YorùbáNER-X: An Explainable Hybrid Transformer–Linguistic Framework for Robust Named Entity Recognition in Low-Resource Yorùbá Text

Named Entity Recognition (NER) is a key activity in Natural Language Processing; it facilitates information extraction, search, answering questions, building knowledge bases, and other downstream applications. Yorùbá NER is still hindered by a lack of labelled data, domain variation, orthographic inconsistency, and variations between standard and regional versions, despite recent advancements in African-language NLP. The empirical foundation for Yorùbá NER has been significantly strengthened by recent resources, especially MasakhaNER 2.0 and the multi-domain YoNER dataset. Nevertheless, recent data also indicate that transformer models suffer performance loss when transported across domains. In this paper, we propose YorùbáNER-X, an explainable hybrid framework that combines explicit Yorùbá linguistic features with transformer contextual representations and Association Rule Mining (ARM). The framework represents orthographic, character-level, lexical, contextual, and entity-indicative information alongside contextual transformer embedding; ARM is used as a supplementary evidence mechanism for recurring linguistic and contextual patterns rather than as a substitute for neural prediction; the framework is assessed using a controlled experimental design that includes standard NER evaluation, diacritic-variation tests, cross-domain evaluation, component ablation, statistical comparison, error analysis, and explainability analysis. The main baselines are XLM-R, an Africa-centric transformer, and OyoBERT. The study offers a repeatable methodological framework for examining whether pre-trained contextual representations in low-resource African-language NER can be enhanced by explicit language-specific information and statistical connections. In this publication, numerical experimental results are purposefully left out and will be added after implementation and assessment.