Sethosa, MRRananga, SMbooi, Mahlatse S2025-10-312025-10-312025-05978-1-905824-75-5DOI: 10.23919/IST-Africa67297.2025.11060561http://hdl.handle.net/10204/14453This research will fill the gap in the availability of hate speech detection models that are inclusive of low-resource languages with a focus on isiZulu. Recent Natural Language Processing studies have concentrated more on high-resource languages, and as a result, languages like isiZulu have been under-represented in this field of research. To bridge this gap, an annotated English dataset has been utilized, leveraging Google translation API to translate the data into isiZulu. We made use of semantic analysis to look for patterns within the labeled categories of data. Without using the categories as target variables in the training process, However, this classification by domain allowed this study to delve into offensive terms with respect to prevalence and context. Machine learning models, such as Support Vector Machines and Random Forests, were trained using TF-IDF (Term Frequency – Inverse Document Frequency) vectorization to achieve state-of-the-art accuracy improvements in Zulu hate speech detection. These results show that a targeted semantically enriched approach that we used in this study enhances the precision of the model, which holds great potential for fine-tuned hate speech detection across multilingual contexts. This research contributes to expanding effective NLP tools for low-resource languages, promoting safer and more inclusive digital spaces.FulltextenisiZuluLow-resource languagesHate speech detectionSemantic analysisNatural language processingNLPGoogle translationSocial media moderationSupport Vector MachineSVMRandom ForestRFConference PresentationN/A