Poster E29, Saturday, August 18, 3:00 – 4:45 pm, Room 2000AB
Network Analysis of Concreteness within Association Modules
Dominick DiMercurio1, Chaleece Sandberg1;1Pennsylvania State University
A large body of evidence from neuroimaging meta-analyses (Binder et al., 2005, 2009; Wang et al., 2010), behavioral studies (Binney et al., 2016), and clinical cases (e.g., Jeffries et al., 2007, 2009) suggests that abstract and concrete words are dissociable; furthermore, models of neuroimaging or behavioral data provide further insight into the possible mental organization of this dissociation. For instance, binary classification methods such as logistic regression (Wang et al., 2013) and support vector machines (SVMs), applied to neuroimaging data, have been used to decode concreteness. On the behavioral end, clustering analyses based on cognitive scores have been used to identify patterns that appear to dissociate concreteness to some degree as well (Troche et al., 2017). To our knowledge, no study to date has used word association to decode concreteness; however, it appears plausible as a word association study (de Groot, 1989) found that concrete words produced more, faster, and less diverse associations than abstract words. Notably, she also found that concrete words strongly cued concrete words, and abstract words cued both types of words. The present study investigates, using free association norms, how abstract and concrete words pattern within semantic domains. We hypothesize that decoding concreteness from features of an association network, as a model of semantic organization, should improve if those features are selected based on measures related to graph communities, which model semantic domains. To test this hypothesis, we developed a 4763-word directed weighted association network from the University of South Florida association norms databases. After extracting communities from the network, we analyzed measures of segregation and integration for each cue and target. For testing the representativeness of the network, we investigated the relative transitions based on concreteness classification and word frequency. In comparison with findings from de Groot, we found that abstract words more strongly cued abstract words and that concrete words more strongly cued concrete words; in addition, we found that target words were of significantly higher frequency than cue words. It is also worth noting that there is a heavy concrete-bias in the norming database, with the bias roughly 6:1 in cue words and 8.5:1 in target words. Overall, we found good performance in SVMs built from distance features to highly segregated cues (69% accurate) or highly integrative targets (64% accurate), comparable to neuroimaging-derived SVMs (60-75% accurate) and higher than SVMs built from distance features to random nodes (50% accurate). From inspection of the communities, some communities seem to be more abstract-biased and some are more concrete-biased; however, the distributions also showed that cue segregation measures were higher for concrete than abstract words and that target integration measures were reverse. In conclusion, these results show that a model based on association norms can discriminate concreteness, semantic domains appear to vary in abstract or concrete content, concrete words tend to be more segregated while abstract words tend to be more integrative, and examining the community structure of semantic networks using graph theory is useful for modeling the organization of the semantic system.
Topic Area: Meaning: Lexical Semantics