Skip to main content

Thesis defence : Ali SAGHIRAN

Thesis defence

On 10 June 2021

Bayesian modelling of reading

Most computational models of reading adopt a two-route architecture that assumes that the knowledge involved in serial processing differs fundamentally from the knowledge involved when processing is parallel. Thus, the reading of new words or pseudowords involves a system of grapheme-phoneme correspondences via the sublexical route that operates serially, whereas the reading of known words relies on the activation of lexical knowledge via the lexical route that is parallel. These models also suppose that the grapheme-phoneme conversion system is explicit, independently stored, and necessary for the reading of pseudowords. They assume, moreover, that this conversion system is preceded by a system of segmentation of the word into subunits to be converted independently.However, other reading models assume that parallel and serial processing involve the same type of knowledge. These two classes of models agree on the explanation of parallel processing in reading but do not agree on the description of serial processing. They, therefore, differ on their explanation of length effects, i.e. the observation of longer response times for long stimuli. Indeed, dual-route models can only interpret these length effects through serial decoding of the sublexical channel. Moreover, even if all models agree on the role of visual attention in serial processing, the corresponding mechanisms are not well described, especially mathematically, in the literature.This thesis aims to evaluate the hypothesis that a treatment involving only learned lexical knowledge about words can account for sublexical relationships between orthographic and phonological units, simulate length effects in different types of tasks, and perform sublexical segmentation of new words.For this purpose, we propose a new probabilistic computational model of reading called "BRAID-Phon". This model is an extension of the computational word recognition model of the "BRAID" model by adding a phonological knowledge sub-model. We use the BRAID-Phon model to study the plausibility of a system based on a "single-route" architecture, i.e. including only orthographic and phonological lexical knowledge to perform a reading simulation. We show the ability of BRAID-Phon to account for length effects on words in three types of tasks (reading, lexical decision, and progressive unmasking) and study the role of implemented visual attention mechanisms on these effects. Finally, we illustrate the need for an attention-controlled segmentation process to perform pseudoword reading

- Directrice de thèse : Sylviane VALDOIS - (Sylviane[dot]Valdois[at]univ-grenoble-alpes[dot]fr)
- Codirecteur : Julien DIARD - (julien[dot]diard[at]univ-grenoble-alpes[dot]fr)

Keywords: Phonological representation, Lexical recognition, Language, Visual attention,

Read the thesis



On 10 June 2021


CNRS - eFRAN Fluence

01/10/2017 - 10/06/2021

Submitted on 20 November 2023

Updated on 20 November 2023