and Linguistics as a retreat-based resource infrastructure model that addresses these constraints. We conceptualise Hundzula not as a once-off event, but as a structured, upstream research infrastructure that facilitates human capacity development, interdisciplinary collaboration between linguistics and NLP, ethical data practices, and the early-stage incubation of language resources for African indigenous languages. Drawing on evidence from multiple iterations of the retreat, we describe the design principles, workflows, and governance mechanisms that support resource development, including training pipelines, human-in-the-loop methodologies, and collaborative project formation. Rather than focusing on already formalised outputs, the paper foregrounds the infrastructural conditions that make such outputs possible within under-resourced contexts. In doing so, the paper shifts attention from outputs to the enabling ecosystems required for their production. We argue that retreat-based infrastructures constitute an essential but under-recognised category of language resources and demonstrate how the Hundzula model can be adapted and replicated in other low-resourced language contexts. The paper contributes a transferable framework for sustainable NLP resource development grounded in African linguistic realities.

Creative Commons (CC) licenses are prevalent in African natural language processing (NLP) corpus releases, but their compatibility implications are rarely examined systematically. CC-BY-SA and CC-BY-NC cannot be combined in a single published dataset; a NoDerivs (ND) clause prohibits redistribution of tokenised or annotated derivatives. This paper presents an empirical audit of license provenance across more than twenty corpus families used in African NLP, applying established compatibility rules to three case-study languages: Kituba/Munukutuba, Zarma, and Moore. Four failure modes are documented with primary-source evidence: outright prohibition (JW300, removed from OPUS after a legal audit confirmed a Terms of Service violation); composite license misrepresentation (WAXAL, whose CC-BY 4.0 claim is contradicted by its HuggingFace dataset card); a ND restriction not reflected in the CC-BY label (Tanzil); and data persistence failure (the Congolese Radio Corpus, where 402 of 405 source URLs are no longer accessible). A due diligence checklist and a survey of legally compliant enrichment opportunities conclude the paper. We argue that lawful data use is an ethical baseline: for African language communities with limited institutional recourse, license violations are not only legal risks but ethical failures that compound existing power asymmetries.