Small language models efficacy prototyped for Oromo word sense disambiguation

Word Sense Disambiguation (WSD) is a fundamental issue in Natural Language Processing (NLP), particularly for low-resource languages with minimal annotated content. The substantial inflection, homonymy, and polysemy of Oromo, a morphologically rich Cushitic language spoken in Ethiopia and the surrounding regions, complicate computational semantic analysis. This study evaluates the effectiveness of lightweight Transformer-based models for Oromo WSD by comparing them with traditional TFIDF-based machine learning and rule-based approaches. Verbs, nouns, adjectives, adverbs, and combined AllWords are the five parts-of-speech categories covered by the 1,919 annotated examples in 154 senses that make up our WSD dataset. Rare classes with less than eight examples were replicated in order to lessen low-frequency affects. The studies employed stratified 5-fold cross-validation, and Accuracy, Macro Precision, Macro Recall, and Macro F1 were used to assess the models. The results show that transformer-based models, including BERT, DistilBERT, XLM-RoBERTa, and multilingual BERT, consistently beat TFIDF and rule-based baselines in every area. According to confusion matrix research, rule-based and TFIDF models are biased toward frequent senses, whereas Transformers produce more balanced, context-aware predictions that effectively distinguish semantically comparable senses. Statistical significance testing using paired t-tests confirms the advantage of transformers (p < 0.05). DistilBERT provides the optimum trade-off between efficiency and performance, whereas XLM-RoBERTa excels with combined POS categorization. These findings demonstrate that by allowing semantic processing in low-resource languages, lightweight multilingual Transformer topologies offer a consistent, efficient, and context-sensitive approach for WSD tasks.