Ojo, Sunday O.Masethe, Mosima Anna2026-06-292026-06-292026-03-31https://hdl.handle.net/10321/6425Submitted in fulfilment of the requirements of the Degree of Doctor of Philosophy in Information Technology, Durban University of Technology, Durban, South Africa, 2026.Meaning conflation deficiency (MCD) is a major issue in natural language processing (NLP) to improve low-resource languages with complex morphologies. This research study tackles the MCD problem in word vector space (WVS) word embedding (WE) models, that struggle with polysemous words. The study explored optimization strategies and graph-based architectures to resolve sense ambiguities in resource-limited environments. A computational experimental design (CED) assessed traditional neural models (BiLSTM, NNLM, CNN), transformer models (BERT, RoBERTa, ELECTRA, BART, XLNet, ELMo), and graph-based approaches (domain-specific, integrated and hierarchical). Models were tested on a sense-annotated SsaL corpus with polysemous terms, focusing on ambiguous words. Experimental results showed significant performance gains through optimization. Optimised BART improved accuracy to 96% from 87%, and XLNet increased precision to 92% from 83%. CNN models with dropout regularization saw the largest F1-score improvement, rising from 69% to 98%, highlighting regularization’s importance for semantic sparsity. McNemar’s test (p < 0.05) confirmed the improvements as significant across all models. The model was optimised using morphology-aware preprocessing, context-sensitive attention, SMOTE to address class imbalance, and tailored dropout regularisation to reduce overfitting. Graph-based approaches excelled, with the domain-specific graph achieving near-perfect scores (98% precision, recall, F1-score, accuracy) by integrating linguistic knowledge. The integrated graph maintained a strong performance (88% across metrics), utilizing sentence similarity, domain indicators, and clustering. The hierarchical graph, with moderate results (83-86%), highlighted the importance of graph granularity in disambiguation. ROC curves (AUC 0.90-1.00) and confusion matrices validated the effectiveness of graph methods in capturing disambiguation cues, outperforming distributed representations with limited data. The analysis showed performance differences among model families. BERT was the strongest baseline (87% accuracy, 60% F1-score), followed by Multilingual BERT (89% accuracy, 55% F1). Multilingual models were inconsistent—Multilingual RoBERTa only had 53% accuracy and 17% F1-score—indicating poor representation of Bantu morphological structures. Models performed well on frequent senses such as “location” but struggled with infrequent senses such as “time”, highlighting issues with class imbalance and data scarcity. Ablation studies highlighted key architectural insights for low-resource settings. The encoder-only BART configuration achieved approximately 97% of the full model’s overall classification accuracy, showing that reduced architectural complexity resulted in only a minor loss in predictive performance, suggesting that computationally efficient implementations can preserve predictive effectiveness while reducing model complexity. XLNet’s permutation-based attention effectively handled long-distance dependencies in agglutinative languages. Techniques such as morphology-sensitive tokenization and focused finetuning minimized misclassification errors and enhanced the detection of infrequent word senses across all model categories. This research contributes significantly to computational linguistics and lowresource NLP by showing that MCD requires multi-faceted optimization beyond architectural choices. The domain-specific graph framework bridges data-driven neural methods with linguistic encoding, achieving state-of-the-art performance via domain knowledge and semantic representation. The study proves that transformer models can adapt to low-resource contexts with careful fine-tuning, highlighting limits in current multilingual models for morphologically complex African languages. Findings impact low resourced language technology beyond SsaL, offering a blueprint for word sense disambiguation in morphologically rich languages with limited data. Semantic processing is feasible with constrained resources, promoting linguistic equity in NLP. Limitations include dataset scale, computational resource constraints for hyperparameter tuning, and domain-specific graph construction needing manual lexicon development. Future research should broaden sense inventories, explore crosslingual transfer learning from Bantu languages, develop few-shot learning for new senses, and enhance disambiguation in applications such as machine translation. Creating morphological tokenization for Bantu languages and multilingual models trained on African languages is crucial for advancing computational semantics low resourced communities.304 penWord Sense DisambiguationMorphologically Rich LanguagesWord sense disambiguationSesotho languageHUMANITIES and RELIGION::Languages and linguistics::Other languages::Bantu languagesNatural language processing (Computer science)Language and languagesMachine learningComputational linguisticsText processing (Computer science)Thesishttps://doi.org/10.51415/10321/6425