HSRG-WSD: A Novel Unsupervised Chinese Word Sense Disambiguation Method Based on Heterogeneous Sememe-Relation Graph

Word sense disambiguation (WSD) plays a crucial role in natural language processing. Unsupervised WSD approaches based on knowledge bases like HowNet offer improved applicability compared to supervised learning, but existing research tends to oversimplify disambiguation and neglect hierarchical sememe relationships which lead to the inability to accurately differentiate between senses with the same combination of sememes radicals. This paper pre-sent an unsupervised Chinese word sense disambiguation method based on a Heterogeneous Sememe-Relation Graph (HSRG) that leverages sememe hierarchical relationships to uncover the intrinsic connections between sememes. Additionally, we incorporate cross-word sememe relationships and semantic dependency relationships, establishing both indirect and direct contextual associations while mitigating the influence of syntactic structures. This integration enhances the representation of ambiguous words and improves disambiguation outcomes. Furthermore, our study innovatively combines the principles of graph contrastive learning with node selection algorithms, employing heterogeneous graph neural networks to effectively represent graph models and facilitate unsupervised selection of accurate sense vertices in HSRG. The proposed model is evaluated on the HowNet-based Chinese WSD dataset, demonstrating superior performance over competitive baselines.