Reproduce “Solving Word-Sense Disambiguation and Word-Sense Induction with Dictionary Examples” on OpenResearch

About this paper

Tadej Škvorc, Marko Robnik-Šikonja

This research introduces a methodology to overcome data scarcity for Word-Sense Disambiguation (WSD) and Word-Sense Induction (WSI) in less-resourced languages. It leverages generative large language models (LLMs) to expand dictionary examples into full sentences, achieving up to 94.2% accuracy for WSD in Slovene.

Key insights

• Generative LLMs can effectively transform short dictionary snippets into high-quality, sense-preserving, and grammatically correct full sentences (98.5% combined accuracy) for a less-resourced language.
• Training a WiC classifier on LLM-augmented dictionary data is a viable approach for performing WSD and WSI, particularly in data-scarce environments where sense inventories are limited.
• The methodology demonstrates robust generalization, improving performance even for entirely unseen words (Pure-OOV scenario) in the WiC task.

Summary by alphaXiv.