A graph-based analysis of semantic types and coercion in contextualized word embeddings
Researchers have developed a novel graph-based approach to analyze how semantic type information is represented within contextualized word embeddings. This method uses metrics like Neighbor Type Probability (NTP) and Neighbor Type Entropy (NTE) to examine the distribution of semantic types in the embeddings' neighborhoods. The study found that sense-enhanced embeddings better capture lexical and contextual type information, enabling the distinction between sentences with matching and mismatching semantic types. AI
IMPACT Introduces a new analytical framework for understanding the nuances of word embeddings, potentially improving downstream NLP tasks.