Setting the Balance between the Lexical and Sublexical Pathways of Dual-Route Models of Reading: Insight from Atypical Dyslexia in Surgical Glioma Patients.

Abstract

Reading is a complex skill that has been acquired only recently in the evolution of humans. This fascinating ability has motivated many neuropsychological models. In particular, it has been proposed that reading is sustained by two parallel and complementary systems (see for example Epelbaum et al., 2008 for a plausible anatomical implementation of this model): a lexico-semantic one and a phonological one (that further requires attentional resources for processing serially the different syllables of a given word). In parallel to advances in neuropsychological models of reading, computational models have been intensively studied. Such models attempt to go further than the usual “boxology”, by providing a simulation on a computer of the different cognitive processes that are involved in reading. One of the very first version in 1989 (Seidenberg and McClelland, 1989) was a connectionist approach, known as the “triangle model”, that already implemented the two routes (a semantic one and a phonological one). Since then, two categories have been competing in order to achieve the best fit between simulations and human datas in healthy and brain-damaged individuals: the dual-route cascade (DRC) (Coltheart et al., 2001) and the connectionist dual-process (CDP) (Zorzi et al., 1998) and its updated versions CDP+ (Perry et al., 2007), CDP++ (Perry et al., 2010), and CDP++-parser (Perry et al., 2013). Both kind of models share the same architecture regarding the lexical route, while differing in the sublexical route. In both cases, the balance between the lexical and sub-lexical route is a critical issue: whenever the lexical route is too strong, pseudowords will be read as a lexical neighbor, and whenever the sub-lexical route is too strong, error rates rise for irregular words (which are regularized by the sublexical route). As clearly explained in the paragraph “Searching parameter space” of Coltheart et al. (2001), the optimal balance is found in the DRC model by trial and error, testing the ability of the system to read a specific pair of irregular word and pseudoword (respectively “chef” and “starn”). Similarly, it is stated in Perry et al. (2007): “The first step was to determine the appropriate balance between lexical and sublexical phonology, which in turn largely depends on the speed at which the serial process of grapheme parsing occurs. These parameters need to be chosen together, because slower grapheme parsing speeds reduce the amount of sublexical phonology in the model, and faster speeds increase it. Performance on irregular words provides a particularly important benchmark for parameter setting.” Importantly, in both models, there is no explicit computational system that resets the balance between the two routes depending on the characteristics of each trial (regular, irregular, and pseudowords).