Proceedings of the Annual Meeting of the Cognitive Science Society

It is presently unclear how localized, word association networkrepresentations compare to distributed, spatial representationsin representing distant concepts and accounting for primingeffects. We compared and contrasted 4 models of representingsemantic knowledge (5018-word directed and undirected stepdistance networks, an association-correlation network andword2vec spatial representations) to predict semantic primingperformance for distant concepts. In Experiment 1, responselatencies for relatedness judgments for word-pairs followed aquadratic relationship with network path lengths and spatialcosines, replicating and extending a pattern recently reportedby Kenett, Levi, Anaki, and Faust (2017) for an 800-wordHebrew network. In Experiment 2, response latencies toidentify a word through progressive demasking showed a lineartrend for path lengths and cosines, suggesting that simpleassociation networks can capture distant semanticrelationships. Further analyses indicated that spatial modelsand correlation networks are less sensitive to directassociations and likely represent more higher-levelrelationships between words.