Enhancing Zero-Shot Word Sense Disambiguation via Clustered Sense Vectors

Recent approaches to Word Sense Disambiguation (WSD) often rely on the use of either clustered sense labels or sense vectors. The former reduces the output label space and helps handle untrained senses while maintaining fine-grained sense distinctions. The latter leverages latent representations to capture contextual information. In this study, we propose Clustered Sense Vectors, a method that integrates both strategies to enhance fine-grained WSD performance. In order to create clustered sense vectors, we convert clustered sense labels into clustered sense vectors. For this, we propose Sense-aware Clustering (SaC), a method for clustering sense labels more effectively. Then, for a WSD system, we present CluBE, a bi-encoder model that utilizes these clustered sense vectors by converting clustered sense labels into representative vectors through cluster vector normalization. The experimental results show that SaC improves clustering quality, and that CluBE not only achieves 1.6%p higher F1 scores but also reduces memory usage by 82% compared to a baseline bi-encoder model, while significantly enhancing zero-shot learning capability by effectively handling untrained word senses through clustered representations.