|
Title
|
|
|
|
Dynamic reading : the processing of orthographic features by dynamical systems
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| During single word reading, words are primarily identified by their visual properties: the letters or characters that make up that word. From a common sense point of view, this process of identification can be considered to be like looking up words in a dictionary. Several phenomena, however, point towards a much larger role for interactive, dynamic processing during visual processing: the identification of a word depends on the other words a person knows, and is not just the result of a visual stimulus. One of these phenomena is the neighborhood effect, which is the effect that words that look more like other words are read more quickly. For example, the word “BOOK” looks like several other words, such as “NOOK”, “ROOK”, “TOOK” and “LOOK”, while “SONIC” only looks like “TUNIC” and “TONIC”. All things being equal, “BOOK” is therefore read faster. The role of the neighborhood effect, and modeling the neighborhood effect using dynamic models of word recognition, is the topic of this thesis. The thesis is divided into two parts, and is prefaced by an introductory theoretical chapter. The theoretical chapter mainly deals with the philosophical preliminaries of the work, and specifically with the notion of representation employed in many theories and models of word reading, and how a specific notion of representation is employed in many theories of word reading, including those governing the neighborhood effect. The first part of this thesis deals with measuring the neighborhood effect, and specifically with the notion of similarity employed in the definition of the neighborhood effect. In short, there exist many featurizations of orthography, also called orthographic codes, all of which result in differences in how words can be said to be orthographically similar. According to some feature sets, the words “TREE” and “THREE”, for example, are very similar, while to others, these words are very dissimilar. In the first chapter, we discuss wordkit, a Python package containing implementations of all these feature sets, along with tools for accessing popular corpora. This toolkit thus facilitates research in word reading. In the second chapter, we compare several feature sets, and introduce a new metric for comparing them, which we call Representation Distance (RD). This metric is the sum of the distances to the 20 closest neighbors for an arbitrary vector space. Previous metrics only operated on string representations, and therefore could not be used to compare theories of orthographic representation. In a regression analysis using French, Dutch and British English corpora, we show that feature sets that are more flexible, i.e., feature sets that assign higher similarity to THREE and TREE, explain less variance in lexical decision latencies. We also show that RD, the newly proposed metric, can be used to analyze hidden state matrices of Multi-Layer Perceptrons. In the third chapter, we expand this analysis, and optimize the different feature sets by jointly optimizing the number of nearest neighbors taken into account in the calculation of RD, and the parameters of the feature sets. We apply this to five alphabetic languages: the three languages used in the previous study, as well as American English and Spanish. We show that all results of the optimizations are in agreement across these languages. We also introduce a feature weighting scheme based on a discrete version of negentropy, the complement of entropy. Finally, we also present results on Hangul, the alphabet of Korean. These results diverge from the results on the other alphabetic languages, and, as such, raise questions of whether the way we measure the neighborhood effect is specific to alphabetic languages. The second part of the thesis deals with the dynamic models that process these orthographic features. In this part, we specifically study the Interactive Activation (IA) model, a dynamic model of word recognition first published in 1981. In a first chapter, we introduce Metameric, a toolkit for the IA model and related models, together with some empirical work demonstrating some of the assumptions inherent in the IA model and IA networks. We specifically show that the assumption that IA models can only use words of a specific length is false, and show how the model can be extended to use words of any length. In the second chapter, we present a working version of the BIA+ model, a bilingual version of the IA model. This model, although very popular, has only been available as a theoretical model, and has not been implemented before. Using a process called rational reconstruction, we use the theoretical version of the BIA+ model to create an implemented version. We show that one of the main theoretical commitments of BIA+, that language information is explicitly represented, is not backed up by the implemented version of the BIA+ model. In the final chapter, we introduce a new model, called SIMBL, which is a simple dynamic model of lexical decision. This model is more flexible than the IA model in terms of which orthographic features it accepts, but still can explain some common results in lexical decision experiments. Finally, we present a short conclusion in which we discuss how these dynamic models could be extended to a fully visually grounded model of reading. |
| |
|
Language
|
|
|
|
English
| |
|
Publication
|
|
|
|
Antwerp
:
University of Antwerp, Faculty of Arts, Department of Linguistics
,
2020
| |
|
Volume/pages
|
|
|
|
180 p.
| |
|
Note
|
|
|
|
:
Daelemans, Walter [Supervisor]
:
Sandra, Dominiek [Supervisor]
| |
|
Full text (open access)
|
|
|
|
| |
|