Accurate and Efficient Statistical Testing for Word Semantic Breadth Using Contextualized Embeddings

arXiv·low signal

Introduces an efficient statistical test for measuring how broadly a word's meaning spreads across contexts, leveraging contextualized token embeddings from language models. Replaces expensive permutation tests with a closed-form approximation that runs orders of magnitude faster. Primarily useful for NLP researchers studying polysemy and lexical semantics at scale.