Computational Modelling of Spoken Word Recognition in the Auditory Lexical Decision Task
Filip Nenadić · ERA: Education and Research Archive (University of Alberta) · 2020
The process of spoken word recognition has been an important topic in the field of psycholinguistics for decades. Numerous models have been created, many of which received their own computational implementation. However, large-scale simulations using these models performed on the same dataset by an independent researcher are rare at best. In the present dissertation, three models of spoken word recognition (TRACE, DIANA, and the discriminative lexicon approach) are tested in their ability to simulate the spoken word recognition process as captured by the auditory lexical decision task. The simulated data comes from the Massive Auditory Lexical Decision project, a large-scale study that enables us to estimate model performance on thousands of English words and compare it with performance of hundreds of human listeners. The main goals of the present work are threefold. The first goal is to assess models' performance in simulating the auditory lexical decision task. The second goal is to learn about the process of spoken word recognition through differences in models and model setups. The third goal is to provide suggestions for model improvement or future model development. The dissertation begins by outlining the history of development and the current state of computational models of spoken word recognition, motivating the conducted research. The central part of the dissertation is split into three separate sections. The first section describes the TRACE model in more detail and the simulations of MALD data performed using TRACE's re-implementations called jTRACE and TISK. The second section describes an implementation of an end-to-end model of spoken word recognition called DIANA and simulations performed using that model. The third section presents the simulations performed using the discriminative lexicon approach to spoken word recognition. Each of these sections includes a separate discussion of the results, focusing predominantly on the model in question. A joint conclusion brings together the findings from these three separate studies and also includes a suggestion to creating a hybrid model using strong aspects of the tested computational models of spoken word recognition.