Polysemysimilarity

Polysemysimilarity is a proposed metric in linguistics and natural language processing that quantifies how similar the different senses of a single polysemous word are to one another. It focuses on the internal relatedness of a lemma’s senses rather than the similarity between different lemmas. The concept is used to analyze lexical complexity, inform sense disambiguation, and support semantic search and lexicography.

Approaches to measuring polysemysimilarity fall into three broad families: knowledge-based, distributional, and context-driven. Knowledge-based methods use lexical resources such as WordNet, FrameNet, or domain ontologies to compare senses by shared hypernyms, synonyms, or semantic paths. Distributional methods embed sense representations in vector spaces; similarity is computed with cosine or other metrics between sense embeddings. Context-driven methods estimate sense similarity from usage patterns across corpora, by examining selectional preferences and contextual co-occurrence profiles.

Data sources include sense-annotated corpora and lexical databases; applications include disambiguation, machine translation alignment, information retrieval optimization, and lexicography. Example: the senses of 'bank' (financial institution) and 'bank' (river edge) tend to show moderate to low polysemysimilarity, whereas 'book' (a bound volume) and 'book' (to reserve) may exhibit higher similarity due to shared administrative or recording semantics.

Challenges include defining appropriate sense granularity, cross-language applicability, and domain shift. The term polysemysimilarity is not universally standardized but appears in discussions of sense inventories and vector-based sense representations as a descriptive label for internal sense relatedness.