The Spreading-Activation Framework Does not Explain the Effects of Degree and Clustering on Spoken Word Recognition

Abstract

AbstractDuring spoken word recognition (SWR), various lexical properties affect speed and accuracy, and computational models offer competing explanations for these effects. Classical psycholinguistics has focused on properties internal to individual lexical items such as the facilitative effects of high probability at various levels of processing, rather than focusing on the relations between lexical items. Even connectionist models, which represent SWR as a dynamic process of activation flowing over a hierarchical network of lexical and sublexical representations, have not focused on that network’s topology nor on how a word’s position in that network affects its recognition.In contrast, the Cognitive Network Science (CNS) approach has addressed how network topology affects a variety of psycholinguistic tasks. For example, speed and accuracy of perception has been found to be affected by a word’s degree on a simple “phonological neighborhood network” (PNN), in which words are connected to similar-sound words. Similar effects have been found for other network properties, and several simulation studies have suggested that (at least some of) these behavioral effects can be explained by a simple spreading-activation algorithm along PNNs, rather than positing a hierarchical network of distributed processing units. In particular, some CNS adherents have claimed that the spreading-activation framework, in combination with the techniques of network science, can explain the effects of degree and clustering on SWR performance.The present study rigorously explores the state-space of the proposed spreading-activation framework. Adopting the agent-based modeling approach and implementing the process in NetLogo provides a powerful tool to explore this state-space. A series of simulations shows that the framework predicts word-degree in the PNN to have an effect on SWR performance in the opposite direction from observed behavioral results, contrary to previous studies. This suggests previous results are an artifact, presumably resulting from network construction techniques and linear regression modeling that did not reflect the system’s complexity. The effects of clustering are also analyzed, which suggest the model’s predictions are not as straightforward as has been suggested by prior studies. Implications for models of spoken word recognition are discussed.