Comparing competing models of retrieval processes

Abstract

In order to correctly parse a sentence, its underlying structure needs to be understood. The functional task of every word in a sentence stands in relation to other words through the notion of dependency, and the task for the person taking in a sentence is to lay such dependency links between the word they are currently attending and one of the previously attended words. How exactly the chosen previously attended word is retrieved from memory is often under specified. Therefore Nicenboim & Vasishth (2018) compared two models with each other in terms of their power to describe the speed/accuracy trade-off of this process. We expand on this work by using these two models with a dataset that includes individuals with aphasia. The first model is the activation-based race model. It assumes that resolving syntactic dependencies is related to the activation of previously retrieved candidate dependants. When confronted with a new item, these candidates accumulate activation over time. The dependant (and thus, the interpretation) associated with the accumulator that first surpasses its threshold is chosen. The second model is the direct access model which assumes instant access to previously retrieved words. The difference in listening times here is explained by a backtrack-and-repair process that may take place when the initial parse is deemed incorrect. The activation-based race model is implicitly assumes that incorrect interpretations are generally associated with longer listening times, whereas the direct access model is ties incorrect interpretations with shorter listening times. The resulting fits on the empirical data show that although the data tells us that the mean listening times for all of its cross sections are shorter for incorrect trials, the direct access model does not perform better. Instead, the resulting fits indicate that both models have problems fitting certain different aspects of the data.