EXEMPLAR BASED WORD MEANING IN THE AGE OF CONTEXTUAL EMBEDDINGS

Abstract

This article discusses a closely related alternative to prototype-based accounts of lexical meaning: exemplar-based and usage-centered approaches to words and senses. Instead of treating a word as a container of a small set of discrete dictionary meanings, exemplaroriented models assume that language users retain memory traces of encountered uses and interpret new instances by similarity to stored experiences. The article traces the cognitive foundations of exemplar models in categorization research and shows how their logic aligns with corpus-driven semantics and the long-standing critique that “word senses” are taskdependent abstractions from usage. It then connects these ideas to contemporary computational semantics, where contextualized language models (e.g., ELMo, BERT) produce token-level representations that vary with context, and where word sense disambiguation is increasingly approached through nearest-neighbor decisions in embedding space or by aligning contextual vectors to knowledge-base senses. Finally, the paper outlines unresolved questions on sense granularity, interpretability, and the relation between psychologically plausible exemplars and engineering-oriented embeddings, arguing that the most productive view treats “senses” as emergent clusters within a continuum shaped by frequency, discourse goals, and domain constraints.