Leveraging Semantic Diffusion for Polysemous Word Disambiguation in Morphologically Rich Low-resourced Languages

Abstract: Word Sense Disambiguation (WSD) remains one of the most challenging problems in Natural Language Processing (NLP), particularly in morphologically rich and low-resource languages. Hausa presents a unique case, where polysemy interacts with morphology to produce highly ambiguous tokens. We introduce the Hausa Polysemy Dataset (HPD), a linguistically curated sense-annotated resource, and propose the Semantic Diffusion Model (SDM), which integrates contextualized transformer encoders with graph-based semantic diffusion to jointly leverage contextual cues, gloss knowledge, and morphological relations. On HPD, SDM achieves an F1-score of 78.5%, outperforming strong baselines including GlossBERT and non-diffusive GNNs. Detailed ablations demonstrate the importance of diffusion, class-balanced focal loss, and gloss pretraining for robust performance on rare senses.