ABSTRACT
Abstraction is a core principle of Distributional Semantic Models (DSMs) that learn semantic representations for words by applying dimensional reduction to statistical redundancies in language. Although the posited learning mechanisms vary widely, virtually all DSMs are prototype models in that they create a single abstract representation of a word’s meaning. This stands in stark contrast to accounts of categorisation that have very much converged on the superiority of exemplar models. However, there is a small but growing group of accounts in psychology, linguistics, and information retrieval that are exemplar-based semantic models. These models borrow many of the ideas that have led to the prominence of exemplar models in fields such as categorisation. Exemplar-based DSMs posit only an episodic store, not a semantic one. Rather than applying abstraction mechanisms at learning, these DSMs posit that semantic abstraction is an emergent artifact of retrieval from episodic memory.
Acknowledgements
This work was supported by NSF BCS-1056744 and IES R305A150546. I would like to thank Randy Jamieson, Melody Dye, and Brendan Johns for helpful input and discussions.
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
1 LSA and word2vec are formally equivalent: Levy and Goldberg (Citation2014) demonstrated analytically how the SGNS architecture of word2vec is implicitly factorizing a word-by-context matrix whose cell values are shifted PMI values.
2 The model’s direction can also be inverted, using the word to predict the context (SGNS) rather than using the context to predict the word (CBOW).
3 The bulk of the evidence used by Tulving to argue for distinct semantic and episodic memory systems was from neuropsychological patients.
4 The model is simply referred to as the “semantics model” in Kwantes’ (Citation2005) original paper, but “Constructed Semantics Model” has become it’s popular name among semantic modelers because semantic representations are constructed on the fly from episodic memory in the model.
5 An exception here is the topic model, which uses conditional probabilities, so it is not subject to metric restrictions of spatial models (e.g., Griffiths, Steyvers, & Tenenbaum, Citation2007).
6 And essentially the same architecture has been used by Goldinger (Citation1998) to explain “abstract” qualities of spoken word representation from episodic memory retrieval.