The multiplex interplay between phonological and semantic networks impacts word production across different types of aphasia

The structure of mental lexicon networks, where similar words are connected, influences several language processes in healthy (e.g., De Deyne, Navarro, & Storms, 2013; Vitevitch, Goldstein, Siew, & Castro, 2015) and clinical (e.g., Castro, Pelczarski, & Vitevitch, 2017) populations, above and beyond traditional psycholinguistic measures. Specific to aphasia, phonological networks were found to influence picture naming performance by people with Broca’s and Wernicke’s aphasia (Vitevitch & Castro, 2015). However, these previous endeavors only examined one type of word-word similarity (e.g., phonological similarity; Vitevitch, 2008). The more recent approach of multiplex lexical networks (Stella, Beckage, & Brede, 2017; Stella, Beckage, Brede, & De Domenico, 2018) allows consideration of semantic and phonological relationships among words simultaneously. We extend the previous analysis of picture naming by people with aphasia to include the multiplex lexical network defined in Stella et al. (2018), where words are connected on four layers representing free associations, synonyms, generalizations, and phonological similarities. We focused on three network measures, namely degree, PageRank, and closeness centrality, for each word on each layer and over the entire multiplex. Degree is a local network measure equal to the number of immediate links of a node. PageRank highlights meso-scale network patterns by identifying the most important nodes in the network. Closeness centrality is a global network measure quantifying how fast information can flow from one node to all others. As multiplex generalizations of these measures we used multidegree (the sum of word degrees across layers), PageRank versatility (De Domenico, Solé-Ribalta, Omodei, Gómez, & Arenas, 2015), and multiplex closeness centrality, respectively. To assess picture naming performance, we conducted logistic regression models predicting the probability of correct picture naming from traditional psycholinguistic predictors (word length, frequency, and age of acquisition), type of aphasia (Anomic, Conduction, Wernicke, Broca, and healthy controls), and each of the network measures (layer-specific and multiplex variants) using archival data. We found that for local and meso-structural measures (i.e. degree and PageRank), the layer-specific variants were more predictive of picture naming performance than the multiplex variants. Conversely, for the global measure of closeness centrality, we found the multiplex variant being twice more predictive of picture naming performance than layer-specific variants. Furthermore, the difference in production likelihood between words with high and low multiplex closeness centrality is considerably greater for people with different types of aphasia than healthy controls, suggesting that this measure is particularly sensitive to picture naming performance in this clinical population (Fig. 1). These findings indicate that picture naming by people with aphasia crucially depends on the global interplay between phonology and semantics, which is captured by the whole multiplex structure. These findings may have particular relevance for the treatment of aphasia. Potentially, selecting words with high multiplex closeness centrality for treatment will lead to greater generalization of treatment to untrained words than when selecting words with low closeness centrality, given that words with high multiplex closeness centrality are 1) already better accessed by people with aphasia and 2) well-connected within the entire multiplex lexical network through semantic and phonological connections.