Distant Concept Connectivity in Network-Based and Spatial Word Representations

AbstractIt is presently unclear how localized, word association network representations compare to distributed, spatial representations in representing distant concepts and accounting for priming effects. We compared and contrasted 4 models of representing semantic knowledge (5018-word directed and undirected step distance networks, an association-correlation network and word2vec spatial representations) to predict semantic priming performance for distant concepts. In Experiment 1, response latencies for relatedness judgments for word-pairs followed a quadratic relationship with network path lengths and spatial cosines, replicating and extending a pattern recently reported by Kenett, Levi, Anaki, and Faust (2017) for an 800-word Hebrew network. In Experiment 2, response latencies to identify a word through progressive demasking showed a linear trend for path lengths and cosines, suggesting that simple association networks can capture distant semantic relationships. Further analyses indicated that spatial models and correlation networks are less sensitive to direct associations and likely represent more higher-level relationships between words.

Return to previous page