Neighborhood density effects have a long history in the psycholinguistic literature as a way of operationalizing phonological or orthographic similarity between words, with well-documented, if small, effects on lexical processing, vocabulary development, and phonological development. However, neighborhood density has important limitations, and it has increasingly become regarded as a somewhat blunt, if useful instrument. One way of sharpening it, as it were, has been pursued by researchers employing the emerging tools of network science. These researchers point out that, when all words in a lexicon are connected to their neighbors, the result is a complex network, and a word’s position in that network reflects much more than its relationship with its immediate neighbors. Several studies have examined one or two network measures in isolation (e.g. degree and clustering coefficient) on highly controlled stimulus sets; however, many network-theoretic measures are conceptually related and are not mathematically independent (Rubinov & Sporns, 2010). This makes it difficult to interpret the separately reported effects for individual measures as indicators of how network structure may be related to the organization and functioning of the mental lexicon.
In the present work, we sought to examine how a wide variety of network-theoretic properties jointly relate to lexical processing. We considered a wide range of network measures (e.g. degree, clustering coefficient, centrality, and efficiency), and used them (alongside word frequency and length as covariates) to predict behavioral measures for a large and diverse database (20,930 words in the English Lexicon Project). We employed statistical methods that enable us to both cope with and explore mathematical relatedness among measures, namely, partial least squares regression and decision trees. A partial least squares regression between lexical properties and behavioral data revealed that frequency, length, and the network properties appear to work together, suggesting that network properties may not play a substantial role in processing that is independent of frequency and length. Further investigation through decision tree analysis elaborates on these findings, suggesting that word frequency is the most reliable of the tested variables for predicting response times, with network properties becoming explanatory under specific conditions. Together, these results suggest that neighborhood density and other related measures of how words can be embedded within orthographic and phonological networks may play a smaller role in processing time than the previous literature may lead one to believe. Our work has implications to guide the development of more refined approaches in future research in networks for representing the mental lexicon.