To read this text, the first thing you did was to orient your eye to the beginning of this sentence. Now you move your eye, stopping on every word, trying to extract letters so as to identify this word among all words you know, to pronounce it, possibly out loud, and then going to the next word. At least, this is the impression you have.
These complex mechanisms, like gaze movement, extraction of orthographic features and pronunciation are domains which have been largely studied, more or less independently, by researchers. For instance, studies on gaze movement have shown that the eye does not actually stop on every word, and that eye positions on the sentence are not random, but the result of planned movements (Vitu, O’Regan, & Mittau, 1990). Similarly, concerning pronunciation, it is likely that you do not need to recognize all the letters in familiar words, in order to read them aloud. This is what is described in dual-route models, such as the Dual Route Cascade model (DRC ; Coltheart et al. 2001) and the Connectionist Dual Process (CDP+ ; Perry et al., 2007).
Since the 80s, researchers have begun to use computers as investigation tools, by implementing computational models. The first such model of word recognition was the Interactive Activation (IA) model of McClelland and Rumelhart (1981). This model has been used to show that word recognition is finely influenced by words that are already known. The IA model is also the basis of recent popular reading models : the Dual Route Cascade model (DRC ; Coltheart et al. 2001) and the Connectionist Dual Process (CDP+ ; Perry et al., 2007). However, these models have a hard time accounting for common recent topics, such as positional letter coding and reading acquisition. The problem of positional coding is difficult, because position is crucial to distinguish words such as “lion” and “loin”. Still, we can raed txets in wihch wdors hvae thier ltetres srcmbeland. This problem is usually solved by models that are specifically designed for it, such as the SERIOL (Whitney, 2001) or SCM models (Davis, 2010). The problem of modeling reading acquisition is more naturally dealt with by connectionist models (McClelland and Seidenberg, 1999). A first attempt at implementing a dual-route model featuring self-learning was performed recently (Ziegler et al., 2014). However, these models involve learning mechanisms that do not correspond to those of young readers ; they are not plausible.
The BRAID project
The first objective of our project is to build a model to study the cognitive mechanisms involved in visual word recognition. To do so, we have developed BRAID, a Bayesian algorithmic model of written word recognition.
To begin with, this model is original because of the used framework. Knowledge involved in the model is expressed as probability distributions, and relationships between these as probabilistic dependencies. Every task that BRAID simulates, such as letter identification or lexical decision, is defined by a probabilistic question. Answers provided by the BRAID model are the results of inference, based on the knowledge put in the model.
Concerning its contents, BRAID is both classical and original. In a classical manner, the model features a parallel and dynamic structure with three layers. The first layer allows dynamic extraction of visual information, dependent on gaze position and acuity. The second layer allows accumulating visual information into internal representations of letters. Finally, the third layer describes lexical knowledge, that is to say the relationship between known words and the letters that they feature.
To this classical structure, BRAID adds first of all a model of lateral interference, at the level of letter perception. The most crucial originality of our model is that it involves a visual attention model, which acts as a filter between the perception of written traces and their memorization. This filter speeds up information accumulation in interesting areas, contrary to less informative areas.
Ongoing research with the BRAID model aim at exploring how it accounts for experimental observations of expert or impaired readers, first, and extending it towards a model of orthographic knowledge acquisition, second.