Thesis defence : Alexandra STEINHILBER

The major theoretical framework for the study of reading acquisition is the self-teaching theory. According to this theory, learning to read is based on incidental learning of novel orthographic forms through successful phonological decoding. Context could play a facilitating role when phonological decoding is partially correct.To date, only two computational models of self-teaching exist. They are based on the dual-route architecture, which assumes different processing for reading known words and novel words. The processing of known words is performed by direct access to the lexicon, whereas the processing of novel words is performed by graphemic segmentation, then application of grapheme-phoneme conversions, independent of lexical knowledge. These models condition the orthographic learning of a novel word on prior knowledge of its phonological form.However, behavioral data question the assumptions made by these models. Several studies discuss the relevance of the grapheme as the primary psycholinguistic unit of decoding and suggest that reading, even of novel words, is performed by analogy to lexical knowledge. Furthermore, behavioral studies show that incidental orthographic learning without prior phonological knowledge is possible.This thesis focuses on modeling reading acquisition and is based on the theory of self-teaching. We evaluate the hypothesis that a lexical processing, not relying on any predefined psycholinguistic unit, but using visual attention to segmentate the stimulus, is able to simulate successful self-teaching. We also hypothesize that orthographic learning is possible even without context and when the phonological form is not previously known, even though these two dimensions are facilitating.We propose a new probabilistic computational model for reading acquisition, named "BRAID-Acq". It has a single-route architecture, and a visual-attentional submodel that allows for spatially arbitrary visual information taking, without necessarily aligning with a presupposed psycholinguistic unit. The model also has a phonological attentional submodel, coupled with its visual counterpart, to relate orthographic and phonological segments. It simulates the dynamics of attentional exploration during processing.The validation of our model was done in two steps. In a preliminary, purely visual version of BRAID-Acq, we show that the model simulates the evolution of oculomotor behaviors across repeated exposure to novel words. We show that the amount of visual attention in the model impacts this evolution, as well as length and lexicality effects. Next, we show that the full BRAID-Acq model is able to correctly read most novel words based on flexible sub-lexical processing, which does not involve graphemic segmentation or grapheme-phoneme conversions. It successfully simulates a variety of self-teaching situations (with and without context, with and without prior phonological form), but using a unique processing. We show that the presence of context and knowledge of phonological form are facilitative but not essential for learning. In particular, our model's context helps disambiguate the reading of novel words when it is difficult, such as when the word is irregular, when the language is opaque, or when the level of orthographic knowledge is low. In conclusion, the BRAID-Acq model successfully simulates self-teaching, which supports our hypotheses.

Jury :
Sylviane Valdois - CNRS - Directrice de thèse
Julien Diard - CNRS - Co-directeur de thèse
Pierre Bessière - CNRS - Rapporteur
Franck Ramus - CNRS - Rapporteur
Marie-Line Bosse - UGA - Examinatrice et Présidente du Jury
Marie Lallier - BCBL - Examinatrice
Fabienne Chetail - ULB - Examinatrice

Read the thesis