GC-PCWR+ for Word Sense Disambiguation

Word sense disambiguation (WSD) is a significant task in the field of natural language processing, focused on determining a word’s specific meaning within a given context. The gloss-context prompt-based contextual word representation (GC-PCWR) method, first proposed, successfully applied prompt learning to the task of word sense disambiguation, yielding promising results. However, it exhibited insufficient adaptability and generalization capabilities. For this purpose, we propose the GC-PCWR+ method that improves GC-PCWR by designing more appropriate prompt templates and better embedding the method. To leverage the capability of pre-trained models for semantic information during the training phase, the GC-PCWR+ method embeds prompt templates into the context encoding. This integration of prompt information with the context of the target word provides more comprehensive semantic insights. We established a more pragmatic dataset based on Collins and additionally performed a series of experiments on template statement complexity and information complexity. Furthermore, our method outperforms baseline models across datasets with diverse annotation systems, validating the impact of varied templates on overall performance.